A Comprehensive Survey of Artificial Intelligence Techniques for Talent Analytics

Recruitment data in pre-employment mainly includes the following types:

Regarding the development of employees within a company, a significant amount of employee data has been accumulated, including training records and individual work outcomes. An overview of employee-related data is shown in Figure 4.

An organizational structure is a system that outlines how activities are directed toward the achievement of organizational aims [330], which plays an important role in decision-making and knowledge management. Generally, an organization is commonly represented as a hierarchical tree structure, which can take on diverse forms, such as matrix, flat, and network structures. Figure 5 shows several common types of organizational structures. Typically, existing studies explore these complex structures from various dimensions, such as reporting lines and in-firm social networks.

Reporting lines are generally the most representative aspect of an organizational structure, which delineates how authority and responsibility are allocated in an organization. Regarding this point, Sun et al. developed an organization structure-ware convolutional neural network to hierarchically extract compatibility features for measuring person-organization fit and its impact on talent management [381, 383]. Nevertheless, due to privacy concerns, mainstream studies utilize in-firm social networks for human resource management. In general, an in-firm social network can be formed from email or Instant Messaging (IM) records across employees. For example, text-based communication has been used by several machine learning classifiers to identify group mood [213]. Besides, Cao et al. leveraged the lasso regression model to explore team viability using text conversations of online teams [62]. In addition to social networks, researchers have also taken other information into account. For example, Ye et al. utilized both email communication and a high-potential talent list to identify employees with high potential [447]. Along this line, Teng et al. further utilized datasets from three sources for organizational turnover prediction, including profile and turnover, social network, and job levels [386].

The source of recruitment data can be categorized into two types: internal data and external data. Correspondingly, data collection methodologies align with these source categories.

After collecting a large amount of recruitment data, it is essential to preprocess the data for downstream applications, especially removing noisy, redundant, irrelevant, and potentially toxic data. In this part, we review the detailed data preprocessing strategies to improve the quality of the collected data according to various data types.

After initial data preprocessing, standard structured data is obtained. However, given the potential for noise introduced during the data acquisition process, coupled with non-uniform acquisition methods, the provided data may not be of high quality. Therefore, a cleanup and debiasing of the data becomes imperative. In this part, we first introduce several data quality issues commonly seen in AI-driven talent analytics. Subsequently, several clean and debias methods are introduced to solve these issues.

Existing data and processing methods in the field of talent analytics continue to face many limitations and challenges, which constrain the development of AI-based approaches. We summarize these issues as follows.

• •

  Lack of benchmark datasets: Most current research in talent analytics heavily depends on proprietary data related to job applicants, employees, and organizations. The sensitive nature of this data often prevents it from being publicly accessible, severely restricting the development of publicly available benchmark datasets. This limitation impedes the standardization of problem definitions and comparative analyses of methodologies, thus slowing down technological advancements in the field. While some job-related data has been made open-source, it remains limited in scale and temporal coverage. Consequently, there is an urgent need to create comprehensive open-source benchmark datasets for pivotal tasks within this field. We present several open datasets, as shown in Table II, primarily related to resume parsing and employee turnover prediction. These datasets can help researchers in develo** standardized datasets for uniform comparisons. Additionally, the data anonymization and de-identification methods require more extensive consideration to facilitate the construction of more open-source datasets for various scenarios.

• •

  Diversity deficiencies in datasets: The datasets commonly used in AI-based talent analytics often lack adequate consideration of diverse populations, especially underrepresented minorities. This deficiency hampers efforts to evaluate and ensure algorithmic fairness. Models developed from such datasets may exhibit biases, leading to inequitable outcomes among various demographic groups. Therefore, it is essential to integrate more diverse datasets to promote fairness and enhance the overall performance of AI algorithms in talent analytics.

• •

  Challenges with subjective data: In many talent analytics scenarios, such as employee promotion decisions, outcomes are frequently based on the subjective judgment of supervisors. This practice can introduce inherent biases and noise into the real datasets collected from these decisions. Traditional data preprocessing techniques often fail to detect and address these anomalies effectively. Therefore, there is a pressing need for specialized noise detection algorithms or the development of robust AI-based talent analytics models that can handle such irregularities. These approaches are essential to ensure the accuracy and fairness of the analytics outcomes.

A job posting comprises both job duties, which describe the responsibilities and tasks of the role to candidates, and job requirements, which outline the professional experience, skills, and domain knowledge that an employer expects from the ideal candidate to perform the role. Job duties are tailored to each specific position. However, identifying the necessary capabilities and aligning them with job duties can be challenging, particularly when recruiters have limited experience and domain knowledge. A research report from Allegis Group indicates that many organizations face difficulties when writing job descriptions, as only 39% of candidates find these descriptions clear and easy to understand [149, 332].

Technically, this task can be viewed as a text-to-text generation problem, where the job duties and job requirements are typically long sequences of text. To address this task, sequence-to-sequence models using an encoder-decoder architecture are commonly employed, as illustrated in Figure 6. For instance, Liu et al. applied two Long-Short Time Memory (LSTM) layers as the encoder and decoder, respectively, to extract the key information from job duties and generate the job requirement, respectively [254]. Since it is important to precisely use and organize skill-related keywords in job requirements, they implemented the decoder in a two-pass manner. The first-pass decoder focuses on predicting skill-related keywords, while the second-pass decoder uses these predicted skills to guide the generation of fluent text. In particular, the attention mechanism is employed to integrate hidden states from the LSTM layer of the encoder with context information, which enhances skill prediction in the decoder. Furthermore, Qin et al. trained a neural topic model in job duties to capture the global topic information [332]. The topic distribution of each job duty is used as the context information to guide the generation of each word in the LSTM-based decoder.

Moreover, Borchers et al. attempted to generate job postings directly from job names using LLMs [52]. Furthermore, the authors employed prompt engineering and gathered high-quality, gender-unbiased training data to enable the generator to produce unbiased job postings.

Recent studies have demonstrated that ChatGPT and other LM-based models can be utilized for text evaluation, achieving performance close to that of human evaluators [143, 298]. This advancement opens new possibilities for the automated assessment of fluency, realism, and other dimensions in generated job postings.

Resume understanding, also known as resume parsing aims to extract semantically structured information from resume documents, which can facilitate a wide range of intelligent talent analysis applications, such as talent searching [161], person-job fitting [334, 473], and talent assessment [239]. As illustrated in Figure 3, a resume typically includes the candidate’s personal information (e.g., name, phone number), educational background (e.g., major), work experience (e.g., company name), and various other relevant details. Specifically, this task can be formalized as follows:

Typically, resume understanding begins by converting various forms of resume files, into textual formats. This conversion is accomplished using technologies or tools such as Optical Character Recognition (OCR) [69] and PDF parsers [443]. Consequently, resume understanding becomes a text-mining task, enabling the extraction of structured information from the textual data. The advancement of resume understanding technology is closely linked to innovations in information extraction techniques. Initially, Kopparapu utilized keyword searches and rule-based matching to extract six major fields of information from resumes: “name”, “software skills”, “qualification”, “experience”, and “email” [216]. Considering the hierarchical structure of resume information, for example, the “personal information” usually includes details such as “name” and “gender”. In [449], Yu et al. first divided a resume into different blocks (e.g., “personal information” and “education information”) using a Hidden Markov Model (HMM). They subsequently employed both Support Vector Machines (SVM) and Hidden Markov Models (HMM) to extract detailed information from these blocks. By leveraging traditional machine learning models such as HMM [449], SVM [449, 89], and Conditional Random Fields (CRF) [321, 74], resume understanding has achieved an accuracy rate of over 80% across most fields.

