acceptodds
Withdrawn from ICLR 2027

Do Neural Networks Preserve Case Structure? Case-Based Decomposition, Interpretation, and Decision Consistency

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural networks increasingly influence consequential decisions, yet the case-based foundations behind these decisions remain hidden. This makes it difficult to understand which learned experiences support a decision and when such decisions can be trusted. Motivated by Case-Based Decision Theory (CBDT), which views decisions as being formed from relevant past cases, we ask whether neural networks preserve a similar case structure during learning. We first show that trained neural networks preserve a recoverable case structure within their learned representations. This structure enables a CBDT-inspired decomposition that traces fitted decisions back to contributions from individual training cases. We then reveal a fundamental gap between the recovered case structure and classical CBDT assumptions. Although neural networks preserve case-based decision structure, the recovered case coefficients do not necessarily satisfy CBDT relevance requirements. Enforcing these stricter constraints changes selected decisions and reduces task-optimal accuracy by 4.2%–9.1% relative to signed reconstruction. Finally, we establish a verifiable decision-consistency certificate that identifies when the recovered case structure agrees with the original network decision. Across three real-world decision settings, certified decisions achieve substantially higher task accuracy (83.7% vs. 35.0% on the Adult Income dataset). Case-level analysis further identifies compact sets that preserve some certified fitted decisions. Together, our results reveal that neural networks can preserve case-based decision structures without fully satisfying CBDT assumptions. More importantly, they show that such imperfect but verifiable structures can still provide a principled foundation for reliable case-level understanding of neural network decisions.

Primary area:
interpretability and explainable AI
Keywords:
Neural Networks, Case-Based Decision Theory, Explainable AI, Model Interpretation, Representation Learning, Data Attribution, Decision Consistency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uHVQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper considers the problem of bridging case-based decision theory (CBDT) with the decisions of a neural network. More specifically, the problem setup is a DNN trained to predict returns for a set of actions, for example the return for each of the actions of "conservative, standard, and expanded approval" (I'm assuming of a loan?) using the Adult Income dataset.

The core of the method, to my understanding, is as follows:

  1. Train the neural network.
  2. Extract the feature representations from the DNN (penultimate layer?) for all training data.
  3. Fit a linear model per possible action with ordinary least squares (OLS) or ridge regression on these representations. (Equation 2)
  4. The prediction of the fitted linear model for a new example can then be decomposed as a weighted sum of the training examples' returns (for that action) (Equation 4). These weights can be negative.
  5. The contribution of each training sample to a decision can then be expressed as the contribution of the sample to the margin between two of the fitted linear models (Equation 5).

The authors then further establish that the found weights must be positive to "admit a Case-Based Decision Theory interpretation", which they can construct using non-negative least squares. This is followed by multiple experiments on four datasets evaluating the method and comparing it to other training sample decomposition methods.

Critical strengths
  1. The paper achieves its goal of bridging CBDT and DNN decisions.
  2. The overall method appears sound.
Critical weaknesses

I have three primary concerns:

  1. Presentation - The paper is difficult to follow, with very little explanation, intuition, or examples to guide the reader. For illustration, a small example of what is meant with "action-specific returns" would greatly improve the readability, as this a niche problem setting not often encountered in the ML literature. Similarly, several concepts are introduced without any preceding explanation as to why it is relevant to the problem. The effect of this is that there's significant work required on behalf of the reader to follow the logic.
  2. Limited experimental evaluation and explanation: the experimental section is hard to parse, with the actual setup in the appendix. Furthermore, the baseline comparison (Table 2) does not establish that this decomposition is a "better" method of assigning contributions to training samples, but simply that it's better correlated with a given reference (Gaussian similarity weighting).
  3. Contribution - The OLS expansion of equations 3 and 4 is a known method, and the practical benefit of the CBDT interpretation compared to other training point decomposition methods is unclear.
Questions for authors
  1. Why is a CBDT interpretation a desirable property? Essentially, it is unclear to me why interpreting the coefficients as CBDT relevance is any "better" than the standard method with potentially negative coefficients. How does the one make it more clear how the neural network has reached a given decision than the other? And what benefit does this interpretation have over other methods such as influence functions?
  2. How does the core method differ from that of prior work? Are there any novel methodological alterations not used before? This is not adequately explained in the current manuscript, and it is difficult to determine the novelty of the method.
  3. Are there no real-world datasets available for this task? All datasets use real-world inputs but synthetic labels.
  4. Can the authors demonstrate that the derived explanations are more true to the neural network's decisions than the baselines? I see no clear evidence of this.
9 comments
AuthorsOpenReview ↗

W1. Presentation and action-specific returns

Thank you very much for your comments and for recognizing the connection between CBDT and neural network decisions. We appreciate your questions about the interpretation, practical benefit, and evidence. Below we address each point and provide the corresponding manuscript revisions and completed comparisons.

