Can Small Language Models be Good Reasoners for
Sequential Recommendation?
Abstract.
Large language models (LLMs) open up new horizons for sequential recommendations, owing to their remarkable language comprehension and generation capabilities. However, there are still numerous challenges that should be addressed to successfully implement sequential recommendations empowered by LLMs. Firstly, user behavior patterns are often complex, and relying solely on one-step reasoning from LLMs may lead to incorrect or task-irrelevant responses. Secondly, the prohibitively resource requirements of LLM (e.g., ChatGPT-175B) are overwhelmingly high and impractical for real sequential recommender systems. In this paper, we propose a novel Step-by-step knowLedge dIstillation fraMework for recommendation (SLIM), paving a promising path for sequential recommenders to enjoy the exceptional reasoning capabilities of LLMs in a “slim” (i.e., resource-efficient) manner. We introduce CoT prompting based on user behavior sequences for the larger teacher model. The rationales generated by the teacher model are then utilized as labels to distill the downstream smaller student model (e.g., LLaMA2-7B). In this way, the student model acquires the step-by-step reasoning capabilities in recommendation tasks. We encode the generated rationales from the student model into a dense vector, which empowers recommendation in both ID-based and ID-agnostic scenarios. Extensive experiments demonstrate the effectiveness of SLIM over state-of-the-art baselines, and further analysis showcasing its ability to generate meaningful recommendation reasoning at affordable costs.
†Corresponding author.
1. Introduction
Sequential recommendation is extensively utilized in a variety of internet applications due to its prominent performance in uncovering a user’s evolving and dynamic interests from his/her chronological interactions (Quadrana et al., 2018). Despite the effectiveness, existing models are often trained on a closed-loop user-item interaction dataset, inevitably suffering from severe exposure bias and popularity bias. Therefore, beyond narrow information present in the original datasets, it is crucial to incorporate open-world knowledge to foster a more comprehensive and generalized understanding of historical behaviors.
Due to the impressive reasoning capability, the recent emergence of Large Language Models (LLMs), such as GPT 3.5/4, has brought a significant breakthrough in various NLP tasks (Qin et al., 2023; Zhang et al., 2023b; Wei et al., 2023; Pan et al., 2023), showing substantial potential in overcoming the isolated nature of real-world sequential recommenders that rely on closed data sources for training (Sun et al., 2019; Chen et al., 2018). These LLMs are trained on massive corpora, granting them to exhibit the remarkable capability of human-like thinking as well as seamless reasoning. Roughly speaking, current LLM empowered recommenders mainly fall into the following two groups: (1) LLM as a ranker, which typically involves prompting the frozen LLM to offer a reasonable ranked list that satisfies the user interests (Hou et al., 2023). However, solely relying on the zero-shot or few-shot learning capability of LLMs is still inferior compared to traditional sequential recommendations that utilize in-domain collaborative knowledge. To address this limitation, (2) LLM as a knowledge enhancer has been proposed, typically following a cascading architecture: the LLM is first instructed to generate rich knowledge (e.g., user preference and factual knowledge on items), followed by a classical recommendation backbone for harvesting in-domain knowledge and collaborative signals. Generally, the bridging of both worlds tends to elicits a more promising performance (Xi et al., 2023). While LLMs for recommendation hold promise, they also face significant challenges that cannot be ignored.
One is the exceptional reasoning capability of LLM within the context of recommendation has not been fully explored. There is a gap between the open-world nature and recommender systems, which means that the recommendation knowledge generated by LLMs may be incorrect or task-irrelevant. Fortunately, with the chain-of-thought (CoT) prompting strategy (Wei et al., 2022; Magister et al., 2022; Hsieh et al., 2023), LLMs can break down complex tasks into a series of intermediate reasoning steps, which can improve the ability to understand behavior patterns and explore user interests. Consequently, there is a strong motivation to leverage the CoT reasoning capability of LLMs in sequential recommender systems, enabling the generation of targeted recommendation-related rationales. For instance, guiding LLMs to reason progressively, similar to a human salesperson, to deduce user interests, narrow down the categories of items that align with their interests, and ultimately recommend specific items within these categories that the user is likely to interact with.
Another significant challenge is the prohibitively high resources are far beyond affordable for real-world recommender systems. The immense size of LLMs demands a considerable amount of memory and computational power, which necessitates specialized infrastructure. For instance, the deployment of the open-source LLaMA2-70B requires eight Nvidia A100 servers. On the other hand, working with closed-source LLMs also involves significant costs. For instance, using ChatGPT as an example, the current approach requires calling its API, which comes with substantial monetary expenses. For instance, in gpt-3.5-turbo, the costs are approximately $0.0015 per 1,000 tokens for input and $0.002 per 1,000 tokens for output. Therefore, a natural question arises:
Can a language model with affordable costs still serve as
an effective reasoning engine for sequential recommendation?
To answer this question, in this paper, we propose a novel Step-by-step knowLedge dIstillation fraMework for recommendation (SLIM), which enables sequential recommendations to enjoy the significant reasoning capabilities of LLMs in a “slim” (i.e., resource-efficient) manner. Specifically, we develop a step-by-step knowledge distillation strategy for sequential recommendations to transfer the reasoning capabilities of LLMs (i.e., teacher) to a “small” language model (i.e., student). This strategy guides the larger teacher model to engage in macro-to-micro thinking for complex recommendation task through CoT prompting. Through the process of distillation, the small student model with only 4% parameters of the large teacher model acquires step-by-step thinking capabilities and evolves into a good reasoner. Subsequently, we directly deploy the small language model as a knowledge generator for sequential recommendation, which can derive high-quality reasoning knowledge highly relevant to recommendation. These knowledge reflect user preferences for categories, brands, and specific items, which can be flexibly integrated with any sequential recommendation backbone, including ID-based and ID-agnostic scenarios. Our key contributions can be summarized as follows:
-
•
To the best of our knowledge, it is the first knowledge distillation framework of LLMs tailored for sequential recommendation.
-
•
We propose SLIM, a novel step-by-step knowledge distillation framework, empowering sequential recommenders with the CoT reasoning capabilities of LLMs in a resource-efficient manner.
-
•
Extensive experiments on three datasets demonstrates the effectiveness of our proposed SLIM. Further analysis reveals that SLIM generates meaningful reasoning at affordable costs.
2. The Proposed Framework
In this section, we propose SLIM, a novel knowledge distillation framework tailored for recommendation, which incorporates the reasoning capabilities of LLMs into recommender systems in a resource-efficient manner. The overview is illustrated in Figure 1.
2.1. Sequential Recommendation Backbone
Sequential recommendation aims at the accurate prediction of users’ next behavior by capturing evolved and dynamic preferences over historical behavior sequences, which has occupied a critical position in various modern information systems (Quadrana et al., 2018). In general, the success of sequential recommendation typically hinges on the meaningful representation of items and effectively encoding behavior patterns.
