-
Towards Enhanced Local Explainability of Random Forests: a Proximity-Based Approach
Authors:
Joshua Rosaler,
Dhruv Desai,
Bhaskarjit Sarmah,
Dimitrios Vamvourellis,
Deran Onay,
Dhagash Mehta,
Stefano Pasquali
Abstract:
We initiate a novel approach to explain the out of sample performance of random forest (RF) models by exploiting the fact that any RF can be formulated as an adaptive weighted K nearest-neighbors model. Specifically, we use the proximity between points in the feature space learned by the RF to re-write random forest predictions exactly as a weighted average of the target labels of training data po…
▽ More
We initiate a novel approach to explain the out of sample performance of random forest (RF) models by exploiting the fact that any RF can be formulated as an adaptive weighted K nearest-neighbors model. Specifically, we use the proximity between points in the feature space learned by the RF to re-write random forest predictions exactly as a weighted average of the target labels of training data points. This linearity facilitates a local notion of explainability of RF predictions that generates attributions for any model prediction across observations in the training set, and thereby complements established methods like SHAP, which instead generates attributions for a model prediction across dimensions of the feature space. We demonstrate this approach in the context of a bond pricing model trained on US corporate bond trades, and compare our approach to various existing approaches to model explainability.
△ Less
Submitted 18 October, 2023;
originally announced October 2023.
-
The Geometry of Reduction: Model Embedding, Compound Reduction, and Overlap** State Space Domains
Authors:
Joshua Rosaler
Abstract:
The relationship according to which one physical theory encompasses the domain of empirical validity of another is widely known as "reduction." Here it is argued that one popular methodology for showing that one theory reduces to another, associated with the so-called "Bronstein cube" of physical theories, rests on an over-simplified characterization of the type of mathematical relationship betwee…
▽ More
The relationship according to which one physical theory encompasses the domain of empirical validity of another is widely known as "reduction." Here it is argued that one popular methodology for showing that one theory reduces to another, associated with the so-called "Bronstein cube" of physical theories, rests on an over-simplified characterization of the type of mathematical relationship between theories that typically underpins reduction. An alternative methodology, based on a certain simple geometrical relationship between dis- tinct state space models of the same physical system, is then described and illustrated with examples. Within this approach, it is shown how and under what conditions inter-model reductions involving distinct model pairs can be composed or chained together to yield a direct reduction between theoretically remote descriptions of the same system. Building on this analysis, we consider cases in which a single reduction between two models may be effected via distinct composite reductions differing in their intermediate layer of description, and motivate a set of formal consistency requirements on the map**s between model state spaces and on the subsets of the model state spaces that characterize such reductions. These constraints are explicitly shown to hold in the reduction of a non-relativistic classical model to a model of relativistic quantum mechanics, which may be effected via distinct composite reductions in which the intermediate layer of description is either a model of non-relativistic quantum mechanics or of relativistic classical mechanics. Some brief speculations are offered as to whether and how this sort of consistency requirement between distinct composite reductions might serve to constrain the relationship that any unification of the Standard Model with general relativity must bear to these theories.
△ Less
Submitted 5 October, 2018;
originally announced October 2018.
-
Naturalness, Wilsonian Renormalization, and "Fundamental Parameters" in Quantum Field Theory
Authors:
Joshua Rosaler,
Robert Harlander
Abstract:
The Higgs naturalness principle served as the basis for the so far failed prediction that signatures of physics beyond the Standard Model (SM) would be discovered at the LHC. One influential formulation of the principle, which prohibits fine tuning of bare Standard Model (SM) parameters, rests on the assumption that a particular set of values for these parameters constitute the "fundamental parame…
▽ More
The Higgs naturalness principle served as the basis for the so far failed prediction that signatures of physics beyond the Standard Model (SM) would be discovered at the LHC. One influential formulation of the principle, which prohibits fine tuning of bare Standard Model (SM) parameters, rests on the assumption that a particular set of values for these parameters constitute the "fundamental parameters" of the theory, and serve to mathematically define the theory. On the other hand, an old argument by Wetterich suggests that fine tuning of bare parameters merely reflects an arbitrary, inconvenient choice of expansion parameters and that the choice of parameters in an EFT is therefore arbitrary. We argue that these two interpretations of Higgs fine tuning reflect distinct ways of formulating and interpreting effective field theories (EFTs) within the Wilsonian framework: the first takes an EFT to be defined by a single set of physical, fundamental bare parameters, while the second takes a Wilsonian EFT to be defined instead by a whole Wilsonian renormalization group (RG) trajectory, associated with a one-parameter class of physically equivalent parametrizations. From this latter perspective, no single parametrization constitutes the physically correct, fundamental parametrization of the theory, and the delicate cancellation between bare Higgs mass and quantum corrections appears as an eliminable artifact of the arbitrary, unphysical reference scale with respect to which the physical amplitudes of the theory are parametrized. While the notion of fundamental parameters is well motivated in the context of condensed matter field theory, we explain why it may be superfluous in the context of high energy physics.
△ Less
Submitted 29 January, 2019; v1 submitted 25 September, 2018;
originally announced September 2018.