W1. Presentation and action-specific returns

Thank you for the comment. We are sorry that the action-specific returns and the purpose of the fitting stages were not explained clearly. We have added a numerical solver case to explain how different actions receive different returns and how these returns form a case contribution. We have also clarified the fitting sequence and the purpose of the subsequent decision conditions.

Section 2.1, illustration of action-specific returns:

Consider choosing between two solver configurations, A and B, for a computational problem. A historical case records the problem's observed features and the runtime of each configuration. Suppose A solves the problem in 10 seconds and B in 30 seconds. Under the return definition r=−log⁡(1+t)r=-\log(1+t), where tt is runtime cost, A receives approximately −2.398 and B approximately −3.434. Thus, the same case has a separate return for each action, and its return difference of 1.036 favors A. For a new problem, CBDT combines such past return differences according to the relevance of past cases. In the measured-runtime experiment, the training return is −log⁡(1+timeout-penalized runtime cost)-\log(1+\text{timeout-penalized runtime cost}), and evaluation reports the cost in seconds.

To make the meaning of an individual case contribution more explicit, we will also clarify the explanation following Eq. (1).

Section 2.1, explanation following Eq. (1):

Each case contribution combines how relevant a past case is to the current problem with how much its returns favor one action over another. In the solver case above, A's return advantage is approximately 1.04. If the case has a relevance weight of 0.2, it contributes approximately 0.21 in favor of A. Other cases can favor B instead. The sum of these contributions forms the overall case-based preference.

To connect the case description to the subsequent mathematical steps, the fitting sequence is stated explicitly in Section 3.1.

Section 3.1, fitting sequence:

The network is trained before the linear return models are fitted. Its learned representation and original score function remain fixed during this second stage. In the original return-supervised experiments, both stages use the training cases' action returns. A supervision comparison changes the first-stage target while retaining the subsequent return fit (Appendix C.4).

The following passages introduce the decision questions before their formal conditions.

Section 4.2, motivation preceding Eq. (9):

To use a fitted preference to explain the network's decision, we check whether score error could reverse its winner. Suppose the fitted scores are 3 for solver A and 1 for B. The fitted winning margin is 2, while twice the maximum score error is 0.8. Even in the worst case, A's network score falls to 2.6 while B's rises to 1.4, so the network still selects A. Let εfid(x)\varepsilon_{\mathrm{fid}}(x) bound this difference for every action: ∣S(x,a)−Q^(x,a)∣≤εfid(x),a∈A.|S(x,a)-\widehat Q(x,a)|\leq\varepsilon_{\mathrm{fid}}(x),\quad a\in\mathcal A. (9) The following result gives the general condition.

Section 5, opening motivation:

The combination requirement in Gilboa et al. (2002, Section 3, A2) motivates retaining a preference supported by two case sets after combining them.

AuthorsOpenReview ↗

W2. Experimental setup and the reference comparison

W2. Experimental setup and the reference comparison

Thank you for pointing this out. We are sorry that the experimental setup and the purpose of the reference comparison were not explained clearly. The original Table 2 (now Table F1 in Appendix F.1) compares our case contributions with attribution-inspired scores under the same trained representation, fitted return models, and action pair. Its reference combines Gaussian input-based relevance with action-return differences. Higher correspondence therefore indicates that a scoring rule more closely reflects this specified case-based preference pattern. The result supports the connection between our representation-based contributions and the CBDT relevance-and-return form. The applicability of the explanation to the original network is examined separately through the decision-agreement condition.

We will clarify these roles and retain the reference comparison as Table F1 in Appendix F.1. To make the experiments easier to follow, we have placed the essential task, fitting, and evaluation settings at the beginning of Section 6, before presenting the results, as follows.