Item Representation. For neural sequential recommendations, the item encoder is the key component which transfers the items to representations. Formally, given an item set , each item may be associated with several optional attributes , such as title, category and brand. The encoder can generate the item representations for each item based on their ID (i.e., ) and attributes (i.e., ):
(1) |
where is the representation of item . Generally, is implemented as a hybrid architecture where a embedding layer aims at tackling ID-like features (e.g., item id), coupled with a text encoder (e.g., BERT (Devlin et al., 2018)) for context embedding based on the item description (e.g., title, category).
Sequential Encoding. To capture the sequential characteristics of user behaviors, the action sequence of user can be organized in chronological order , where represents the -th item that the user interacted with. Next, each item in is firstly fed into (denoted as ), followed by the a sequential encoding.
(2) |
where denotes the representation of sequence . is sequence encoder, which can be implemented with the Attention (Vaswani et al., 2017) or other neural architectures (Zhang et al., 2019; Sherstinsky, 2020). Based on the sequence , our objective is to predict the next item that the user is likely to interact with at the -th step.
Prediction and Optimization. After generating the above representations, we can obtain the final prediction at time with dot product or MLP layer followed by a sigmoid activation function (He et al., 2017), where each element indicates how likely the item should be recommended to the target user . Finally, the model is trained with binary cross-entropy (De Boer et al., 2005) loss as follows:
(3) |
Note that these classical sequencical models typically perform recommendation based on the user action sequences and item attributes (e.g., title, category and brand), lacking the reasoning power that have recently emerged in LLMs.
2.2. Step-by-Step Knowledge Distillation for Recommendation
Despite the remarkable reasoning ability of LLMs, it is non-trivial to adapt LLMs to empower the traditional recommender systems. The challenge arises from two aspects: (1) Complex behavior patterns of users are difficult to understand directly by LLMs. (2) The large size and high inference latency of LLMs exacerbates resource-consuming. Therefore, we propose step-by-step knowledge distillation to transfer the reasoning capabilities of LLMs to a smaller LLaMA2-7B (Touvron et al., 2023) model specialized for the recommendation tasks.
In detail, our distillation strategy consists of two straightforward steps: Firstly, we employ CoT prompting related to user behavior to guide the LLM (i.e., teacher) in thinking step-by-step and generating natural language rationales that support its predictions in the recommendation scenario. Secondly, these rationales are subsequently utilized as labels to fine-tune the downstream smaller language model (i.e., student), enabling it to approach the reasoning capabilities of the larger model in the recommendation domain. Finally, the fine-tuned smaller model acts as the ultimate knowledge generator, offering reasoning knowledge to the recommender systems.
2.2.1. Extracting Recommendation Rationales from LLMs
Traditional sequential recommendations rely solely on scores to perform recommendation and ignore the intermediate reasoning steps, which constrains the accuracy and explainability of recommendations. Inspiringly, the emergence of LLMs has led to a significant breakthrough in understanding human and generating rationales step-by-step for recommendation through CoT prompting. Therefore, we guide LLMs in generating critical reasoning, encompassing user preferences, interested categories/brands and specific items, all of which are essential for providing appropriate recommendations.
To achieve this objective, we utilize zero-shot CoT prompting to elicit LLMs in extracting this reasoning information from user behaviors. Specifically, given the user set and the behavior dataset , we construct a user sub-set and a small behavior sub-dataset by random sampling, where and . Then, we have meticulously designed a CoT prompt template to facilitate in-depth reasoning according to user behavior by LLMs. As illustrated in Figure 2, our instructions consists of three progressive steps:
-
•
Summarize user preferences based on the historical behavior sequences.
-
•
Recommend categories or brands to the user based on the summarized preferences.
-
•
Recommend products that align with the recommended categories/brands.
The template starts with a macro perspective and gradually zooms in to a micro perspective, which guides LLMs in a step-by-step thinking process and ensures that the output of the LLM conforms to a specific format. Through this process, LLMs can effectively leverage extensive open-world knowledge to infer the aspects that users are likely to be interested in the future. As shown in Figure 2, through CoT prompting, the teacher model can generate informative recommendation rationales in response.
Subsequently, we further fill the template with user historical behaviors in small dataset to generate corresponding CoT prompts for the teacher LLM model. With this incremental CoT prompts , LLMs will generate corresponding recommended rationales for each input . Although the current recommendations overlook these rationales, they are crucial for achieving more efficient and explainable recommender systems. Besides, unlike previous studies (Xi et al., 2023), we do not concentrate on inferring item factual information (e.g., the director of movie items), as most real-world items (e.g., food, home and kitchen items) are difficult to associate with trustworthy open-world knowledge.
2.2.2. Fune-tuning Smaller Models with Recommendation Rationales.
By guiding the thought process of LLMs step-by-step, we can comprehend complex behavior patterns of users and generate high-quality recommendation rationales. However, their large scale and computational overhead make them unsuitable for recommendation scenarios that require low latency. For instance, serving a single 175 billion LLM necessitates a minimum of 350GB of GPU memory (Zheng et al., 2022). Despite recent study (Xi et al., 2023) attempting to mitigate this issue by offline inference, it’s still unaffordable to generate recommendation rationales for all users in the real-world scenario.
To this end, we leverage knowledge distillation to transfer the recommendation reasoning capabilities of larger teacher models to smaller student models, thereby reducing the computational overhead. Considering that complex prompts can improve the reasoning quality of large models but greatly increase the understanding difficulty of small models, we design simplified template based on the template , as showed in Figure 3. Subsequently, we generate simplified prompts as input, and collect the rationales generated by teacher LLMs as the expected output labels to fine-tune the smaller student model. As a result, for a given input instruction , we train the smaller model with parameters to generate the corresponding recommendation rationale . Formally, we optimize the negative log-likelihood of conditional language modeling objective as follows:
(4) |
where is the - token of the , represents the tokens before . To conserve resources, we employ the LoRA (Hu et al., 2021) for parameter-efficient model fine-tuning. This approach involves training only a small set of additional parameters instead of the entire model. Through experimental validation, we demonstrate that the generated rationales maintain a comparable quality to models with 25 times the model size, despite using a limited number of training samples and a smaller model size. As illustrated in Figure 3, the student model responses showed a step-by-step reasoning ability similar to that of the teacher model. For instance, the student model initiates by logically inferring the user’s intent by leveraging its recommendation-related CoT. Subsequently, it offers potential game genres that align with the user’s interests. Ultimately, several specific games are recommended to the user.
Overall, by utilizing recommendation rationales as labels instead of generating pseudo-labels for recommended results from LLMs, we enhance the smaller language model with step-by-step reasoning capabilities similar to the reasoning process of the larger model.