Although these traditional models achieve good performance, they incur the cost of extensive feature engineering. With the development of deep learning and its broad application in information extraction, many researchers have shifted to utilizing deep learning to construct more effective resume understanding models [27, 327, 198, 240, 262]. Most deep-learning approaches view resume understanding as a Named Entity Recognition (NER) task. Ayishathahira et al. utilized a Bidirectional LSTM-CNN architecture to perform resume information extraction [27]. Van et al. leveraged Convolutional Neural Network (CNN) to learn the character-level representations of words in resumes, and utilized the classic neural NER architecture, BiLSTM-CRF, to enhance resume information extraction performance [327]. Meanwhile, ** et al. integrated a highway network and self-attention mechanism into BiLSTM-CRF [198]. Subsequently, the introduction of pretrained language models such as BERT has significantly improved the representation of resume text, further enhancing the performance of deep learning-based resume understanding models [240, 262]. In experiments with a Chinese resume dataset, the methods have achieved an F1 score of up to 96%.

Prior research has primarily transformed resume documents into plain text inputs, thereby overlooking the multimodal information inherent in these files, such as layout information, which is crucial for resume understanding. Wei et al combined the pre-trained model RoBERTa with a Graph Convolutional Network (GCN) to extract structured information from resumes, taking into account layout and positional features [411]. Recently, the integration of multi-modal features with pretrained models for document understanding has gained widespread adoption [196, 428, 427, 429]. Inspired by the BERT architecture, LayoutLM was the first to enhance the masked language modeling task by incorporating the 2D coordinates of each token as layout embeddings [428]. This approach enables the model to capture interactions between text and layout information, leading to an improvement of approximately 15% compared to text-only methods. Subsequently, advancements in methods such as DocFormer [20], LayoutLMv2 [427], LayoutLMv3 [185], and LayoutXLM [429] have progressively enhanced the performance of multimodal pretrained models in document understanding. These technologies have also been extensively adapted for resume understanding. Given that resume documents are more text-centric and voluminous compared to traditional document types like receipts, Yao et al. proposed a hierarchical multi-modal pre-training model tailored for long document understanding [443]. Additionally, to tackle the scarcity of annotated training data for resumes, the authors developed a distantly supervised sequence labeling method. This method, trained within a self-distillation based self-training learning framework, significantly improved the model’s performance in resume understanding with limited training data. Jiang et al. introduced a novel layout-aware multi-modal fusion transformer that encodes resume segments by integrating textual, visual, and layout information [196]. Additionally, they developed a multi-granularity sequence labeling task to address the hierarchical relationships among the fields to be parsed.

Talent searching is designed to identify suitable candidates based on search queries provided by recruiters or hiring managers [209]. Formally, considering the candidate set $\mathcal{R}=\{R_{1},..,R_{n}\}$, where each candidate $R_{i}$ denotes his/her resume consisting of work experience, education experiences, and so on ¹¹¹¹11Here, $R_{i}$ usually refers to the structured data derived from the process of resume understanding.. Then, talent searching can be approached as an information retrieval task ¹²¹²12When the query directly corresponds to a job position, it aligns with the person-job fitting task, which is detailed in the following subsection and will not be further discussed here.:

The key to talent search lies in measuring the fitness between the search query $q$ and each resume $R_{i}$. Manad et al. established connections between queries and candidates through skill matching. They extracted skill information from both queries and resumes, then ranked candidates based on scores that reflect the level of skill proficiency demonstrated in the resumes [274]. Wang et al. focused on queries composed of competency keywords, such as skill and activity keywords. They utilized the BERT model to extract competency keywords from resumes and applied the weighted average method to calculate candidates’ scores for different competency keywords. Meanwhile, the authors enhanced the effectiveness of talent search by leveraging a structured Competence Map (CMAP) to explore the relationships between various competency keywords [407]. In [19], Anca et al. developed several independent classifiers based on traditional machine learning models, such as KNN and LDA, to categorize candidate attributes in areas such as Education (e.g., University and Faculty), Programming Languages (e.g., JavaScript and Python), and Foreign Languages (e.g., English, French, German), thus enabling recruiters to search suitable candidates based on these aspects.

As talent search systems accumulate substantial historical data, researchers attempt to enhance retrieval performance through supervised training. For instance, Ozcaglar et al. proposed a two-level ranking system to integrate structured candidate features by combining Generalized Linear Mixed (GLMix) models and Gradient Boosted Decision Tree (GBDT) models, using recruiter actions as supervised information to learn candidate ranking [311]. Geyik et al. involved recruiters’ immediate feedback on recommendation results to cluster candidates according to recruiters’ intents [139]. Subsequently, they developed a multi-armed bandit-based approach to select the appropriate intent cluster for each recruiter, which was then used to rank the candidates within that cluster. Considering that queries in talent search scenarios can be quite complex, combining several structured fields (such as canonical skills and company names) and unstructured fields (such as free-text and keywords), Ramanath et al. have developed deep learning-based embedding models for both queries and candidates, and they explored learning to rank approaches with DNN models [340]. Yang et al. further proposed a three-stage cascaded ranking model, integrating DNN and BERT, to enhance a deep learning-based talent search system, and accounted for personalized preferences of different users in the final stage [440].

Considering the challenge recruiters face in crafting effective search queries, Ha et al. developed a novel talent search system on the LinkedIn platform that allows users to select ideal candidate examples as queries [161, 162]. They extracted keywords related to titles, skills, and companies from these candidates to reconstruct queries, built training data for a Query-By-Example retrieval system using historical data from a Query-By-Keyword system, and employed the Coordinate Ascent algorithm to optimize the model.

Person-job fitting (PJF) has always been a crucial task in the recruitment process, focusing on measuring the degree of match between a job posting—which includes job duties and requirements—and a candidate’s resume, which details their work and educational experiences. Effective PJF assists employers in selecting the right talent for appropriate positions, thereby enhancing employee performance and job satisfaction. This ultimately contributes to the mutual success of both the organization and its employees. However, the traditional solution relies heavily on recruiters’ domain expertise and subjective judgment, making it challenging to achieve both efficiency and objectivity [334]. Recently, with the accumulation of extensive historical recruitment data through HRIS like ATS, many researchers have leveraged AI technology to develop supervised models for PJF. This advancement has significantly improved the efficiency and effectiveness of the recruitment process [334, 435, 286, 472]. Specifically, the task of Person-Job Fitting can be formally defined as follows:

Early research efforts in the field of PJF using AI-based technologies can be traced back to [273]. Malinowski et al. proposed that the compatibility between a job and a candidate often hinges on underlying factors not explicitly stated in the job posting or the candidate’s resume. To address this issue, they developed a latent variable model to represent job requirements and candidate abilities, and formulated PJF as a bilateral matching problem from a demands-abilities perspective. Since then, several studies have explored the use of AI-based technologies to extract job and candidate profiles from textual data. For instance, Zhu et al. introduced a CNN based neural network to extract representation vectors from job postings and resumes [485]. They then evaluated the fit between a candidate’s qualifications and the job requirements by measuring the similarity between these vectors, achieving an average AUC performance of around 75% on real-world data from 2013 to 2016. In [475], the authors utilized CNN combined with an attention layer to enhance the representation of textual information for both positions and candidates. Qin et al. used LSTM to model the text sequence and applied ability-aware attention strategies to measure the importance of each job requirement or candidate’s ability on the final PJF decision, resulting in an improvement of approximately 10% compared to previous methods [334, 335]. Similarly, in [46] the authors treated job postings and resumes as multi-sentence documents and utilized Bidirectional Gated Recurrent Units (BIGRU) and word-level attention to model sentences and documents. Wang et al. utilized co-attention mechanisms and Graph Neural Networks (GNN) to better leverage historical recruitment data, thereby enhancing the representations of jobs and resumes and subsequently improving PJF performance [409]. Luo et al. expanded upon these efforts by integrating LSTM, CNN, and attention models to process various types of structured textual information, such as skills and experiences [263]. They proposed an adversarial learning-based framework to enhance the expressive representations of this data.