Section 6, experimental setup:

We use a controlled CBDT setting to evaluate score reconstruction and correspondence with specified generating references. PJM, Adult, and Default use real inputs with constructed action returns to compare case reconstructions under a common task setup. CSP-2010 supplies recorded runtimes for two solver configurations. We report agreement with the signed fitted model, agreement with the original network, and task outcomes separately. The reference comparison in Appendix F.1 and the removal experiments in Appendix F.2 address different properties of the case contributions.

To explain the purpose of the original Table 2, we will state what the reference comparison evaluates in Section 6.1 and specify how its scores are compared in Appendix F.1.

Section 6.1, explanation of the reference comparison:

The fitted case contributions align more closely with the specified Gaussian relevance-and-return reference than the adapted attribution scores (Appendix F.1). This comparison tests the ranking and support direction of individual cases under the controlled reference. Agreement with the original network is evaluated separately.

Appendix F.1, comparison setup:

We compare case contributions and attribution-inspired scores on the same trained representation, fitted return models, and action pair. The reference combines each training case's Gaussian input-based relevance with its return difference for that action pair. Closer correspondence indicates that a ranking better reflects this specified case-based preference pattern.

Appendix F.1, revised table caption:

Table F1: Case-contribution correspondence with the specified Gaussian input-based relevance-and-return reference over 500 sampled training cases. Ranking is evaluated by Top-30 Jaccard and nDCG, while sign agreement measures support direction.

AuthorsOpenReview ↗

W3. Contribution beyond the known expansion

W3. Contribution beyond the known expansion

Thank you for the comment. We agree that the OLS expansion in Eqs. (3) and (4), together with its equivalent kernel form, is standard. Our contribution is not the expansion itself. We use CBDT, a theory of human rational decision-making based on relevant past cases, to examine whether neural networks designed to model human decisions can be explained through their training cases.

This leads to three connected questions. First, can the learned representation provide a case-level explanation with the nonnegative relevance required by CBDT? Second, under what condition does this fitted explanation apply to the original network's decision? Third, when case sets are pooled, can common preferences be retained to maintain decision stability? Joint fitting addresses the first question, the sufficient condition addresses the second, and the pooling analysis addresses the third.

These contributions therefore concern the feasibility, applicability, and stability of CBDT-based explanations of neural network decisions, rather than a new least-squares expansion.

We have clarified this progression in the Introduction and explicitly distinguished the standard mathematical foundation from the questions addressed in Section 3.

Introduction, contributions:

CBDT explains human decisions through experience with relevant past cases. Neural networks also learn from past cases, which motivates us to use CBDT to examine how their learned representations can explain neural network decisions. Our contributions are threefold (Figure 1):

First, we connect trained representations to the CBDT view by decomposing fitted preferences into individual training-case contributions, revealing how past returns support or oppose each choice. Since CBDT relevance requires nonnegative weights, we show why feature-only reconstruction can change fitted choices and develop joint fitting to reduce these changes while improving task-optimal accuracy.

Second, at the network level, we establish a sufficient condition for the fitted case-based explanation to apply to the original network's decision. We use its margin-to-error ratio to rank test cases and identify a smaller subset with higher task-optimal accuracy.

Third, extending the CBDT case-based explanation to pooled data, we derive a requirement for neural network decision consistency: preferences supported by both case sets should be retained after pooling. On a real-world task, our method corrects preference reversals after pooling while retaining an explicit case decomposition, and retains all common choices, improving neural network decision stability.

Section 3, opening paragraph:

The ordinary least-squares fit and its equivalent kernel expansion are standard. We use them to express fitted action preferences through training-case returns, and examine their CBDT relevance interpretation and agreement with the original network decision.

Reviewer_j2sLOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to understand if a NN model's output can be interpreted in a way consistent with choice in CBDT. That is, each action is scored via summed outcomes weighted by reference. To align this with a pre-trained neural net, the framework freezes the last layer and fit a least squares model based on each action's payoff.

Critical strengths

The paper has strong experimental hygiene with code and multiple seeds The paper nicely unifies thinking about neural nets and decision theory, which seems relevant in the real world.

Critical weaknesses

The weights come from fitting on a reconstruction task over a small number of features but ignores a high number of payoffs. As the paper's formulation indicates that score errors are attributable to payoff residuals, it would be nice to see if the same set of weights can just optimize both objectives.