2.3. Empowering Recommender with Reasoning Knowledge
With the help of step-by-step knowledge distillation, small language models can become efficient reasoners. However, traditional sequential recommendation models cannot directly utilize the rationales of natural language forms. Thus, in this section, we explore how to apply the recommendation rationales generated by small language models to the sequence recommendation model, enabling it to efficiently combine the reasoning ability of LLMs in a resource-efficient manner. Specifically, we introduce two application approaches. The first approach is ID-based, where we treat the rationales text as supplementary knowledge and combine them with ID-based recommendation backbone to improve the traditional closed-loop learning dependent on user-item interactions. The second approach is ID-agnostic, where we encode the rationale text of user behaviors and the description text of candidate items as the representations of user and item, respectively. This allows us to make recommendations based on text similarity.
2.3.1. Encoding Recommendation Rationales
Owing to the efficient reasoning power of small language model, each user behavior sequence can be associated with corresponding CoT rationales , while each item can be associated with attribute descriptions (e.g., title, category, brand). Then we leverage pre-trained language models (PLMs) to learn text representations, enabling the measurement of semantic distance in vector space. Concretely, we adopt the text encoder to map the text on both the item side and the sequence side into a unified semantic space:
(5) |
where and represent the text representations of item descriptions and recommendation rationales, respectively. The encoder can be flexibly configured as any frozen or trainable text encoding model, which we instantiate with BERT (Devlin et al., 2018) in this work. Due to the step-by-step thinking process of language model, the representation encodes rich reasoning knowledge from open-world at both macro-level (i.e., general user preference) and micro-level (i.e., specific recommended item).
2.3.2. Utilizing Recommendation Rationales
Traditional sequential recommendation methods learn the user sequence representation based only on the user-item interaction history, resulting in an information-enclosed model. To alleviate this issue, we enhance the traditional recommender systems with the rationale representations obtained from Eq. (5), which is derived from open-world knowledge and deep reasoning about the user’s behavior patterns. Specifically, we leverage it through the following two approaches.
Empowering ID-Based Recommendation. To disrupt the closed systems of sequential recommendation, we integrate the rationale representations into traditional recommendation backbone, effectively combining the open-world reasoning knowledge with the collaborative signal of traditional recommendations. Specifically, we propose an information fusion layer to combine the meaningful text representations (i.e., and ) with the original embeddings in the backbone model as follows:
(6) |
where is the ID embedding of item, is the sequence representation obtained from SeqEncoder (i.e., Eq (2)) in backbone model, denotes the concatenation operation, transforms the text representations to the same dimension with the ID embeddings, and is an fusion layer that enables the model to learn and incorporate flowing information from both sources. Without loss of generality, we implement and with linear layers.
Empowering ID-Agnostic Recommendation. Recent studies have revealed that sequential models that focus on text modeling exhibit superior generalization abilities and are more effective in handling cold-start items (Li et al., 2023; Hou et al., 2022). Therefore, we explore a direct utilization of rationale representations in ID-agnostic recommendation scenarios. In this case, the representations of item text and rationale are directly transformed into a unified space as follows:
(7) |
where denotes the transformation layer, which we implement using linear layers. Since contains step-by-step reasoning knowledge about user preferences, the model can recommend item with matching item-side information to the user and provide explainable recommendation rationales of natural language form.
3. Experiments
Methods | Games | Food | Home | ||||||
---|---|---|---|---|---|---|---|---|---|
NDCG@10 | Hit @10 | Hit @20 | NDCG@10 | Hit @10 | Hit @20 | NDCG@10 | Hit @10 | Hit @20 | |
GRU4Rec | 17.61 ± 0.18 | 30.87 ± 0.56 | 42.39 ± 0.62 | 9.10 ± 0.30 | 15.27 ± 0.58 | 19.51± 0.24 | 2.19 ± 0.21 | 4.17 ± 0.39 | 7.53 ± 0.64 |
27.33 ± 0.53 | 44.06 ± 0.79 | 56.53 ± 1.24 | 17.75± 0.78 | 31.10 ± 1.09 | 45.01 ± 1.46 | 12.19 ± 1.02 | 26.76 ± 2.58 | 49.10 ± 3.82 | |
27.70 ± 0.47 | 45.13± 0.56 | 57.70 ± 0.37 | 17.97 ± 0.70 | 31.78 ± 1.47 | 46.88 ± 2.10 | 13.59 ± 1.05 | 30.30 ± 2.01 | 55.85 ± 3.55 | |
SLIM | 28.37 ± 0.41 | 45.68 ± 0.53 | 58.09 ± 0.58 | 18.32 ± 0.53 | 32.56 ± 1.30 | 46.92 ± 1.82 | 15.64 ± 0.51 | 34.33 ± 1.53 | 62.93 ± 3.46 |
Improv. | 3.81 | 3.68 | 2.76 | 3.21 | 4.69 | 4.24 | 28.3 | 28.29 | 28.17 |
SASRec | 22.73 ± 0.28 | 37.77 ± 0.52 | 51.53 ± 0.39 | 26.78 ± 0.24 | 35.78 ± 0.36 | 43.32 ± 0.55 | 2.66 ± 0.22 | 5.56 ± 0.72 | 14.93 ± 1.53 |
27.46 ± 0.19 | 44.88 ± 0.63 | 58.90 ± 0.38 | 30.95 ± 0.38 | 44.98 ± 0.53 | 55.61 ± 1.12 | 5.58 ± 0.10 | 11.09 ± 0.16 | 20.69 ± 0.77 | |
31.58 ± 0.35 | 50.83 ± 0.62 | 63.45 ± 0.71 | 32.65 ± 0.15 | 48.01 ± 0.48 | 59.25 ± 0.56 | 5.95 ± 0.32 | 11.83 ± 0.62 | 22.43 ± 0.54 | |
SLIM | 31.43 ± 0.39 | 51.11 ± 0.82 | 64.10 ± 0.26 | 32.80 ± 0.40 | 48.27 ± 0.64 | 59.30 ± 0.89 | 6.01 ± 0.19 | 12.01 ± 0.38 | 22.29 ± 0.85 |
Improv. | 14.46 | 13.88 | 8.83 | 5.98 | 7.31 | 6.64 | 7.71 | 8.3 | 7.73 |
SRGNN | 16.45 ± 0.22 | 29.29 ± 0.14 | 40.99 ± 0.52 | 10.99 ± 2.07 | 20.32 ± 4.30 | 32.14 ± 6.55 | 5.04 ± 0.83 | 13.48 ± 2.23 | 37.22 ± 3.85 |
21.54 ± 0.64 | 36.77 ± 1.05 | 49.11 ± 1.54 | 11.91 ± 0.71 | 21.39 ± 1.91 | 33.63 ± 3.41 | 11.61 ± 1.14 | 25.22 ± 2.46 | 43.85 ± 3.58 | |
22.35 ± 1.48 | 37.69 ± 1.53 | 51.29 ± 0.59 | 12.92 ± 0.78 | 23.80 ± 1.60 | 37.22 ± 2.75 | 11.25 ± 1.42 | 24.28 ± 3.19 | 44.05 ± 5.97 | |