Recently, the rapid advancement of pretrained language models like BERT has significantly enhanced textual data representation, enabling researchers to develop more effective PJF models [226, 2, 250, 360, 208, 123, 152, 147]. For instance, Abdollahnejad et al. serialized concatenated each pair of job $J_{j}$ and resume $R_{i}$ as inputs to BERT, using the representation of the $[\text{CLS}]$ token at the end the sequence to signify the joint representation of the application $S_{i,j}$ [2]. Subsequently, they employed a fully-connected layer to predict the recruitment label $y_{i,j}$, ultimately fine-tuning the BERT-based PJF model with historical interaction data. A similar model structure has also been employed by Kaya et al. [208]. Lavi et al. fine-tuned the BERT model using the Siamese networks and matched jobs and candidates with both classification and similarity-based objectives [226]. Distinct from previous approaches that directly utilized BERT’s architecture, Fang et al. proposed a skill-aware, prompt-based pretraining framework that significantly enhanced the representation of recruitment domain corpora [123]. Compared to BERT and other backbone models, this framework effectively improved performance on several downstream tasks including PJF.

Based on a wealth of high-quality historical data, the aforementioned supervised PJF models are powerful. However, in many recruitment scenarios, job-resume interaction data is sparse and noisy, which significantly undermines the effectiveness of PJF models [46, 45, 450, 73]. To address this, researchers have approached the problem from various perspectives. For instance, in [45], the authors constructed a Job-Resume Relation Graph to enhance the representation of jobs and resumes, where the edges in the graph were created based on historical interactions, category labels, and keywords. Additionally, they designed a Multi-View Co-Teaching Network to learn text-based and relation-based matching modules simultaneously. Bian et al. addressed this issue from the perspective of transfer learning, utilizing potential semantic relations among different job categories and domain adaptation technologies to enhance the performance of PJF models in target domains with sparse interactions [46]. In [450], the authors employed data augmentation techniques based on EDA [410] and ChatGPT [95] to construct additional synthesized job-resume pairs, effectively addressing the issue of data sparsity.

In addition to capturing textual features from application data, various studies have incorporated additional related information to enhance the performance on the PJF task. For instance, it has been noted that certain numerical and categorical attributes of jobs and candidates—such as career level, education, company name, tag, and region—are also crucial and beneficial [197, 169, 168]. Consequently, several neural network-based approaches have been designed to capture the comprehensive interaction of different types of data, such as Factorization Machine [197], CNN [169], and Self-Attention [168]. Furthermore, many researchers have enhanced the performance of PJF models by incorporating knowledge graphs from the recruitment domain as side-information [444, 47, 481]. Moreover, user interaction records on online recruitment services serve as important complementary features to represent both jobs and candidates. Specifically, Yan et al. and Jiang et al. integrated historical interviewed applicants for a job posting and historically applied jobs for a particular talent to complement and enhance the representation learning for jobs and candidates [435, 197]. The hidden idea is that historical interview choices and job applicants reveal the preference of jobs and candidates for each other. Meanwhile, in [178], the authors leveraged job seekers’ click behaviors and search histories on recruitment websites to better model their job-seeking intentions, thereby enhancing the performance of the PJF task. Inspired by existing multi-behavior recommendation methods, Saito et al. utilized job seekers’ interactions—including viewing, favoriting, and applying—to implement a multi-behavior job recommendation system [348, 349]. In [128], the authors integrated the supply-demand situation of various companies at specific times as supplementary labor market information into the PJF model. They utilized hierarchical reinforcement learning to address the dynamic components of the PJF task, particularly by incorporating the temporal factors of the application process.

Furthermore, some online recruitment systems offer recommendation services to both employers and job seekers. In response, researchers have developed joint models to address these dual recommendation scenarios [437, 129, 181, 478]. For instance, Yang et al. proposed a unified dual-perspective interaction graph that incorporates two distinct nodes for each candidate (or job) to characterize both successful and failed matching [437]. Subsequently, the authors employed BERT and GCNs to construct the PJF model. It was trained using a quadruple-based loss and a dual-perspective contrastive learning loss to enable bidirectional recommendation. In [129], the authors investigated the dynamic preferences of users, including browsing, clicking, and online chat behaviors, within dual recommendation scenarios. They introduced a Bilateral Cascade Multi-Task Learning framework to implement dynamic PJF effectively.

Job interview aims to assess the fitness of candidates and the job positions by evaluating their skills and experiences. A critical aspect of this process is designing appropriate questions to comprehensively assess the competencies of potential employees. In this phase, personalized question recommendation has emerged as a feasible approach, that selects the right questions from the question set based on the candidate and the job position. Along this line, Qin et al. proposed to recommend personalized question sets for various applications, taking into consideration both job requirements and candidates’ experiences [336]. In particular, to enhance the performance of the recommendation system, a knowledge graph of job skills was built using query logs from the biggest search engine, Baidu. Datta et al. further took into account the difficulty of interview questions and designed an interview assistant system using knowledge graphs and Integer Linear Programming techniques [102]. To cope with the substantial risk of bias arising from the subjective nature of traditional in-person interviews, Shen et al. proposed to learn the representative perspectives of in-person interviews from the successful job interview records in history [364]. With the help of topic models, they represented job postings, resumes, and interview assessment reports in an interpretable way. The potential relationships among them are also mined to recommend questions or skills that should be estimated during interviews [364, 363]. Substantially, Shi et al. proposed an automated system for recommending personalized screening questionnaires based on job postings [365]. They encoded the job posting with the BERT model, selected the question templates with a Multi-Layer Perception (MLP) Classifier, and extracted necessary parameters from the templates using feature-based regression models. In [333], the authors introduced a scalable, skill-oriented interview question generation model that leverages external knowledge from an online Knowledge-Sharing Community (KSC). Additionally, they developed a GNN-based interview question recommendation system that tailors interview questions to user queries.

Assessment scoring is another critical problem in talent management, evaluating the competency of candidates or employees based on their performance. Typically, this problem can be formulated as a binary classification problem as follows:

Substantial effort has been invested during the interviewing phase [303, 78, 80, 172, 171, 77, 174, 264]. Within this context, a major area of research involves using recorded face-to-face job interviews or simulated asynchronous video interview (AVI) data to assess candidates’ competencies in various aspects. For instance, Nguyen et al. were the first to focus on the automated prediction of employment interview outcomes based on both audio and visual nonverbal cues from the interviewee and interviewer [307]. Naim et al. extracted 82 features from 138 recorded interview videos, covering three dimensions: prosodic, lexical, and facial information [303, 304]. They employed regression models such as Support Vector Regression (SVR) and Lasso to predict overall interview scores. Chen et al. focused on monologue-style responses to Structured Interview (SI) questions and constructed an AVI dataset [78, 80]. Subsequently, they utilized the Linguistic Inquiry Word Count (LIWC) [324] to extract lexical features from text-based interview content, used automated speech analysis and transcription [79, 456] to generate features assessing various dimensions of speaking, such as fluency, rhythm, intonation & stress, and pronunciation, and extracted visual features related to facial expressions using Visage SDK’s FaceTrack and FACET SDK. Finally, several shallow classification models, including Support Vector Machine (SVM) and Random Forest (RF), were employed to predict interviewee traits such as agreeableness, conscientiousness, emotional stability, extraversion, and openness. Hemamou et al. collected a corpus of over 7,000 candidates who participated in asynchronous video job interviews for real positions, recording themselves answering a set of questions [172]. Utilizing this extensive dataset, they designed a hierarchical attention model to predict candidates’ hireability. Additionally, to pinpoint the relevant segments of an answer, the authors developed attention mechanisms that extract fine-grained temporal information from the responses [171].