-
Inter-Theory Relations in Physics: Case Studies from Quantum Mechanics and Quantum Field Theory
Authors:
Joshua Rosaler
Abstract:
The relationship that is widely presumed to hold between physical theories and their successors, in which the successors in some sense explain the success of the theories they replace, is known commonly as 'reduction.' I argue that one traditional approach to theory reduction in physics, founded on the notion that a superseded theory should simply be a mathematical limit of the theory that superse…
▽ More
The relationship that is widely presumed to hold between physical theories and their successors, in which the successors in some sense explain the success of the theories they replace, is known commonly as 'reduction.' I argue that one traditional approach to theory reduction in physics, founded on the notion that a superseded theory should simply be a mathematical limit of the theory that supersedes it, is misleading as a general picture of the relationship whereby one theory encompasses the domain of empirical validity of another. I defend an alternative account that builds upon a certain general type of relationship between dynamical systems models describing the same physical system. I demonstrate how this particular relationship resembles the methodological prescriptions set out by Ernest Nagel's more general approach to reduction across the sciences.
After clarifying these points of general methodology, I go on to apply this approach to a number of particular inter-theory reductions in physics involving quantum theory. I consider three reductions: first, connecting classical mechanics and non-relativistic quantum mechanics; second, connecting classical electrodynamics and quantum electrodynamics; and third, connecting non-relativistic quantum mechanics and quantum electrodynamics. In all cases, a certain core set of mechanisms, employing decoherence together with variations of Ehrenfest's Theorem, serves to underwrite the occurrence of approximately classical behavior. For concreteness, I consider two particular realist interpretations of quantum theory - the Everett and Bohm theories - as potential bases for these reductions. However, many of the technical results concerning these reductions pertain also more generally to the bare, uninterpreted formalism of quantum theory.
△ Less
Submitted 20 February, 2018;
originally announced February 2018.
-
Interpretation Neutrality in the Classical Domain of Quantum Theory
Authors:
Joshua Rosaler
Abstract:
I show explicitly how concerns about wave function collapse and ontology can be decoupled from the bulk of technical analysis necessary to recover localized, approximately Newtonian trajectories from quantum theory. In doing so, I demonstrate that the account of classical behavior provided by decoherence theory can be straightforwardly tailored to give accounts of classical behavior on multiple in…
▽ More
I show explicitly how concerns about wave function collapse and ontology can be decoupled from the bulk of technical analysis necessary to recover localized, approximately Newtonian trajectories from quantum theory. In doing so, I demonstrate that the account of classical behavior provided by decoherence theory can be straightforwardly tailored to give accounts of classical behavior on multiple interpretations of quantum theory, including the Everett, de Broglie-Bohm and GRW interpretations. I further show that this interpretation-neutral, decoherence-based account conforms to a general view of inter-theoretic reduction in physics that I have elaborated elsewhere, which differs from the oversimplified picture that treats reduction as a matter of simply taking limits. This interpretation-neutral account rests on a general three-pronged strategy for reduction between quantum and classical theories that combines decoherence, an appropriate form of Ehrenfest's Theorem, and a decoherence-compatible mechanism for collapse. It also incorporates a novel argument as to why branch-relative trajectories should be approximately Newtonian, which is based on a little-discussed extension of Ehrenfest's Theorem to open systems, rather than on the more commonly cited but less germane closed-systems version. In the Conclusion, I briefly suggest how the strategy for quantum-classical reduction described here might be extended to reduction between other classical and quantum theories, including classical and quantum field theory and classical and quantum gravity.
△ Less
Submitted 19 November, 2015;
originally announced November 2015.
-
"'Formal' vs. 'Empirical' Approaches to Quantum-Classical Reduction"
Authors:
Joshua Rosaler
Abstract:
I distinguish two types of reduction within the context of quantum-classical relations, which I designate "formal" and "empirical". Formal reduction holds or fails to hold solely by virtue of the mathematical relationship between two theories; it is therefore a two-place, a priori relation between theories. Empirical reduction requires one theory to encompass the range of physical behaviors that a…
▽ More
I distinguish two types of reduction within the context of quantum-classical relations, which I designate "formal" and "empirical". Formal reduction holds or fails to hold solely by virtue of the mathematical relationship between two theories; it is therefore a two-place, a priori relation between theories. Empirical reduction requires one theory to encompass the range of physical behaviors that are well-modeled in another theory; in a certain sense, it is a three-place, a posteriori relation connecting the theories and the domain of physical reality that both serve to describe. Focusing on the relationship between classical and quantum mechanics, I argue that while certain formal results concerning singular limits as Planck's constant goes to zero have been taken to preclude the possibility of reduction between these theories, such results at most block reduction in the formal sense; little if any reason has been given for thinking that they block reduction in the empirical sense. I then briefly outline a strategy for empirical reduction that is suggested by work on decoherence theory, arguing that this sort of account remains a fully viable route to the empirical reduction of classical to quantum mechanics and is unaffected by such singular limits.
△ Less
Submitted 19 November, 2015;
originally announced November 2015.