SLIM | 23.77 ± 0.20 | 39.81 ± 0.52 | 52.34 ± 0.63 | 12.38 ± 0.51 | 22.98 ± 1.30 | 36.44 ± 1.72 | 12.29 ± 1.39 | 26.51 ± 2.71 | 47.01 ± 3.61 |
Improv. | 10.35 | 8.27 | 6.58 | 3.95 | 7.43 | 8.36 | 5.86 | 5.11 | 7.21 |
Methods | Games | Food | Home | ||||||
---|---|---|---|---|---|---|---|---|---|
NDCG@10 | Hit @10 | Hit @20 | NDCG@10 | Hit @10 | Hit @20 | NDCG@10 | Hit @10 | Hit @20 | |
SLIM-Step1 | 13.78 ± 0.59 | 26.08 ± 0.91 | 41.71 ± 0.62 | 13.62 ± 0.22 | 24.90 ± 0.59 | 38.15 ± 0.69 | 4.25 ± 0.06 | 9.53 ± 0.27 | 18.91 ± 0.77 |
SLIM-Step2 | 16.78 ± 0.66 | 30.09 ± 0.89 | 45.75 ± 0.59 | 13.71 ± 0.48 | 24.23 ± 0.78 | 36.89 ± 0.77 | 4.75 ± 0.29 | 10.30 ± 0.55 | 19.99 ± 0.61 |
SLIM-Step3 | 20.20 ± 0.31 | 35.57 ± 0.54 | 50.04 ± 0.47 | 15.69 ± 0.26 | 26.69 ± 0.68 | 39.19 ± 0.98 | 4.83 ± 0.27 | 10.39 ± 0.38 | 21.05 ± 0.43 |
21.75 ± 0.58 | 38.05 ± 0.88 | 53.73 ± 0.79 | 19.08 ± 0.71 | 32.34 ± 0.74 | 44.63 ± 0.90 | 4.63 ± 0.28 | 10.15 ± 0.62 | 20.11 ± 1.01 | |
SLIM | 21.99 ± 0.22 | 38.33 ± 0.32 | 53.59 ± 0.72 | 18.13 ± 0.5 | 30.84 ± 0.54 | 44.04 ± 0.66 | 4.49 ± 0.24 | 9.96 ± 0.47 | 19.67 ± 0.59 |
3.1. Experimental Settings
3.1.1. Datasets.
We conduct our experiments on three categories from the Amazon Review dataset: Video Games (Games), Grocery and Gourmet Food (Food), and Home and Kitchen (Home). More information about these datasets are presented in Appendix A.1.
3.1.2. Baselines.
We adopt three widely used sequencial recommendation models as the backbone, i.e., GRU4Rec, SASRec, SRGNN. More information about these backbones are shown in Appendix A.2. For each backbone, we examine the performance of its Item Feature Extensions: denoted as , , and . These extensions concatenating the item ID vector and item description text vector as the input, resulting in enhanced item representations. We also introduce another ChatGPT Feature Extension of each backbone: , which directly input the rationales generated by the teacher model into Eq (5) without distillation. The implementation details of each methods shown in Appendix A.3.
3.1.3. Evaluation Metrics
The details of evaluation shown in Appendix A.3. We utilize three widely-adopted metrics for evaluation: NDCG@10, Hit Rate@10, and Hit Rate@20. The average scores of 5 runs and the standard deviation are reported. Following the strategy in (Kang and McAuley, 2018), we randomly sample 100 negative items for each user and rank these items alongside the ground-truth item. The rankings of these 101 items are then used to evaluate.
3.2. Overall Performance
3.2.1. Improvement over Backbone Models in ID-based scenarios.
Our SLIM is highly flexible and can be integrated with any type of sequential recommendation backbone. Firstly, we evaluate the performance of the SLIM across various backbones. The results of these comparisons are presented in Table 1. We have the following observations: (1) Compared to all backbones and their item feature extensions, the proposed SLIM achieves state-of-the-art (SOTA) performance across all datasets. This further substantiates the effectiveness of our model in enhancing traditional recommendations. Notably, SLIM achieves a relative improvement of over the in terms of Hit Rate@10 on the Home dataset. These improvements are attributed to the meaningful rationales generated from our distilled student model, which contains a wealth of knowledge that benefits recommendations as a valuable supplement to closed collaborative signals. (2) Surprisingly, in most cases (22 / 27), SLIM in each backbone outperforms the ChatGPT feature extensions , achieving a relative improvement of in terms of Hit Rate@10 on the Home dataset with the GRU4Rec backbone. While SLIM’s knowledge is distilled from the teacher model ChatGPT, the lack of control over closed-source models may result in the generation of correct but irrelevant responses to recommendations. This indicates that our smaller model can further prioritize the information relevant to recommendations after distillation. Despite being smaller in scale, it greater effectiveness in recommendations.
3.2.2. Performance in ID-agnostic scenarios.
To establish a more efficient and generalizable model, we evaluate the performance of SLIM in ID-agnostic scenarios, i.e., we solely based on matching CoT-based sequence embeddings and text-based item embeddings as Eq (7), named Text Matching. The results are shown in Table 2. We also obtained interesting findings: (1) In comparison to models that generate rationales based on the teacher model (), SLIM outperforms it in of cases, despite having only of the parameters compared to ChatGPT. This demonstrates that even with limit training samples (1000-2000) and a smaller model size, SLIM can generate high-quality recommendation rationales that are highly competitive with ChatGPT. (2) Additionally, this straightforward matching approach exhibits superior performance compared to all ID-based backbones listed in Table 1. This indicates that high-quality text from both the sequence and item side can lead to promising recommendations, even without meticulous design of the text encoder. (3) To verify the effectiveness of each step in the rationales, i.e., the user interest of Step1, the item category of Step2, and the specific product of Step3, we evaluate them separately in Text Matching. It is worth noting that the ranking of recommendation performance consistently follows the pattern of Step3 ¿ Step2 ¿ Step1 in all cases. Surprisingly, on the Home dataset, Step3 even surpasses the performance achieved using the entire Rationale. These results suggest that the smaller model trained with CoT prompting is capable of step-by-step thinking, similar to human reasoning. As the chain of thought evolves, the information relevant to recommendations will be inferred. However, the performance of the Step1 is not satisfactory, possibly because the macroscopic information in this step fails to align well with the microscopic information on the item side, such as titles and categories. Nevertheless, the first step still plays a crucial role as the foundation for subsequent reasoning processes and ensures the interpretability of the model.