With the European Union’s approval of the world’s first AI technology regulation, researchers are now reconsidering the ethical risks associated with AI-based assessment scoring methods used in interview processes. For instance, unlike previous work that utilized multimodal information from interview recordings, Chen et al. solely employed automatic speech recognition transcriptions from AVI data [77]. They introduced GNNs to construct dependency relations between questions, enabling the learning of a model that scores multiple question-answer pairs. Besides, Singhania et al. were the first to investigate fairness concerning gender and race in video interviews [372]. Moreover, in [173], the authors proposed an approach using adversarial methods to remove sensitive information from the latent representations of neural networks, which was applied to interview assessments to promote fairness in job selection.

In addition to the interviewing phase, there is still a focus on predicting competency based on employee profile information. For instance, Liu et al. and Hong et al. applied the SVM model to predict the competence level of civil servants and highway construction foremen, respectively [252, 177]. Furthermore, Li et al. evaluated the performance of traditional classifiers, including SVM, RF, and Adaboost, in competency evaluation and found that prediction results based solely on structured personal static data were suboptimal [234]. For improved performance, incorporating more unstructured and dynamic data, such as textual data or social networks, may represent a significant direction for future research.

• •

  Advantages of AI technology: (1) Existing AI-based assessment scoring models offer an automated and objective method for evaluating candidates’ abilities in professional skills, social skills, and other overall competencies. (2) AI-based assessment scoring models can analyze candidates’ performances in simulated interview scenarios, aiding them in honing their interview skills and increasing their chances of success in actual job interviews.

• •

  Limitations and future directions: (1) Although some researchers are increasingly aware of the potential ethical risks associated with the use of data and models in AI-based assessment scoring, there is still a lack of comprehensive research in this area. (2) Most existing research focuses on training supervised models trained with original scores from interviewers or managers for automated scoring predictions. However, these studies often overlook the inherent biases embedded within the data. Therefore, further research is needed on how to train unbiased assessment scoring models from biased data. (3) Cognitive diagnosis has been widely validated in computational education for learning the knowledge profiles of students and predicting their future exercise performance [474, 451]. However, research concerning the application of these techniques to competency diagnostics in talent assessment contexts, such as interviews, remains limited.

The organization offers a diverse array of training programs, including technical, project management, quality, leadership, specialized, and soft skills training. Hence, it is crucial to assist employees in selecting the training courses that align closely with their background and objectives [377]. Course Recommendation aims to provide personalized courses based on the different preferences and needs of users in various aspects. Much research in this direction is concerned with student education [317, 118, 247, 110, 97, 491]. Recently, some effort has also been posed in the talent management field [377, 402, 401, 320, 480, 439]. There is a growing focus on the advancement of enterprise Learning Management Systems (LMS), designed to enhance both individual and organizational performance by customized online training programs aimed at augmenting employees’ skills and knowledge [480]. Typically, we formalize this problem as follows:

As a recommendation task, most of the methods in the recommendation system can be utilized to conduct course recommendations. However, unlike traditional item recommendation, where the decision is determined by the users’ rating, more attention should be paid to mining employees’ competencies and their needs for further development from side information such as the employee’s profiles. Therefore, Wang et al. used a topic model to extract the latent interpretable representations of the employee’s current competencies from their skill profiles, as well as a recognition mechanism to explore the personal demands from learning records [402]. Then, they integrated the collaborative filtering algorithm with the Variational AutoEncoder (VAE) to develop an explainable course recommendation system. In addition to learning records, Srivastava et al. also introduced work history into the learning content recommendation and defined a Markov Decision Process (MDP) to extract the past training patterns [377]. Moreover, Sharvesh et al. proposed an adaptive recommendation model that incorporates employees’ styles and goals to suggest the most suitable courses, considering their interactions and performance [320]. Zheng et al. proposed a novel generative recommendation framework called Generative Learning for Adaptive Recommendations (GLAD) [480], which integrates reinforcement learning. It concurrently considers enhancing employee performance and ensuring the rationality of generated recommendations. Yang et al. introduced a contextualized knowledge graph embedding to recommend training courses to the talent in an explainable manner [439].

• •

  Advantages of AI technology: (1) The AI-based course recommendation system integrates employee and course profiles to provide tailored course recommendations, thereby enhancing learning outcomes for employees. (2) Several AI-based course recommendation systems not only offer personalized course suggestions to employees but also offer suitable justifications for these recommendations, thus facilitating employees’ decision-making in selecting training courses. (3) The AI-based course recommendation system recommends courses to employees and also assesses their competency level, thereby facilitating their subsequent career development.

• •

  Limitations and future directions: (1) While many AI-based course recommendation models prioritize courses based on employee preferences, these preferences may not always align with the employee’s career development needs. As a result, the recommended courses may not effectively improve work performance [480]. (2) While existing AI-based course recommendation systems offer personalized recommendations to employees, they often fail to consider the complex motivations driving employees’ course selections [439]. Further research is needed to understand the diverse motivations behind employees’ course choices. (3) Most current course recommendation systems prioritize student course recommendations, leaving limited research on recommending courses for employee training. As employee training advances, there is a pressing need for further research on employee course recommendation systems.

Promotions serve two essential roles in the organization, that is, assigning individuals to the jobs for which they are best suited and providing incentives for lower-level employees  [33]. Traditionally, promotion is decided by only managers. Promotion prediction is mainly used to better judge the development potential of employees, so as to better work allocation efficiency and better output. This prediction or judgment is generally decided through questionnaires, which is labor-intensive. To this end, AI-based research is concerned with identifying the features that are correlated with promotion and applying machine learning methods to predict promotion. Typically, the problem can be formulated as a classification task:

In this phase, some basic machine learning algorithms have been adopted to solve this problem like KNN, SVM, random forest, Adaboost and so on [189, 192]. Generally, the majority of machine learning methods achieve an accuracy of over 90% in promotion prediction. At the same time, to solve the data imbalance issue in promotion data, various sampling methods have been adopted like random oversampling, under-sampling, or hybrid sampling combining multiple sampling methods [347, 358]. With increasingly complex features introduced, Yuan et al. suggested that work-related interactions and online social connections are strongly predictive and correlate with promotion and resignation [453]. Along this line, Long et al. used Random Forest to predict promotions based on demographic and job features [259]. However, these methods may not fully capture the complexity and dynamism of career development. To address this, Li et al. proposed a novel survival analysis approach, where predicting promotion events is transformed into estimating the expected duration of time until promotion occurs [233]. Liu et al. defined the development potential with various dimensions like the number of expertise [253]. Through dimensions distribution and clustering algorithms, the model could divide development status into four categories so that promotion prediction results could be generated. Guarino et al. designed a method for optimizing the growth path with Deep Q-Learning [153]. In this case, experience results, which are the reward value, are evaluated by the competences development. Through this system, employees’ development potential is judged so that we can confirm whether this employee should be promoted.