The paper also shows limited evidence that highlighted cases are particularly more important. For instance, removing the ten strongest supporting cases does not seem to differ from removing random cases in the Appendix experiment. The paper would benefit from discussing this to contextualize the paper's relevance.

The presentation of the result that the case structure is satisfied follows circularly from the objective's construction. The network is trained to optimize via squared error the evaluation payoffs. The paper isn't convincing that the networks preserve case structure, just that the training objective can be cleanly optimized. Because the readout and training phases are not separated, it is difficult to draw meaningful conclusion in practice.

More minor comment, but the paper presents a standard least squares with a equivalent kernel as a new formulation.

Questions for authors

Please see weaknesses.

\epsilon(x) requires evaluating network(x). in this case, you can just check directly if two models agree. in this case, what is the value? is it possible to get \epsilon wtihout evaluating the network?

6 comments
AuthorsOpenReview ↗

W1. Joint fitting of features and payoffs

Thank you for recognizing the connection between neural networks and decision theory and for identifying the points that need clearer explanation. Your observation about the score error left by feature-only fitting prompted us to revise the joint fitting procedure. We have also clarified the meaning of individual case contributions and the relationship between network training and the subsequent fitting, and stated more precisely what the least-squares decomposition and decision-consistency condition establish. The revised analysis further examines whether preferences supported by two case sets are retained after pooling. Our responses and the corresponding manuscript revisions are provided below.

W1. Joint fitting of features and payoffs

Thank you for this insightful suggestion. To address the score error left by feature-only reconstruction, we now fit the same nonnegative case weights to both the learned representation and the fitted action scores, while keeping the trained network unchanged. The corresponding addition to Section 3.5 explains why both fitting errors need to be considered and gives the joint objective as follows.

Section 3.5, Joint Fitting of Representations and Action Scores:

Let R=ΦV+ER=\Phi V+E collect the training returns and their linear-fit residuals, and let q^(x)=V⊤ϕ(x)\widehat q(x)=V^\top\phi(x) collect the fitted action scores. For case weights uu, the score error is R⊤u−q^(x)=V⊤(Φ⊤u−ϕ(x))+E⊤uR^\top u-\widehat q(x)=V^\top(\Phi^\top u-\phi(x))+E^\top u.

The first term reflects the error in reconstructing the learned representation, while the second reflects the contribution of the linear-fit residuals to the score error. We therefore fit both objectives using one nonnegative case-weight vector: min⁡u≥0∣Φ⊤u−ϕ(x)∣22sϕ2+η∣R⊤u−q^(x)∣22sr2.\min_{u\geq0}\frac{\\|\Phi^\top u-\phi(x)\\|_2^2}{s_\phi^2}+\eta\frac{\\|R^\top u-\widehat q(x)\\|_2^2}{s_r^2}.

Here, sϕs_{\phi} and srs_r are the root-mean-square sizes of the training representations and action returns. Dividing by these values puts the two fitting errors on comparable numerical scales. We use η=1\eta = 1 and also compare η=0.1\eta=0.1 and 10, score-only fitting, and nearest-case fitting (Appendix D.4). The trained network and the separate linear return models remain fixed, and the targets are their fitted action scores.

Section 6.2, Table 2 and accompanying analysis:

Table 2: Joint nonnegative fitting versus feature-only fitting. Disagreement is measured against the signed fitted model on 200 cases per seed. Task-optimal accuracy is measured on 300 cases per seed, including those 200. Values are five-seed means, with paired 95% confidence intervals for accuracy gains in percentage points (pp).

TaskDisagreement (%) Feature → JointAccuracy (%) Feature → JointAccuracy gain (pp) [95% CI]
PJM28.1 → 0.059.60 → 65.535.93 [0.77, 11.10]
Adult14.7 → 0.176.47 → 79.873.40 [0.51, 6.29]
Default18.9 → 0.068.40 → 73.074.67 [1.60, 7.74]

Feature-only fitting closely reconstructs the learned representation but can still change the fitted choice. By fitting the representation and fitted action scores with the same nonnegative case weights, joint fitting reduces disagreement to 0–0.1% and improves task-optimal accuracy across all three tasks (Table 2). Thus, the decision changes introduced by feature-only fitting can be largely avoided while retaining the nonnegative weights required for the CBDT relevance interpretation.