3.3. Merits of SLIM
Potentially Good Interpretability for the Recommendation Results. Figure 4 illustrates a sample where SLIM successfully recommends the ground-truth, while SASRec fails. The target next item in this sample is a long-tail item that only appears once in the training set. As a result, traditional ID-based models struggle to capture adequate collaborative signals. However, SLIM’s generated rationales are able to deduce the user’s preferences, which align closely with the characteristics of the target item “Milk Protein Rich Nutrition Bar”, such as the categories of “Healthy snacks”. More significantly, SLIM showcases its remarkable reasoning capabilities and extensive domain knowledge by accurately inferring that users are likely to purchase “Protein Bars”. In this manner, the textual information from both the sequence side and item side aligns well, leading to a high similarity in the vector space. Moreover, SLIM generates rationales in human-understandable natural language. The rationales provided in Step 1 and Step 2 offer justifications for the recommendation of “Protein Bars” by SLIM. For each recommended item, SLIM can provide a natural language explanation, enhancing the interpretability of the recommendation process.
Consistent Improvement for User with Different Sparsity. To investigate the impact of interaction data sparsity, we group users based on the sparsity level of their interactions and evaluate the performance of SLIM separately on different user groups. Specifically, we sorted users based on their interaction frequency, and then divided them equally into five user groups. Subsequently, SLIM and SASRec are trained separately on the interaction data of each user group, and compare their recommendation performance on different user groups. The results, as depicted in Figure 5, that SLIM consistently outperforms SASRec across all user groups, and SLIM exhibits greater improvement on the relatively sparse user group compared to the dense user group . This suggests that our method’s improvement is stable and robust, effectively mitigating the issue of sparsity in sequential recommendation.
Impressive Capability of Alleviating Popularity Bias. In the field of recommender systems, popularity bias means that popular items are recommended even more frequently than their popularity would warrant. This bias intensifies the long-tail effects in real-world recommendation domains. To analyze the impact of our proposed SLIM on popularity bias, we count the frequency of items in the training data and recommendation results. As depicted in Figure 6 (the results of two additional datasets are presented in Figure 7 in the Appendix), our method effectively recommends tail items compared to the traditional method SASRec, which focuses on recommending popular head items. Experimental results confirm that our method significantly mitigate the popularity bias.
Costs | LLM as Ranker | LLM as Enhancer | SLIM |
---|---|---|---|
Offline Time (s) | ✗ | 5.54 (API call) | 6.11 |
Online Time (s) | 5.54 (API call) | ✗ | ✗ |
Model Size | 175B | 175B | 7B |
Deployment | Hard | Hard | 1 A100 |
API Costs/Input | $0.0015/1K tokens | $0.0015/1K tokens | ✗ |
API Costs/Output | $0.002/1K tokens | $0.002/1K tokens | ✗ |
Significantly More Affordable Compared to SOTAs. SLIM has demonstrated promising performance. In this part, we will analyze the efficiency of the model through a comprehensive cost analysis across multiple dimensions, including time cost, model size, deployment difficulty and API monetary cost. Specifically, we compare two representative recommendation models based on the generation capabilities of LLM. (1) LLM as a ranker (Hou et al., 2023). (2) LLM as a knowledge enhancer (Xi et al., 2023). We do not take into account the costs associated with backbones, as the backbone model is typically variable and the cost is generally negligible compared to LLM. Due to the two-stage nature of the LLM as a knowledge enhancer approach, it involves offline knowledge generation based on LLM and online inference. The term “Offline/Online Time” refers to the average response time of the closed-source ChatGPT API for the compared methods. Conversely, our method corresponds to the average inference time on a single Nvidia A100 GPU. It is worth mentioning that despite deploying SLIM on only one GPU, we achieve a comparable time cost compared to the API call duration of ChatGPT, which requires significant resource consumption for deployment. From Table 3, it can be concluded that SLIM is a highly efficient model compared to existing LLM-based recommendations. It possesses acceptable inference latency, minimal model size, and can be deployed on limited resources. Additionally, being based on an open-source model, it does not incur any financial cost.
Specifically, we study the data efficiency of fine-tuning the student models, and we conclude that only 1000 samples is enough for a promising performance, the details are shown in Appendix B.1.
4. Related Work
4.1. Sequential Recommender Systems
The core idea of existing sequential recommendations lies in initially formalizing user behavior as a chronologically-ordered interaction sequence with items, followed by designing diverse behavior encoders to learn behavior patterns that accurately depict user interests (Wang et al., 2019). GRU4Rec (Hidasi et al., 2015) is one of the earliest attempts to learn evolving patterns for user behaviors using Gated Recurrent Units (GRU). With the rapid development of deep learning (Wang et al., 2017), there have also been emerging many neural networks as behavior encoders, including Convolutional Neural Networks-based methods (Tang and Wang, 2018), Attention-based methods (Kang and McAuley, 2018), and Graph Neural Networks-based methods (Wu et al., 2019; Sun et al., 2024; Wang et al., 2022). To enhance the transferability of the sequence modeling, recent studies have begun exploring ID-agnostic text-based modeling approaches (Li et al., 2023; Hou et al., 2022), such as Recformer (Li et al., 2023), which proposes formulating each item as a “sentence” and designing a Transformer-based language model to learn user preference representations. However, these methods rely on limited textual information provided by the recommendation dataset, which restricts the model’s capabilities due to its isolation from rich open-world knowledge. Recently, the emergence of LLM that utilize massive training corpora and large model sizes has disrupted the traditional closed-loop of user-item interaction in recommendations (Liu et al., 2023a, b).
4.2. LLM Enhanced Recommender Systems
The utilization of LLMs, with their human-like understanding and generation capabilities, introduces new knowledge spaces in recommendations (Lin et al., 2023a; Fan et al., 2023; Wu et al., 2023). To integrate LLM’s generation capabilities into recommendations, the current methods can be primarily categorized into two mainstream trends based on the different roles LLM plays in the recommendation pipeline. The first trend involves utilizing LLM as a ranker or scorer (Zhang et al., 2023a; Liu et al., 2023b; Dai et al., 2023; Hou et al., 2023; Zhang et al., 2023c; Kang et al., 2023). For instance, (Hou et al., 2023) explores the zero-shot ranking capabilities of LLM in recommendation. This requires careful design of prompts that involve a predefined list of candidate items for the limited re-ranking stage. (Zhang et al., 2023c) proposes to view recommendation as instruction following by LLMs. In this approach, 39 instruction templates are manually designed for LLMs. However, these methods often exhibit limited performance because the frozen LLMs are typically trained on open-world corpora that lack domain-specific collaborative signals from recommendations. To incorporate collaborative information, recent studies have started exploring another trend, which involves utilizing LLM as a knowledge enhancer to complement traditional recommendations (Liu et al., 2023a; Xi et al., 2023). For example, (Xi et al., 2023) explores the acquisition of user preferences and item factual knowledge from ChatGPT, and utilizes them to enhance traditional Click-Through Rate (CTR) prediction. (Liu et al., 2023a) proposes to employ open-source LLM as content encoders and utilize closed-source ChatGPT to enrich the training data. (Lin et al., 2023b) proposes a LLM-based augmentation technique to enrich training samples. While promising, existing work has not fully leveraged the step-by-step reasoning capabilities of LLM in the recommendation scenario. Furthermore, current approaches often rely on the use of large model sizes to achieve improved reasoning capabilities. Although techniques like prestoring can be used to deploy only the inference model, these models still require larger model sizes in either offline or online stages, which may not be feasible in real-world recommender systems.