In the stable development of any organization, employee turnover stands as a critical influencing factor. Whenever a fully integrated employee departs from the organization for any reason, it leads to varying degrees of structural incompleteness, thereby resulting in a decline in productivity [1]. However, early identification of employee turnover tendencies, combined with strategic preparation, can substantially alleviate the productivity losses linked to employee turnover. In recent years, an expanding body of literature on talent management and organizational behavior has emerged, with a plethora of machine learning methods being employed to predict employee turnover, yielding remarkable results [146, 7]. Mathematically, the turnover prediction problem can be formulated as follows:

Many data mining techniques have been introduced to investigate employees turnover prediction [10, 373, 477, 133, 144]. For instance, Nagadevara et al. used five data mining techniques to predict turnover, including MLP, Logistic Regression, Decision Tree, etc [302]. The results reveal that absenteeism and lateness, job content, demographics, and experience in the current team are strong predictors of turnover. Considering the potential noise introduced by a multitude of employee characteristics, some studies employ techniques such as feature filtering algorithms [432, 200] and feature weighting [452, 72] to eliminate or reduce unnecessary information. Meanwhile, oversampling techniques such as synthetic minority over-sampling technique, adaptive synthetic, and borderline synthetic minority oversampling technique have been validated to effectively address the issue of data imbalance in turnover prediction [266, 214]. Additionally, some scholars classify employees before conducting turnover prediction to effectively enhance predictive accuracy by accounting for inter-individual variability [193, 416]. Tailoring feature modeling to the specific characteristics of particular industries has also yielded favorable outcomes in turnover prediction, spanning industries such as apparel [344], IT [38, 292] and hospitality [203].

Besides predicting turnover based on static features, considerable attention is invested in capturing the dynamic factors, especially the evolving neural network-based methods, which have demonstrated significant expressive ability [233]. In this phase, Teng et al. investigated the contagious effect of employee turnover on an individual and organizational level, respectively [387, 386]. Specifically, they developed two LSTM cells to process peers’ turnover sequence and environmental change, as well as a global attention mechanism to evaluate the heterogeneous impact on potential turnover behavior. The experiments conducted on the dataset provided by a high-tech company in China demonstrate the effectiveness of their proposed framework, including profile information and turnover records. Similarly, the authors in  [465] utilized vast career trajectory data to build a heterogeneous company job network, integrating macro-level company position information into individual career transition prediction using the Dual-GRU model. On this basis, Hang et al. further modeled employee turnover from both internal and external views [165]. For the internal component, they captured the influence of close collaborators and colleagues with similar skills using a graph convolutional network. From the external-market view, they connected employees and external job markets through shared job skills. Finally, both internal and external information is fed into BiLSTM and survival analysis for turnover predictions. Furthermore, Subha et al. employed principal component analysis for feature extraction, followed by turnover prediction using CRF-BiLSTM-CNN, yielding significantly improved performance [380].

For instance, Arambepola et al. explored the influence of job-specific factors on job satisfaction level by combining both the employee’s background data and company-related factors, where several classifiers, including Random Forest, Logistic Regression, and SVM, were used for the prediction of the job satisfaction level of software developers [22]. Saha et al. proposed to assess job satisfaction by leveraging large-scale social media, i.e., employees’ Twitter post dataset, where word frequency statistics, lexical analysis, and sentiment analysis have been conducted to extract features from textual data [346]. Accordingly, multiple classifiers, such as SVM and MLP, were used to predict employees’ job satisfaction. Furthermore, Saleh et al. utilized an online questionnaire for data collection and employed the artificial neural network and decision tree for predicting employee job satisfaction [350]. In addition to the aforementioned classification models, Mcjames et al. developed a causal inference machine learning approach to identify practical interventions for improving job satisfaction [284]. This approach was based on the TALIS 2018 dataset, which provides a representative sample of school teachers. Devi et al. achieved improved prediction of employee job satisfaction by employing a weighted averaging ensemble method with logistic regression [112].

In the rapidly evolving labor market, the deployment of AI for career mobility prediction significantly enhances personalized career planning. By analyzing extensive datasets on career trajectories in the labor market, AI-based methods can provide tailored recommendations that align with future employment opportunities. Following [465, 433], we define the career mobility prediction problem as follows:

Generally, as illustrated in Figure 7, the task of career mobility prediction can be addressed through time series analysis, where the career path (also known as career trajectory) is treated as a sequence of events [236, 289, 433]. Most existing work focuses on predicting next-step career mobility, i.e., $K=1$. Specifically, Li et al. were the first to design a contextual LSTM model that integrates the profile context and career path dynamics simultaneously to predict the next company/position of talents [236]. To provide a fine-grained prediction, researchers have developed several methods for modeling career trajectories that predict both the next employer and the corresponding job duration. Meng et al. employed a hierarchical neural network structure with an embedded attention mechanism to characterize internal and external job mobility [289]. Moreover, Wang et al. introduced a temporal encoding mechanism that handles dynamic temporal information [400]. Macro-level job transition behavior may also impact individual career choices. To address this, Zhang et al. constructed a heterogeneous company-position network based on massive career trajectory data and integrated macro information from the company-position into personal career move prediction [465]. Meanwhile, He et al. expanded the prediction tasks to include forecasting salary levels and company sizes, utilizing LSTM and CNN to construct the model [170]. Recently, several studies have focused on employing pre-training techniques to model career trajectory representations, aiming to enhance performance in career mobility prediction [108, 459]. For instance, Decorte et al. developed CareerBERT, a model built upon the BERT framework, to analyze sequences of work experiences—including job titles and descriptions—alongside corresponding ESCO occupation sequences within career trajectory data [108]. They utilized contrastive learning to fine-tune this pre-trained model. Following this, the authors introduced a map** network designed to predict future ESCO occupations based on sequences of work experiences.

Unlike previous work that focused solely on predicting the “immediate” next career move, Yamashita et al. utilized the transformer architecture to predict one’s future career pathway as a “sequence”—specifically, the next $K$ steps of career movement [433]. Furthermore, some researchers have incorporated various trajectory rewards, such as company ratings, staying probabilities, and salary ranges, into the prediction of future career paths [154, 155, 26]. For instance, Guo et al. propose an intelligent sequential career planning system via stochastic subsampling reinforcement learning, which is capable of finding globally optimal career paths for talents [154].

In the real scenario, abundant talent data can be utilized to construct an organizational network, unveiling the intricate relationships among employees as they form project teams or forge alliances across different groups. An illustration of an organizational network is shown in Figure 8. However, extracting useful information to construct the network and obtaining meaningful knowledge to support managerial decisions are still difficult. Traditionally, organizational network is usually constructed by questionnaires like formal or informal relationships, which are labor-intensive and subjective. To improve the accuracy and efficiency of organizational network construction, many researchers are focusing on using network embedding techniques to model the organizational network and capture knowledge from this [326, 150]. Without loss of generality, the generalized organization network can be defined as follows:

The network embedding describes the network as low-dimension vectors and further serves the downstream applications. For instance, Ye et al. proposed a multiplex attentive network embedding approach for modeling organizational networks in a holistic way [446]. In their work, the organizational network is composed of multiple communication interactions among employees. They generated embedding for employees based on the random walk strategy [326] with k-core and approximated shortest path algorithm. Furthermore, they proposed a relational transition-based approach to represent each department. In this way, the learned representation can be leveraged for several talent management tasks, including employee turnover, performance prediction, and department performance prediction. Besides, Teng et al. exploited the network fusion technique for organizational turnover prediction [386]. They concentrated on modeling the relationships among organizations. Specifically, they demonstrated the correlation between the topology of organizational networks and organizational turnover. To this end, they constructed a turnover similarity network based on the multiple organizational social networks, and took advantage of the GNN to learn comprehensive knowledge from these topological structures, which were further used for organizational turnover prediction. Apart from that, there are also some other applications. For example, Dong et al. studied the problem of cross-group community search on the labeled graph, namely Butterfly-Core Community (BCC) search [114]. Specifically, they demonstrated the BCC problem is NP-hard, and proposed an approximated algorithm to solve it. The proposed algorithm was evaluated on a real-world organizational network from Baidu Inc, which can effectively find communities formed by cross-group collaborations given two employees with different positions.