AuthorsOpenReview ↗

W2. The meaning of highlighted cases

W2. The meaning of highlighted cases

Thank you for pointing out that the removal experiment does not show a larger margin change for the ten highest-ranked cases than for random removal. This highlights an important distinction between a case’s contribution to the current fitted preference and the change in that preference after the case is removed.

In Eq. (5), ΔQ^(x;a,b)=∑i=1nαi(x)Δri(a,b)\Delta\widehat Q(x;a,b)=\sum_{i=1}^{n}\alpha_i(x)\Delta r_i(a,b), each case contributes a term determined by its coefficient and its return difference between two actions. The sum of these terms gives the original fitted margin. Thus, the contribution measures how an individual past case supports or opposes the current fitted choice and by how much. Removal and refitting instead measure how the fitted margin changes after cases are removed and the linear return models are fitted again. We have clarified these two quantities and their respective roles in the manuscript.

Section 3.3, explanation of case contributions following Eq. (5):

The sign and magnitude of each case contribution show whether and how strongly it supports one action over another in the fitted model. Together, these contributions show how past return differences form the fitted preference. Section 4 specifies when this fitted preference selects the same action as the original network.

Appendix F.2, interpretation of the existing removal comparison:

Removing the ten highest-ranked supporting cases and refitting gives a normalized margin reduction of 0.0019 [−0.0044, 0.0082], with a difference of 0.0015 [−0.0048, 0.0078] relative to random removal. This comparison reports how the fitted margin changes after removal and refitting, complementing Eq. (5), which decomposes the original fitted margin into individual case contributions.

AuthorsOpenReview ↗

W3. Separate fitting stages and the meaning of the decomposition

W3. Separate fitting stages and the meaning of the decomposition

Thank you for pointing out that the original presentation did not sufficiently distinguish network training from the subsequent fitting of the linear return models. The two stages are separate in our implementation, although both use the same continuous-return targets. We first train the network to predict action-specific returns. After training, we hold its learned representation and original score function fixed and fit a separate linear return model for each action. The decomposition therefore expresses the fitted action scores on the trained representation.

To examine whether the subsequent case-based explanation follows automatically from network training, we changed the first-stage supervision to optimal-action labels while retaining continuous-return fitting in the second stage. Network task accuracy rises from 94.5% to 98.5%, while agreement with the subsequent fitted return models falls from 98.8% to 89.1% (Appendix C.4). This shows that higher task accuracy alone does not ensure that the case-based explanation applies to the network's decision.

Section 3.1, clarification of the fitting sequence:

The network is trained before the linear return models are fitted. Its learned representation and original score function remain fixed during this second stage. In the original return-supervised experiments, both stages use the training cases' action returns. A supervision comparison changes the first-stage target while retaining the subsequent return fit (Appendix C.4).

Section 6.1, supervision comparison:

When the network is trained on the best-action label rather than continuous returns, task accuracy increases from 94.5% to 98.5%, while fitted–network agreement falls from 98.8% to 89.1% (Appendix C.4). This distinguishes task accuracy from agreement with the fitted explanation.

The revised organization connects the three contributions to their practical roles. Section 3 explains how individual past returns support or oppose a fitted preference and how joint fitting reduces decision changes while retaining nonnegative CBDT relevance. Section 4 establishes when the fitted case-based explanation applies to the original network's decision and uses this condition to identify decisions with higher observed task-optimal accuracy. Section 5 extends the same CBDT case-based explanation to pooled case sets and addresses preference reversals, improving decision stability when cases are combined. Together, these contributions make it possible to inspect why a network favors an action, determine when its case-based explanation applies to the network, and maintain common choices when past cases are pooled.

The pooling analysis extends this same case-based explanation to two sets of past cases. Two case sets can separately support the same preference, while their combined fitted preference reverses it. In our full-rank construction, the separate margins are 0.465 and 0.635, whereas the combined margin is −0.134 (Section 5, Appendix A.8). This motivates our third contribution, which derives when the CBDT requirement to retain preferences supported by both case sets holds and examines a correction that restores common choices using recorded solver runtimes.