5. Conclusion
In this paper, we propose SLIM, a method that enables sequential recommender systems to leverage the substantial reasoning capabilities of LLMs in a resource-efficient manner. We design a step-by-step knowledge distillation module to transfer the step-by-step reasoning capabilities in recommendation from a larger teacher model to a smaller student model (with approximately 4% of the parameters of the teacher model). This smaller model evolves into a proficient reasoner, which can be directly deployed as a “slim” knowledge generator for sequential recommendation. Consequently, this knowledge can be flexibly integrated with any sequential recommendation backbone and utilized in both ID-based and ID-agnostic scenarios. The experimental results demonstrate that SLIM significantly improves the performance of sequential recommendation backbones. It also achieves promising results in ID-agnostic scenarios without relying on any backbone. Furthermore, additional analysis experiments highlight that the costs associated with SLIM are affordable and have the potential to enhance the interpretability of recommendations. A possible future direction is to design customized knowledge encoders to further capture the information from smaller models.
Acknowledgements.
The research was supported in part by the National Natural Science Foundation of China (No. U22B2019, 62322203, 62172052, 62276248) and BUPT Excellent Ph.D. Students Foundation (CX2022220).References
- (1)
- Anelli et al. (2021) Vito Walter Anelli, Alejandro Bellogín, Antonio Ferrara, Daniele Malitesta, Felice Antonio Merra, Claudio Pomo, Francesco Maria Donini, and Tommaso Di Noia. 2021. Elliot: A comprehensive and rigorous framework for reproducible recommender systems evaluation. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2405–2414.
- Chen et al. (2018) Xu Chen, Hongteng Xu, Yongfeng Zhang, Jiaxi Tang, Yixin Cao, Zheng Qin, and Hongyuan Zha. 2018. Sequential recommendation with user memory networks. In Proceedings of the eleventh ACM international conference on web search and data mining. 108–116.
- Dai et al. (2023) Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering ChatGPT’s Capabilities in Recommender Systems. arXiv preprint arXiv:2305.02182 (2023).
- De Boer et al. (2005) Pieter-Tjerk De Boer, Dirk P Kroese, Shie Mannor, and Reuven Y Rubinstein. 2005. A tutorial on the cross-entropy method. Annals of operations research 134 (2005), 19–67.
- Devlin et al. (2018) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018).
- Fan et al. (2023) Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Jiliang Tang, and Qing Li. 2023. Recommender systems in the era of large language models (llms). arXiv preprint arXiv:2307.02046 (2023).
- He et al. (2017) Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web. 173–182.
- Hidasi et al. (2015) Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015).
- Hou et al. (2022) Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards universal sequence representation learning for recommender systems. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 585–593.
- Hou et al. (2023) Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2023. Large language models are zero-shot rankers for recommender systems. arXiv preprint arXiv:2305.08845 (2023).
- Hsieh et al. (2023) Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. arXiv:2305.02301 [cs.CL]
- Hu et al. (2021) Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021).
- Kang and McAuley (2018) Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM). IEEE, 197–206.
- Kang et al. (2023) Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed Chi, and Derek Zhiyuan Cheng. 2023. Do LLMs Understand User Preferences? Evaluating LLMs On User Rating Prediction. arXiv preprint arXiv:2305.06474 (2023).
- Li et al. (2023) Jiacheng Li, Ming Wang, ** Li, **miao Fu, Xin Shen, **gbo Shang, and Julian McAuley. 2023. Text Is All You Need: Learning Language Representations for Sequential Recommendation. arXiv preprint arXiv:2305.13731 (2023).
- Lin et al. (2023a) Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, et al. 2023a. How Can Recommender Systems Benefit from Large Language Models: A Survey. arXiv preprint arXiv:2306.05817 (2023).
- Lin et al. (2023b) Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2023b. Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendation. arXiv preprint arXiv:2308.11131 (2023).
- Liu et al. (2023b) Junling Liu, Chao Liu, Renjie Lv, Kang Zhou, and Yan Zhang. 2023b. Is chatgpt a good recommender? a preliminary study. arXiv preprint arXiv:2304.10149 (2023).
- Liu et al. (2023a) Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2023a. ONCE: Boosting Content-based Recommendation with Both Open- and Closed-source Large Language Models. arXiv:2305.06566 [cs.IR]
- Magister et al. (2022) Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2022. Teaching small language models to reason. arXiv preprint arXiv:2212.08410 (2022).
- Pan et al. (2023) Wenbo Pan, Qiguang Chen, Xiao Xu, Wanxiang Che, and Libo Qin. 2023. A Preliminary Evaluation of ChatGPT for Zero-shot Dialogue Understanding. arXiv:2304.04256 [cs.CL]
- Qin et al. (2023) Chengwei Qin, Aston Zhang, Zhuosheng Zhang, Jiaao Chen, Michihiro Yasunaga, and Diyi Yang. 2023. Is ChatGPT a General-Purpose Natural Language Processing Task Solver? arXiv:2302.06476 [cs.CL]
- Quadrana et al. (2018) Massimo Quadrana, Paolo Cremonesi, and Dietmar Jannach. 2018. Sequence-Aware Recommender Systems. arXiv:1802.08452 [cs.IR]
- Sherstinsky (2020) Alex Sherstinsky. 2020. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. Physica D: Nonlinear Phenomena 404 (2020), 132306.
- Sun et al. (2019) Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management. 1441–1450.
- Sun et al. (2024) Li Sun, Zhenhao Huang, Zixi Wang, Feiyang Wang, Hao Peng, and Philip Yu. 2024. Motif-aware Riemannian Graph Neural Network with Generative-Contrastive Learning. arXiv preprint arXiv:2401.01232 (2024).
- Tang and Wang (2018) Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining. 565–573.
- Touvron et al. (2023) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023).
- Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017).
- Wang et al. (2019) Shou** Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z Sheng, and Mehmet Orgun. 2019. Sequential recommender systems: challenges, progress and prospects. arXiv preprint arXiv:2001.04830 (2019).
- Wang et al. (2017) Xiao Wang, Peng Cui, **g Wang, Jian Pei, Wenwu Zhu, and Shiqiang Yang. 2017. Community preserving network embedding. In Proceedings of the AAAI conference on artificial intelligence, Vol. 31.
- Wang et al. (2022) Yuling Wang, Hao Xu, Yanhua Yu, Mengdi Zhang, Zhenhao Li, Yuji Yang, and Wei Wu. 2022. Ensemble multi-relational graph neural networks. arXiv preprint arXiv:2205.12076 (2022).
- Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35 (2022), 24824–24837.