High-potential talents (HIPOs) possess leadership abilities, business acumen, and a strong drive for success, making them more likely to emerge as future leaders within organizations when compared to their peers [231]. Identifying HIPOs has always been a major issue in human resource management, it plays significant roles in the execution of organizational strategy and the optimization of organizational structure [371]. At the same time, they are strategic assets for companies to achieve sustainable competitive advantages [457]. Due to their strategic importance, high-potential talent identification and retention are considered as the most critical component of business strategy.

The traditional methods for HIPOs identification usually rely on subjective selection of HR experts. They primarily focus on evaluating certain talent factors, such as communication skills, teamwork, and self-learning [325, 137]. However, these manually selected factors may lead to unintentional bias and inconsistencies [313]. Recently, with the development of ONA, objective data-driven HIPOs identification has become possible. The rationale behind this is that HIPOs usually perform more actively and have higher competencies than their peers to accumulate their social capital during their daily work [447]. We can detect HIPOs implicitly through social information. Formally, the HIPOs identification problem based on the organizational network can be formulated as:

To solve this problem, Ye et al. proposed a neural network-based dynamic social profiling approach for quantitative identification of HIPOs, which focuses on modeling the dynamics of employees’ behaviors within the organizational network [447]. In particular, they applied GCN and social centrality analysis to extract both local and global information in the organizational network as social profiles for each employee. Then they adopt LSTM with a global attention mechanism to capture the profile dynamics of employees during their early careers. Finally, they evaluated their model on real-world talent data, which clearly validates the effectiveness and interpretability of the proposed model. This method combines longitudinal social network data to embed employees’ work performance during different periods, reflecting the employees’ development and growth over time like social capital increasing. Furthermore, there are another network construction methods using the cooperation experience. For example, as for high-potential scholars identification, Yin et al. constructed the innovation potential with the disruption of technology and science of each paper, then used the co-authorship network in each year to model the development of scholars. Combining social capital theory and LSTM, they measured social capitals in different years to predict future innovative potential [448].

Team formation is an important issue in organizational management. On the one hand, team formation is effective for responding to urgent tasks, on the other hand, team formation mainly could help improve the efficiency of the entire organization. In brief, team formation could be defined as discovering a team of experts that collectively cover all the required skills for a given project, as shown in Figure 9. Whereas it is proven to be NP-hard [224], this requirement still needs to be solved in many real-world scenarios, such as team discovery in a social network, which contains professionals who provide specialized skills or services.

Definition 4.2 (Team Formation). Given a team collaboration in the $t$-th time slice as $\mathcal{C}_{t}=\left\{(s,e)_{t}\mid s\subseteq\mathcal{S},e\subseteq\mathcal{E}\right\}$, where $\mathcal{S}$ and $\mathcal{E}$ represent the sets of skills and experts. The target of Team Formation is to minimize costs in teamwork (e.g., communication costs) or to succeed in their goals.

For instance, Kargar et al. proposed a method to find the object team with minimal communication cost as well as personnel cost of the project [205]. Specifically, they used a graph to model a social network where nodes represent experts and formulate the task as a constrained bi-criteria optimization problem. Since it is proved that the problem of minimizing the combined cost function is still NP-hard, the authors efficiently solve the problem with an approximation algorithm and three heuristic algorithms in polynomial time. Later, Zihayat et al. took both communication cost and experts’ authority into account and proposed greedy algorithms to solve the optimization problem [489]. Since these team formation algorithms are based on very different criteria and performance metrics, Wang et al. implemented these algorithms using a common platform and evaluated their performance with several real datasets [405]. However, these studies have limitations in terms of scalability and fail to effectively manage the dynamic nature of expert networks. To this end, instead of searching over the graph representation of the expert network, Hamidi et al. searched for variational distributions of experts and skills in the context of a team [164]. To be specific, they employed a variational Bayesian neural network to form the optimal team, contributing to a better performance than prior state-of-the-art.

In many scenarios, the purpose of team formation is to anticipate future expert teams by combining sequences of expert skills and the evolution of collaborative relationships over time. Hamidi et al. defined an optimal team as a team observed in the past, and proposed a variational Bayesian neural network to estimate the map** function from the skill power set to the expert power set, finally forming an optimal expert team e for a required subset s of skills [163]. Fani et al. presented a streaming-based neural model training strategy to estimate the map** function from the stream of the collaborative set and the skill subset to the expert subset to find the best possible team [124]. When recommending future teams by learning the expert and skill distributions of past successful teams, it may happen that a few experts perform most of the successful collaborations while the majority of experts rarely participate. The long-tailed distribution of the training data tends to lead to the overfitting of the model, to address the problem, Dashti et al. proposed three negative sampling heuristics [100].

Another scene that draws wide attention is the fast-growing Online Labor Marketplaces, which provide a sharp decrease in communication costs. For example, Liu et al. first implemented team formation in crowdsourcing markets with consideration of the impact of teamwork [256]. The study designs a mechanism that combines the greedy selection rule and a special payment scheme, obtaining various desirable properties, such as efficiency, profitability, and truthfulness. In addition, Barnabo et al. considered the fairness of algorithms related to these online marketplaces [40]. They formalized the Fair Team Formation as the problem of finding the cheapest team that can complete the task and, at the same time, that counts the same number of people from two not overlap** classes. Consequently, four algorithms are designed to solve the problem and experiment on real-world data to confirm their effectiveness. Yet, most of the works focus on the offline version of the team formation problem, i.e., the tasks to be completed are a-priori known. To this end, Anagnostopoulos et al. implemented the problem of online cost minimization, where the goal is to minimize the overall cost (paid on hiring, outsourcing, and salary costs) of maintaining a team that can complete the arriving tasks [18]. Moreover, the study considers a more complex case of outsourcing, i.e., hiring, firing, and outsourcing decisions can be taken by an online algorithm leading to cost savings with respect to alternatives.

Team optimization means the optimization of existing teams, which is beneficial for responding to changes in team tasks. Indeed, there are two ways to optimize teams, including team member replacement and team expansion.

Definition 4.3 (Team Optimization). For a team $T$, the purpose of team optimization is to recommend the group of candidate members $i$ based on their fit with the existing team $T$ when the team is expanding or to recommend the group of new members (sub-teams) $i$ to replace departing members (sub-teams) $i^{\prime}$ in the team when members are replacing them using the GNN approach.