Reviewer_zzMTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper describes an approach for explainability where the authors attempt to recover “case structure” from network representations. The preference of a decision aa vs bb is described in terms of ΔS(x;a,b)\Delta S(x; a,b); for simplicity, I consider a binary classification case in this description. The next step fits a linear readout on the network representation ϕ(x)\phi(x) and decomposes the fitted preference, ΔQ^(x;a,b)\Delta\hat Q(x;a,b), using explanatory instances from the training set. The main decomposition uses all training instances, and considers the usage of few instances in additional experiments. The explanation is expressed with a linear combination of the differences of the training returns, Δri(a,b)=ri(a)−ri(b)\Delta r_i(a,b)=r_i(a)-r_i(b), using some coefficients αi(x)\alpha_i(x).

Alternatively, from my understanding the fitted readout can be also described using bias-free LS-SVMs: Q^(x,a)=∑iβi(a)k(xi,x)\hat Q(x,a)=\sum_i\beta_i^{(a)}k(x_i,x), where xix_i are training instances, and kk is the kernel induced by the network representation, <ϕ(x),ϕ(x′)><\phi(x),\phi(x')>. The solution is fitted using a least-square formulation.

Furthermore, the authors consider assumptions and requirements, showing what happens in unoptimal and best conditions. The authors validate their results against three tabular datasets with constructed action returns, and against three adapted baselines (inspired by influence functions, representer-style, single-checkpoint gradient), with the baseline comparison performed in a controlled synthetic setting.

Critical strengths

S1: The approach is genuinely interesting and has potential. The overall simplicity of the method has its strength, as its mathematical foundations are well understood and can be expanded in a proper framework.

S2: The approach has theoretical guarantees that help rely on the methodology, specifically regarding agreement between the fitted readout and the original network’s selected action.

Critical weaknesses

W1: The procedure outlined to recover case structure does not establish whether the recovered coefficients represent a case structure retained by the network during learning. In particular, since the proposed αi(x)\alpha_i(x) are uniquely specified by the method, it is possible to consider different allocations producing identical predictions. The results do not establish why the proposed allocation should be interpreted as a case structure recovered from the network. The authors examine these settings: in Table 2 the authors show better agreement with an input space-based reference than the adapted baselines, while E.7 shows that some decisions can be reproduced with few cases. However, the paper does not address significantly why the reference represents the network case contributions.

W2: The removal experiment in F.2 does not establish that highly ranked cases have stronger causal effects on the fitted margin under removal. When the authors remove the ten highest-ranked supporting cases and refit their readout, keeping the neural representation fixed, the margin reduction is not statistically distinguishable from a random removal approach. This also contributes to the previous point: the experiment does not provide additional evidence that the proposed ranking better identifies cases where removal changes the fitted margin.

Questions for authors

Q1: What happens when you take higher-dimensional representations for ϕ(x)\phi(x)? Is the training time still manageable? Is the trend of the reported metrics increasing or decreasing?

Q2: How do you distinguish preservation of the original case contributions from preservation of their summed scores? (For example, consider three cases with r1=(2,0)r_1=(2,0), r2=(0,2)r_2=(0,2), and r3=(1,1)r_3=(1,1), with true relevance weights ω(x)=(x1,x2,x3)\omega(x)=(x_1,x_2,x_3). A network could retain only the action scores, ϕ(x)=(2x1+x3,2x2+x3)\phi(x)=(2x_1+x_3,2x_2+x_3). The inputs (1,1/2,0)(1,1/2,0) and (1/2,0,1)(1/2,0,1) have identical representations and scores (2,1)(2,1), although their generating case weights differ. Those weights ω(x)\omega(x) cannot be recovered from the representation, yet an exact fitted decomposition and the decision-consistency condition can still hold. What justifies calling the case structure ‘’preserved’’ in this situation?

5 comments
AuthorsOpenReview ↗

W1. What the case decomposition establishes

Thank you for your careful reading and for recognizing the potential of the case-based perspective and the decision-agreement guarantee. Your comments have helped us clarify what our case decomposition establishes and distinguish the fitted case contributions from the original contributions during network learning. We address each point below and provide the corresponding manuscript revisions and supporting results.

W1. What the case decomposition establishes

Thank you for pointing out that the uniqueness of the coefficients under our rule does not establish that they are the original case contributions retained during network learning. As your case shows, different case contributions can produce the same scores.