- Wei et al. (2023) Xiang Wei, Xingyu Cui, Ning Cheng, Xiaobin Wang, Xin Zhang, Shen Huang, Pengjun Xie, **an Xu, Yufeng Chen, Meishan Zhang, Yong Jiang, and Wenjuan Han. 2023. Zero-Shot Information Extraction via Chatting with ChatGPT. arXiv:2302.10205 [cs.CL]
- Wu et al. (2023) Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. 2023. A Survey on Large Language Models for Recommendation. arXiv preprint arXiv:2305.19860 (2023).
- Wu et al. (2019) Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, Vol. 33. 346–353.
- Xi et al. (2023) Yunjia Xi, Weiwen Liu, Jianghao Lin, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, Rui Zhang, and Yong Yu. 2023. Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models. arXiv preprint arXiv:2306.10933 (2023).
- Zhang et al. (2023a) Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023a. Is chatgpt fair for recommendation? evaluating fairness in large language model recommendation. arXiv preprint arXiv:2305.07609 (2023).
- Zhang et al. (2023c) Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2023c. Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach. arXiv:2305.07001 [cs.IR]
- Zhang et al. (2019) Si Zhang, Hanghang Tong, Jiejun Xu, and Ross Maciejewski. 2019. Graph convolutional networks: a comprehensive review. Computational Social Networks 6, 1 (2019), 1–23.
- Zhang et al. (2023b) Wenxuan Zhang, Yue Deng, Bing Liu, Sinno Jialin Pan, and Lidong Bing. 2023b. Sentiment Analysis in the Era of Large Language Models: A Reality Check. arXiv preprint arXiv:2305.15005 (2023).
- Zhao et al. (2021) Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, et al. 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. In proceedings of the 30th acm international conference on information & knowledge management. 4653–4664.
- Zheng et al. (2022) Lianmin Zheng, Zhuohan Li, Hao Zhang, Yonghao Zhuang, Zhifeng Chen, Yan** Huang, Yida Wang, Yuanzhong Xu, Danyang Zhuo, Eric P Xing, et al. 2022. Alpa: Automating inter-and Intra-Operator parallelism for distributed deep learning. In 16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22). 559–578.
Appendix A Details of Experimental Settings
A.1. Dataset
We conduct our experiments on three categories from the Amazon Review dataset: Video Games (Games), Grocery and Gourmet Food (Food), and Home and Kitchen (Home). In this dataset, reviews are considered as interactions. We follow the same preprocessing procedure from (Kang and McAuley, 2018; Hou et al., 2023), i.e., users and items with fewer than five interactions were filtered out. The interactions of each user were sorted in ascending order based on timestamps to construct the corresponding historical interaction sequences. To cope with the significant cost associated with ChatGPT, we conduct random sampling of complete user interactions from each dataset. Following the sampling process, the resulting sizes of the Games, Food, and Home datasets are as follows: the number of users [3000, 2000, 2000], the number of items [9647, 11190, 15740], and the number of interactions [26222, 17668, 20000]. For item description text, we utilize the titles from all datasets and the sales type from the Food and Home datasets. Following (Kang and McAuley, 2018; Hou et al., 2023), we employ the leave-one-out strategy. Specifically, for each interacted sequence , the most recent interaction is used for testing, the second most recent interaction for validation, and all remaining interactions for training. During testing, the input sequences consist of both training and validation interactions.
A.2. Backbone Models
The detailed information for each backbone is as follows:
-
•
GRU4Rec (Hidasi et al., 2015) is a pioneering method that uses Recurrent Neural Networks to model user behavior sequences.
-
•
SASRec(Kang and McAuley, 2018) is typical self-attention based framework designed to capture the user’s preferences within a sequence.
-
•
SRGNN(Wu et al., 2019) is graph-based model designed to capture the transition information between items in user sequences.
A.3. Implementation Details
We employ the powerful close-source ChatGPT as the teacher model and the open-source LLaMA2-7B (Touvron et al., 2023) as the student model. The OpenAI’s API gpt-3.5-turbo111https://openai.com/ is utilized for generating the rationales of the teacher model. We follow the parameters provided in llama-recipes222https://github.com/facebookresearch/llama-recipes for fine-tuning LLaMA2-7B. Specifically, for the Games, Food, and Home datasets, data sizes of 1000, 2000, and 2000 were used for fine-tuning LLaMA2-7B, respectively. Additionally, we set the maximum length of generated tokens to be 300 during the inference process for LLaMA2-7B. The implementation of SASRec is based on the PyTorch version code333https://github.com/pmixer/SASRec.pytorch. For GRU4Rec and SRGNN, we utilize the RecBole (Zhao et al., 2021), which is a widely-used open-source recommendation library. All embeddings obtained from Eq. (5) have a dimensionality of 768. The implementation of SLIM ensures that the embedding dimensions remain consistent with the original backbone. We carefully search hyper-parameters of all the baselines to get the best performance.
Appendix B More Experimental Results
B.1. Impact of Different Data Size for Fine-tuning Student Models
In this part, we analyze the efficiency of the model by analyzing the data size for fine-tuning the student models. In particular, we validate the optimal fine-tuning data size required for student models under different backbones, as illustrated in Figure 8. The recommendation performance demonstrates a trend of initially increasing and then decreasing with an increase in the data size across all backbones. Remarkably, the optimal performance is consistently achieved with 1000 fine-tuning samples. This finding suggests that SLIM can effectively perform the fine-tuning process using minimal data from the teacher model. This is also supported by the observation that fine-tuning based on CoT is more efficient compared to fine-tuning based on labels (Magister et al., 2022).
B.2. Impact of Different Student Model Size for Distilling
We explore the utilization of additional LLMs, specifically Bloomz-560M and Bloomz-1B, as student models. The outcomes of these experiments are comprehensively presented in Table 4 (ID-based) and Table 5 (ID-agnostic). Our findings are noteworthy: 1) Across diverse backbones, the incorporation of any distilled student model consistently enhances recommendation performance, underscoring the efficacy of our proposed SLIM. 2) As the parameter scale expanded from 560M to 7B, the student model reveals an substantial upward trend in recommendation performance. This suggests that the limited reasoning ability of smaller-sized LLMs complicates the distillation process for sequential recommendation tasks. Consequently, in practical scenarios, striking a judicious balance between efficiency and performance becomes imperative. Furthermore, the decision to refrain from further distilling larger Bloomz models was driven by their architectural similarities to LLaMA, coupled with the impractical expenses entailed when deploying them in practical recommendation scenarios.