Two key problems within the scope of team optimization are team member replacement and team expansion, as illustrated in Figure 10. Specifically, team member replacement is put forward first by [237], which aims to find a good candidate to best replace a team member who becomes unavailable to perform the task. To tackle this problem, they introduced the concept of graph kernels that takes into account the interaction of both skill and structure matching requirements. Furthermore, the study proposes a series of effective and scalable algorithms for this problem. In [242], the authors further took the synergy between skill similarity and structural similarity into consideration, instead of considering the two aspects independently. Later, [241] proposed a new graph kernel that evaluates the superiority of candidate subteams in a holistic way, which can be freely adapted to the needs of the situation. An effective pruning strategy is used in the algorithm to reduce the kernel computation by exploiting the similarity between the structures of the candidate teams, which can output more human-agreeable recommendations compared to previous studies. What’s more, [180] used a clustering-based GNN framework to capture team network knowledge for flexible subteam replacement; and incorporated a self-supervised positive team comparison training scheme into the model for improved team-level representation learning and unsupervised node clustering to reduce candidate objects for fast computation.

In addition, some effort has been paid on team expansion. Zhao et al. formally defined the problem in collaborative environments and proposed a neural network-based approach, considering three important factors (team task, existing team members, and candidate team member) as well as their interactions simultaneously [476]. However, most works on team optimization treat teams as a static system and recommend a single action to optimize a short-term objective. To this end, Zhou et al. proposed a deep reinforcement learning-based framework to continuously learn and update its team optimization strategy by incorporating both skill similarity and structural consistency [482].

Person-organization fit (P-O fit) refers to the compatibility between employees and their organizations. In fact, P-O fit has been widely recognized as an effective indicator of proactive talent management, and it has a significant impact on outcomes such as work attitudes, turnover intentions, and job performance [225]. In the domain of organizational behavior, most studies measure P-O fit based on the similarity between organizational profile and employees’ profile. Figure 11(a) presents the classical P-O fit modeling process. At first, experts collect information with questionnaires to extract employee and organization profiles and manually design metrics. Then, the statistical methods are applied to measure the congruence between an employee and an organization as P-O fit score. However, this process is labor-intensive and subjective, which is difficult to apply to real-world applications. To this end, AI-driven techniques are proposed to automatically extract the profiles and model P-O fit in a dynamic, quantitative, and objective manner. More specifically, the P-O fit problem is defined as follows:

To solve this problem, Sun et al. proposed a new P-O fit modeling process based on AI technology [381], as shown in Figure 11(b). Specifically, they first extracted features automatically from collected employees’ in-firm data and generated person profiles by dimension reduction on these features. Then, they exploited the organization profile by combining the organization’s structure with the profiles of the employees and extracted a unique environment profile for each employee based on their corresponding positions. Finally, they applied a deep neural network to achieve a more complicated map** from person and environment profiles to a P-O fit representation. To capture the dynamic nature of P-O fit and its consequent impact, they exploited an adapted Recurrent Neural Network with an attention mechanism to model the temporal information of P-O fit. Later, in [383], Sun et al. further proposed the attentional features extraction layers that can distinguish individualized relation-level and individual-level influence differences for different nodes on the organizational tree. This largely enhanced the performance of person-organization compatibility modeling and improved the interpretability. Combining the person-organization fit theory, Artar et al. used K-Nearest Neighbors algorithm to cluster the employees through P-O fit representations like the number of former employers, and the number of years in the organization [24].

Job title benchmarking (JTB), as an important function in C&B, aims at matching job titles with similar expertise levels across various organizations (i.e., companies), which provides precise and substantial facilitation of job and salary calibration/forecasting for both talent recruitment and job seekers. Traditional JTB mainly relies on manual market surveys, which are expensive and labor-intensive. Recently, the popularity of online professional networks helps to accumulate massive career records, which provides the opportunity for a data-driven solution. Formally, JTB can be defined as follows:

To handle this problem, Zhang et al. proposed to construct a Job-Graph by extracting information from large-scale career trajectory data, where nodes represent job titles affiliated with the specific companies and edges represent the numbers of transitions between job titles [463]. They redefined JTB as a link prediction task on the Job-Graph by assuming that the benchmarked job title pairs should have a strong correlation with the link. Along this line, they proposed a collective multi-view representation learning model to represent job titles from multiple views, including graph topology view, semantic view, job transition balance view, and job transition duration view. Subsequently, they devised a fusion strategy to generate a unified representation from multi-view representation. Finally, they leveraged the similarity between these representations as an indicator for job title benchmarking. Following this line, subsequent studies have refined the job transition network. Specifically, Zhu and Hudelot [486] proposed enhanced job transition network with the bi-directional edges when job nodes have the same tag, for example,“automotive shop manager” and “purchasing manager” have the same tag named “manager”. At the same time, bi-directional edges between a job title and a tag, representing the “has/in” relationship, have been added into the job transition network. Furthermore, this research uses an open dataset from a Kaggle competition containing a collection of working experiences.

There is another way to solve the problem of job title benchmarking, which is the way of embedding. In these researches, job title generally is embedded with the auxiliary information like skill, employees and so on. Zbib et al. embedded the auxiliary skills under job titles, through the feedback based on the similarity calculation with the text encoding of the job title itself, the reasonable skill embedding under the job title can be learned and different job titles can be integrated [455]. Finally, the effect is verified by skill-based retrieval and text-based retrieval. Malandri et al. used the job contents under job title and generated embedding of job titles to cluster similar jobs [272]. Besides, Zha et al. aggregated job titles by decomposing the semantics of different modules of job titles [458].

The information provided by the job transition graph and embedding can also be combined to perform job title benchmarking. For example, JAMES proposed by Yamashita et al. combined the hyperbolic graph embedding generated from the job transition records, BERT embedding of original job titles, and syntactic embedding generated by the average similarity of original job titles and occupations in job taxonomy like European Skills/Competences, qualifications, and Occupations (ESCO) taxonomy [434]. Liu and Ge jointly trained the job embedding and employee embedding, which used the job contents. Specifically, the job context learning module is designed to use the job embedding to generate employee embedding, using the job transition information of the employee, where the observed context of positions are positive and a sample of negative positions through negative sampling is negative [250].

Job salary benchmarking (JSB) refers to the process by which organizations obtain and analyze labor market data to determine appropriate compensation for their existing and potential employees [48]. Traditional approaches for JSB mainly rely on the experience from domain experts and market surveys provided by third-party consulting companies or governmental organizations [199, 329]. However, fast-develo** technology and industrial structure lead to changes in positions and job requirements, making it difficult to conduct salary benchmarking in a timely manner in dynamic scenarios.

In recent years, the prevalence of emerging online recruiting services, such as Indeed and Lagou, has provided the opportunity to accumulate vast amounts of job-related data from a wide range of companies, enabling a new paradigm of compensation benchmarking in a data-driven manner. Formally, the job salary benchmarking problem can be formulated as follows:

To address this problem, some scholars used the statistic machine learning methods to combine the company characteristics to predict the average salary of industries or economic activities [279]. Then, from the job-company salary matrix view, where each entry indicates the corresponding salary of a given job-company pair, the JSB problem can be regarded as a matrix completion task. Generally, matrix factorization (MF) is a widely used method for handling this task. It aims to factorize an incomplete job-company salary matrix into two lower-rank latent matrices, and use their dot product for estimating the possible salary of the missing entries. However, the intuitive method is too general to meet the various special needs of C&B professionals. To this end, Meng et al. proposed an expanded salary matrix by expanding the original job-company salary matrix with locations and time information for the fine-grained salary benchmarking [288]. Then they designed a matrix factorization-based model for predicting the missing salary information in the expanded salary matrix by integrating multiple confounding factors, including company similarity, job similarity, and spatial-temporal similarity. Further, Meng et al. designed a nonparametric Dirichlet-process-based latent factor model for JSB, which learns representation for companies and positions to alleviate the data deficiency problem. By conducting experiments on two large-scale real-world data, the effectiveness and interpretability of the proposed model have been proved [287]. Similarly, the matrix equation method is used in [186] to minimize the unbiased salary, company competitiveness in the salary of the same job, and inflation of the same job. Following another way of thought, some scholars proposed to construct the auxiliary network of skill requirements under job postings to model the relationship between the job and salary. Specifically, the auxiliary network can learn the representation of jobs through graph learning, so as to achieve similar jobs with similar salary learning effects, so as to effectively predict the salary of different jobs [382].