Our result is a case decomposition of the fitted action scores on the network's trained representation. It determines how the fitted preference is allocated across individual training-case returns under a specified coefficient rule. Specifically, for a current situation xx and a training case xix_i, Eq. (3) defines αi(x)=ϕ(xi)⊤G−1ϕ(x),G=Φ⊤Φ\alpha_i(x)=\phi(x_i)^\top G^{-1}\phi(x),\quad G=\Phi^\top\Phi. Here, ϕ(xi)\phi(x_i) and ϕ(x)\phi(x) are the representations produced by the trained network for the training case and current situation, respectively. Thus, each coefficient is determined by how the network represents the current situation in relation to each past case. Because it does not depend on the action, the same coefficient applies across actions. Under full-column-rank OLS, Proposition 1 further establishes that these coefficients give the unique minimum-Euclidean-norm solution satisfying the representation constraint.

Substituting these coefficients into Eq. (4) gives Q^(x,a)=∑i=1nαi(x)ri(a).\widehat Q(x,a)=\sum_{i=1}^{n}\alpha_i(x)r_i(a). Consequently, the fitted action scores are expressed through the returns of individual training cases. This makes it possible to inspect which past cases support or oppose a fitted choice and by how much, connecting the network's learned representation to the CBDT view of decisions formed from relevant past experience. Section 4 further checks when this fitted choice agrees with the original network. Together, the representation-based coefficients and their decomposition of fitted scores provide a case-level explanation of how the network's learned preference is supported by past returns.

To clarify the meaning of case structure, we will revise its definition in Section 2.3 as follows:

"To study this relationship, we fit a separate linear return model for each action on the trained representation. Here, case structure refers to the allocation of fitted action-score differences across training-case return differences, using coefficients determined by the trained representation and the fitting rule. This shows which past cases support or oppose the network's fitted preference."

We will also change the title of Proposition 1 to “Case decomposition of fitted decision scores” and clarify the paragraph following it in Section 3.2 as follows:

"Equation (4) expresses each fitted decision score as a sum of training-case returns using coefficients common across actions. Under full-column-rank OLS, Proposition 1 shows that α(x)\alpha(x) is the unique minimum-Euclidean-norm vector satisfying Φ⊤α(x)=ϕ(x)\Phi^\top\alpha(x)=\phi(x). The corresponding ridge decomposition used in the experiments is given in Appendix A.3."

To clarify what Table F1 (originally Table 2) evaluates, we will identify its reference as the specified Gaussian input-based relevance-and-return construction. The table compares the case contributions with this reference, while Eqs. (3) and (4) and Proposition 1 establish how the allocation is obtained from the trained representation. The comparison is retained in Appendix F.1 as follows:

"Case-contribution correspondence with the specified Gaussian input-based relevance-and-return reference over 500 sampled training cases."

Appendix E.7 provides a further check of whether the case decomposition can give a shorter explanation of the fitted choice. Cases are ordered by decreasing ∣αλ,i(x)∣|\alpha_{\lambda,i}(x)|, and the reported prefix is the shortest one after which every longer prefix selects the same action as the complete fitted score. This shows how many leading cases are needed to retain the fitted choice, linking the individual case contributions to a concise explanation of the network's decision.

To further clarify what the case structure preserves, we examine how it behaves when case sets are combined. Under a fixed representation and full-column-rank OLS, proportional Gram matrices make each set’s case contributions scale by a positive factor, preserving their supporting or opposing signs and preferences supported by both sets (Section 5, Appendix A.8). This establishes when the case-based explanation retains common preferences after pooling, extending the decomposition beyond an individual case set.

AuthorsOpenReview ↗

W2. Contribution and effects under removal & Q1. Higher-dimensional representations and computation

W2. Contribution and effects under removal

Thank you for pointing out the need to clarify what the removal experiment establishes. This raises an important distinction between a case’s contribution to the current fitted preference and the change in that preference after the case is removed. In Eq. (5), ΔQ^(x;a,b)=∑i=1nαi(x)Δri(a,b)\Delta\widehat Q(x;a,b)=\sum_{i=1}^{n}\alpha_i(x)\Delta r_i(a,b), each case contributes a term determined by its coefficient and its return difference between two actions. The sum of these terms gives the original fitted margin. Thus, the first contribution identifies how individual past returns support or oppose a fitted choice and quantifies their respective contributions. The removal experiment examines how the fitted margin changes after cases are removed and the model is refitted. We will clarify these two roles so that the case decomposition and its practical value are established independently of the removal analysis.

Introduction, first contribution (original p. 2, lines 098–099):

First, we connect trained representations to the CBDT view by decomposing fitted preferences into individual training-case contributions, revealing how past returns support or oppose each choice. Since CBDT relevance requires nonnegative weights, we show why feature-only reconstruction can change fitted choices and develop joint fitting to reduce these changes while improving task-optimal accuracy.

Section 3.3, explanation of case contributions following Eq. (5) (original p. 5):

The sign and magnitude of each case contribution show whether and how strongly it supports one action over another in the fitted model. Together, these contributions show how past return differences form the fitted preference. Section 4 specifies when this fitted preference selects the same action as the original network.

Appendix F.2, clarification following the existing removal results (original p. 31):

Removing the ten highest-ranked supporting cases and refitting gives a normalized margin reduction of 0.0019 [−0.0044, 0.0082], with a difference of 0.0015 [−0.0048, 0.0078] relative to random removal. This comparison reports how the fitted margin changes after removal and refitting, complementing Eq. (5), which decomposes the original fitted margin into individual case contributions.

Q1. Higher-dimensional representations and computation

Thank you for asking how representation dimension affects both the reported results and computation. Increasing the dimension enlarges the matrices used to calculate case coefficients, making it important to examine how the reported metrics and computation change at higher dimensions. In response, we have conducted experiments at dimensions 16, 32, 64, and 128, each with five training seeds, and measured both the reported metrics and computation time. The results will be included in Appendix C.3 as follows.

Appendix C.3, new Table C4.

Table C4. Effect of representation dimension on score reconstruction, case correspondence, decision consistency, and computation. Values are five-seed means in the controlled setting with 500 training and 200 test cases. Case-reference comparisons use the same 96 selected test cases per seed across dimensions. Per-case times are means of within-seed medians measured on one thread. Training times correspond to the stopping iterations shown.

Measure163264128
Score reconstruction R2R^20.99980.99970.99980.9996
Correlation with generating support0.99620.99760.99930.9995
Top-30 Jaccard with the specified reference0.39030.31530.22860.1845
Decision-consistency coverage (%)94.393.393.192.3
Fitted–network action agreement (%)98.898.798.497.7
Training time (s)1.131.261.391.67
Mean stopping iterations1945190816271366
Feature extraction through case ranking (ms per case)0.07420.07480.08050.0884

The results show that score reconstruction and decision-consistency coverage remain high as representation width increases, while correspondence with the specified case reference declines. Training remains below two seconds in the tested 500-case setting. These short runtimes do not establish feasibility at substantially larger scales.

The corresponding addition to Section 6.1 is as follows:

Regarding higher-dimensional representations and computation, score reconstruction remains high and decision-consistency coverage stays above 92% across widths of 16–128, while correlation with generating support increases despite a decline in Top-30 correspondence with the specified case reference. Training takes less than two seconds in the tested 500-case setting (Appendix C.3).

AuthorsOpenReview ↗

Q2.1 The counterexample and generating case weights

Q2. The counterexample and generating case weights

Thank you for this counterexample. It helps clarify precisely what can be obtained from a trained representation. We will use case structure for the decomposition of fitted preferences under the specified coefficient rule, and revise the statements that suggested recovery of the original contributions during network learning. The same scores (2,1)(2,1) can arise from different generating weights, with margin contributions (2,−1,0)(2,-1,0) and (1,0,0)(1,0,0). When their representations are identical, our coefficient rule cannot distinguish those generating weights.

This does not prevent the trained representation from defining a case structure under our coefficient rule. As shown in Eq. (3), αi(x)=ϕ(xi)⊤G−1ϕ(x)\alpha_i(x)=\phi(x_i)^\top G^{-1}\phi(x), the coefficients are determined by the learned representations of the current situation and training cases. Equation (4), Q^(x,a)=∑i=1nαi(x)ri(a)\widehat Q(x,a)=\sum_{i=1}^{n}\alpha_i(x)r_i(a), then uses these coefficients to decompose the fitted scores into individual case returns. Proposition 1 specifies the allocation under the OLS constraint, and Section 4 establishes when the resulting fitted preference selects the same action as the original network. Thus, the structure established here is the representation-based allocation of fitted decisions to past case returns under the specified fitting rule. It provides an inspectable explanation of how past returns support or oppose a fitted choice. Section 4 determines when that choice agrees with the original network, and Section 3.4 separately states the condition for interpreting the case weights as nonnegative relevance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.