Games | |||
Methods | NDCG@10 | Hit @10 | Hit @20 |
GRU4Rec | 17.61 ± 0.18 | 30.87 ± 0.56 | 42.39 ± 0.62 |
27.33 ± 0.53 | 44.06 ± 0.79 | 56.53 ± 1.24 | |
Bloomz-560M | 27.80 ± 0.55 | 44.79 ± 0.58 | 57.23 ± 0.40 |
Bloomz-1B | 28.22 ± 0.33 | 45.54 ± 0.87 | 57.48 ± 0.76 |
LLaMA2-7B | 28.37 ± 0.41 | 45.68 ± 0.53 | 58.09 ± 0.58 |
SASRec | 22.73 ± 0.28 | 37.77 ± 0.52 | 51.53 ± 0.39 |
27.46 ± 0.19 | 44.88 ± 0.63 | 58.90 ± 0.38 | |
Bloomz-560M | 30.50 ± 0.33 | 49.96 ± 0.59 | 63.08 ± 0.26 |
Bloomz-1B | 30.93 ± 0.29 | 50.67 ± 0.33 | 63.35 ± 0.48 |
LLaMA2-7B | 31.43 ± 0.39 | 51.11 ± 0.82 | 64.10 ± 0.26 |
SRGNN | 16.45 ± 0.22 | 29.29 ± 0.14 | 40.99 ± 0.52 |
21.54 ± 0.64 | 36.77 ± 1.05 | 49.11 ± 1.54 | |
Bloomz-560M | 22.36 ± 0.90 | 37.48 ± 1.26 | 49.65 ± 0.87 |
Bloomz-1B | 23.05 ± 0.66 | 38.47 ± 0.81 | 50.34 ± 0.72 |
LLaMA2-7B | 23.77 ± 0.20 | 39.81 ± 0.52 | 52.34 ± 0.63 |
Games | |||
---|---|---|---|
Methods | NDCG@10 | Hit @10 | Hit @20 |
Bloomz-560M | 19.31 ± 0.32 | 34.15 ± 0.36 | 49.55 ± 0.88 |
Bloomz-1B | 20.92 ± 0.28 | 36.81 ± 0.37 | 51.77 ± 0.77 |
LLaMA2-7B | 21.99 ± 0.22 | 38.33 ± 0.32 | 53.59 ± 0.72 |
B.3. Examining the Generalization Capacity of SLIM.
We demonstrate the generalization ability of SLIM through the following experiments. Specifically, we distill a student model on the Home dataset and subsequently deployed it for direct inference on two other datasets, namely Games and Food. The results are presented in Table 6 (ID-based) and Table 7 (ID-agnostic), where it can be observed that , the dataset-transferred model, not only does not exhibit a significant decline in performance compared to SLIM
but, in some instances, surpasses SLIM’s performance at Hit@10 and Hit@20. This experimentation leads us to the conclusion that the student model within the SLIM framework demonstrates robust transferability across different datasets. Even without specific dataset fine-tuning, the student model is effectively endowed with incremental reasoning capabilities. Therefore, the SLIM enhances the student model’s ability to generalize and transfer knowledge across recommendation domains, indicating potential cost savings as the knowledge enhancer exhibits a level of generalization.
Games | Home | |||||
---|---|---|---|---|---|---|
Methods | NDCG@10 | Hit @10 | Hit @20 | NDCG@10 | Hit @10 | Hit @20 |
GRU4Rec | 17.61 ± 0.18 | 30.87 ± 0.56 | 42.39 ± 0.62 | 9.10 ± 0.30 | 15.27 ± 0.58 | 19.51± 0.24 |
28.17 ± 0.46 | 45.22 ± 0.43 | 57.39 ± 0.58 | 17.80 ± 0.45 | 31.70 ± 0.71 | 47.16 ± 1.29 | |
SLIM | 28.37 ± 0.41 | 45.68 ± 0.53 | 58.09 ± 0.58 | 18.32 ± 0.53 | 32.56 ± 1.30 | 46.92 ± 1.82 |
SASRec | 22.73 ± 0.28 | 37.77 ± 0.52 | 51.53 ± 0.39 | 26.78 ± 0.24 | 35.78 ± 0.36 | 43.32 ± 0.55 |
31.39 ± 0.23 | 50.83 ± 0.3 | 63.37 ± 0.51 | 32.68 ± 0.36 | 48.20 ± 0.83 | 59.74 ± 0.98 | |
SLIM | 31.43 ± 0.39 | 51.11 ± 0.82 | 64.10 ± 0.26 | 32.80 ± 0.40 | 48.27 ± 0.64 | 59.30 ± 0.89 |
SRGNN | 16.45 ± 0.22 | 29.29 ± 0.14 | 40.99 ± 0.52 | 10.99 ± 2.07 | 20.32 ± 4.30 | 32.14 ± 6.55 |
23.64 ± 0.35 | 39.74 ± 0.65 | 51.63 ± 0.62 | 12.83 ± 0.48 | 23.85 ± 1.27 | 36.91 ± 1.60 | |
SLIM | 23.77 ± 0.20 | 39.81 ± 0.52 | 52.34 ± 0.63 | 12.38 ± 0.51 | 22.98 ± 1.30 | 36.44 ± 1.72 |
Games | Home | |||||
---|---|---|---|---|---|---|
Methods | NDCG@10 | Hit @10 | Hit @20 | NDCG@10 | Hit @10 | Hit @20 |
21.15 ± 0.46 | 37.33 ± 0.45 | 53.15 ± 0.44 | 18.70 ± 0.39 | 31.07 ± 0.71 | 43.92 ± 0.55 | |
SLIM | 21.99 ± 0.22 | 38.33 ± 0.32 | 53.59 ± 0.72 | 18.13 ± 0.50 | 30.84 ± 0.54 | 44.04 ± 0.66 |
EFD@10 | EPC@10 | |||||
---|---|---|---|---|---|---|
Methods | Games | Food | Home | Games | Food | Home |
SASRec | 0.5810 | 0.6550 | 0.0765 | 0.0511 | 0.0531 | 0.0059 |
SLIM | 0.8171 | 0.9197 | 0.1835 | 0.0690 | 0.0732 | 0.0135 |
B.4. Quantitative Evaluation of Mitigating Popularity Bias
To quantitatively evaluate the models’ effectiveness in mitigating popularity bias, we calculated EFD@10 and EPC@10 using the open-source recommendation framework Elliot (Anelli et al., 2021) for both SLIM and the corresponding backbone, as presented in Table 8. All these metrics serve as indicators of a system’s capability to recommend relevant long-tail items. From Table 8, it is evident that SLIM outperforms the corresponding backbone across every metric. This strongly supports our claim that SLIM excels in alleviating popularity bias, aligning with the phenomenon observed in Figure 6 and Figure 7.
B.5. Case Study
To explain and analysis of the phenomenon wherein the performance of is inferior to SLIM in certain cases, i.e., the results shown in Table 1 and Table 2, we incorporate responses from both GPT-3.5 and the distilled student model LLaMA2-7B for the same user behavior. The specific cases are outlined as Figure 9. The comparison of these responses leads to the conclusion that the smaller student model, following CoT fine-tuning, can more effectively infer key information from user behavior and better match the text of the target item. In contrast, GPT-3.5 falls short in achieving this level of performance.