The task of talent flow prediction mainly revolves around anticipating changes in the labor market, thereby offering guidance for talent strategies. AI-driven techniques can enhance the accuracy and flexibility of such predictions. Formally, the talent flow prediction problem can be defined as follows:

To solve this problem, Zhang et al. designed a dynamic latent factor-based Evolving Tensor Factorization (ETF) model for predicting future talent flows [466]. In detail, they used $U^{t}_{i}$, $V^{t}_{j}$, $W^{t}_{k}$ to represent the latent vectors of origin company $i$, destination company $j$ and job position $k$ at time slice $t$, and evolve them to time slice $t+1$ for predicting talent flows at $t+1$. This model also integrates several representative attributes of companies as side information for regulating the model inference. The authors also proposed a Talent Flow Embedding (TFE) model to learn the bi-directional talent attractions of each company [464]. Subsequently, they explored the competition between different companies by analyzing talent flows using data from OPNs. In detail, the objective of this latent variable model is to learn two attraction vectors $S_{u}$ and $T_{u}$ from talent flow network $G$, where $S_{u}$ is the source attraction vector of company $u$ and $T_{u}$ is the target attraction vector of company $u$. The pair dot of $S_{u}$ and $T_{v}$ indicates the talent flow from company $u$ to company $v$. The experimental results show the pairwise competitive relationships between different companies. Xu et al. enriched the sparse talent flow data by exploiting the correlations between the stock price movement and the talent flows of public companies [430]. They developed a fine-grained data-driven RNN model to capture the dynamics and evolving nature of talent flows, utilizing the rich information available in job transition networks. In addition to leveraging OPNs data, Liu et al. transformed questionnaire data into graph data and achieved accurate talent flow prediction using the Graph Attention Network [249]. This model incorporates an attention mechanism to mitigate information overload and decrease the network’s reliance on temporal and spatial factors.

Many researchers also explored other various flow pattern analysis tasks, such as competitiveness analysis hop** behavior analysis, and talent circle detection. In [86], the author initially gathered job-related information from various social media data. Subsequently, they developed a model called JobMiner, which focuses mainly on employing graph mining techniques to mine influential companies and uncover talent flow patterns. This method provides a better understanding of company competition and talent flow in professional social networks. Yu Cheng further developed machine learning and analytical techniques for the purpose of mining OPNs data [85]. From OPNs, they can mine influential companies with the related company groups and evaluate the company’s influence and competitiveness. Oentaryo et al. developed a series of data mining methodologies to analyze job-hop** behavior between different jobs and companies using publicly available OPNs data [309]. In detail, they used a weighted version of PageRank to measure the competitiveness of jobs or companies, and construct some metrics to measure the relationship between many properties of jobs and the propensity of hop**. Li et al. undertook the quantification of the dynamic attractiveness of companies to various types of talent [238]. Specifically, they utilized job title clustering to identify prevalent talent categories and then integrated the PageRank algorithm with job titles and talent flow graphs to determine companies’ attractiveness to specific talent categories over time. Then, Oentaryo et al. enhanced the data mining framework for analyzing talent flow patterns [310]. The results show that the factors influencing employee turnover can mainly be divided into four categories: employee personal factors, organizational factors, external environmental factors, and structural factors. Xu et al. developed a talent circle detection model and designed the corresponding learning method maximizing the NDCG to detect a suitable circle structure [425]. Each talent circle includes organizations with similar talent exchange patterns. Formally, a talent circle is a subset of neighbors of an ego node. In one circle, nodes are closely connected and similar to each other. The circles can be denoted as $\{C_{m}\in\mathbb{C}\}$, where $m=1,2,...,M$ and $C_{m}\subseteq V$. $V$ represents the set of all organizations. The circles can be overlapped and the appropriate talent circles mean similar flow patterns and are closely connected. The detected talent circle can be used to predict talent exchange in the future and improve the recommendation in talent recruitment.

Forecasting talent demand is pivotal for both job seekers and employers as it offers a crucial insight into economic trends. Leveraging AI-driven techniques can significantly boost the precision and adaptability of these forecasts. Formally, the talent trend forecasting problem can be defined as follows.

The demand trend mainly focuses on the number of job recruitment, some researches used classifier method SVM and decomposition methods STL (Seasonal and Trend decomposition using Loess) to analyze the demand time series characteristics from the web data and the official data [261]. These researches show that web data can reflect the trend of the labor market. Karakatsanis et al. suggested a data mining-based approach for identifying the most in-demand occupations in the modern job market [204]. In detail, a Latent Semantic Indexing (LSI) model was developed for online job posts with job description data. The analysis results can highlight job trends most in-demand and identify occupational clusters. Zhang et al. provided Talent Demand Attention Network (TDAN), which can forecast fine-grained talent demand in the labor market [467]. Specifically, they constructed multiple-grained levels of information (e.g., market level, company level, job level et al.) and the intrinsic attributes of both companies and job positions from recruitment job post data. Then, they designed a transformer-based attentive neural network to automatically utilize this information to forecast the demand trend of each job in each company. Guo et al. introduced a Dynamic Heterogeneous Graph Enhanced Meta-learning (DH-GEM) framework to predict fine-grained talent demand-supply jointly [154]. They employed a Demand-Supply Joint Encoder-Decoder (DSJED) and a Dynamic Company-Position Heterogeneous Graph Convolutional Network (DyCP-HGCN) to capture the correlation between demand-supply sequences and company-position pairs. Additionally, they proposed a Loss-Driven Sampling based Meta-learner (LDSM) to optimize long-tail forecasting tasks with limited training data.

The topic trend primarily focuses on text mining and language modeling techniques applied to job postings. For instance, Marrara et al. designed a language modeling approach for discovering novel occupations in the labor market, which can help the company catch the new trend of recruitment [277]. Zhu et al. developed MTLVM, which is a sequential latent variable model [484]. This model can capture sequential patterns of recruitment states. Moreover, it can automatically learn the latent recruitment topics by the Bayesian generative framework. In detail, it uses $c_{e,t}$ to represent the latent recruitment state of the company $e$ at time step $t$. Then the transition probability between different states is learned to analyze the evolving rules of the recruitment trend, and the topic model is deployed to reveal the trend of different recruitment topics. Azzahra et al. categorized job vacancies into groups like Administration, Finance, IT, and Marketing using a multilabel-classification method with word2vec and an RNN model [29]. Mahdavimoghaddam used NER tools to explore online social topics related to jobs [268]. They assessed social content’s effectiveness in predicting future job requirements, analyzed the relationship between work-related emotions online and social demographics, and identified potential impacts of community support on users’ well-being in the job market.

The skill requirements can be inferred from job postings, and analyzing these requirements can provide valuable assistance in talent selection, job description formulation, and other related tasks. The task of skills requirement prediction task can generally be formulated as: