acceptodds
Under review as a conference paper at ICLR 2027

Beyond Feature Subspaces: Query-Conditioned Projection Learning for Outlier Interpretation

Abstract

Most outlier interpretation methods answer why a point is anomalous by returning a feature subspace: an axis-aligned subset of the observed coordinates in which the query stands out. We show that this format can be expressively inadequate, but in a way that depends on dimensionality and on the search budget in a manner no method can infer a priori. Enumerating every non-empty coordinate subspace of all ADBench Classical benchmarks, we find that of outliers admit no subspace that isolates them from every inlier when , where only – candidates exist, yet when , where exhaustive search succeeds. Beyond exhaustive search is impossible (MNIST: subsets), and under the cardinality budget that any practical search must adopt, of outliers remain unseparated for . Pooled over all benchmarks ( outliers) the rate is under the strict criterion and under a th-percentile reference. Where the search is exhaustive the failures cannot be attributed to pruning or heuristics. We therefore study TMQ, whose parameters are re-estimated for that query alone, so the contrast may be a weighted feature combination rather than a coordinate subset. Sampled quadruplets (query, other outlier, near inlier, far inlier) are embedded; a global attention branch and four leave-one-position-out local branches estimate latent contributions; and three margin constraints organize inlier–inlier, query–outlier and query–inlier distances. We characterize the exact zero-loss distance interval and the subtype condition under which its inter-outlier lower bound is useful. Extensive experiments across twelve benchmarks, nine baselines and seven detector references show that TMQ attains the best pooled mean on six of seven metrics.

Primary area:
interpretability and explainable AI
Keywords:
Outlier Interpretation, Latent Space, Quadruplet Attention

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nFNUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies instance-level outlier interpretation and questions whether axis-aligned feature subsets adequately express the contrast between an anomalous query and normal reference observations. It first measures subspace-separability failures across 47 ADBench datasets, using exhaustive enumeration for datasets with at most 20 features and restricted search budgets at higher dimensionalities. It then proposes TMQ, which learns a separate linear projection for each query using quadruplets consisting of the query, another outlier, a near inlier, and a sampled non-near inlier. One global and four leave-one-position-out attention branches estimate latent weights, while a three-hinge objective imposes a hierarchy among inlier–inlier, query–outlier, and query–inlier distances. The learned projection is subsequently mapped to an input-feature ranking.

Critical strengths
  • A substantial comparative record with promising reported results. The paper provides dataset–detector-level results, component ablations, parameter sensitivity, and resource measurements rather than relying only on a pooled average.

  • An explicit query-specific contrast with a clearly characterized relational objective. TMQ makes its reference structure explicit: the query is compared with both normal observations and another detected anomaly. The global and omission-context branches provide an understandable way to condition the learned projection on different reference configurations.

Critical weaknesses
  • The most immediate issue concerns the relationship between precision and F1. Appendix G.1 states that a weight-based method is converted to a predicted feature set PP by selecting the top k=∣G∣k=|G| features, where GG is the reference set. It explicitly notes that precision equals recall because ∣P∣=∣G∣|P|=|G|.
    Under that protocol, for each query with a nonempty reference set, precision=recall=F1. Thus, these quantities must also coincide when calculated from the same query-level predictions using the same aggregation. However, Table 3 reports TMQ precision of 0.616, whereas Table 4 reports F1 of 0.571, despite describing a matching pooled dataset–detector aggregation.

  • The specific contribution of TMQ—and its claimed move beyond feature subsets—is not adequately isolated. The manuscript itself notes that ATON already uses query-conditioned embedding learning with feature-weight readout, while WANDER already uses quadruplets and attention. The main incremental claim therefore concerns TMQ’s particular global/local attention parameterization. However, the WANDER comparison changes complete pipelines rather than holding the quadruplets, objective, optimizer, projection dimension, and readout fixed. The ablations provide partial evidence: the parameter-matched MLP replacement is useful, and its performance drop is substantial. Nevertheless, removing the global or local branches changes AUPR by only 0.004 or 0.002, respectively. The largest drop comes from replacing the quadruplet objective with a triplet objective, but this does not isolate the distinction from WANDER, which also uses quadruplets.

Questions for authors

See Weaknesses.

7 comments
AuthorsOpenReview ↗

Beyond Feature Subspaces: Query-Conditioned Projection Learning for Outlier Interpretation

Response to Reviewer nFNU: Q1

The precision in Table 3 and the F1F_1 score in Table 4 are reported for feature-weight evaluation and subspace evaluation, respectively. As defined in the manuscript's Evaluation Metrics paragraph (Appendix G.1), G\mathcal{G} denotes the ground-truth subspace and P\mathcal{P} the predicted subspace. We follow the distinction in Xu et al., \emph{Beyond Outlier Detection: Outlier Interpretation by Attention-Guided Triplet Deviation Network} (WWW 2021, Secs. 5.1.2 and 5.1.5).

In the feature-weight setting, each method produces feature importance weights. We obtain P\mathcal{P} by selecting the top-kk features with k=∣G∣k=|\mathcal{G}|, and report precision =∣G∩P∣/∣P∣=|\mathcal{G}\cap\mathcal{P}|/|\mathcal{P}| and Jaccard, together with AUPR and AUROC for the full ranking. Recall equals precision because ∣P∣=∣G∣=k|\mathcal{P}|=|\mathcal{G}|=k, so recall is redundant in this comparison; F1F_1 is not a reported metric in Table 3.

In the subspace setting, P\mathcal{P} is the predicted explanation subspace, and Table 4 reports F1=2∣G∩P∣/(∣G∣+∣P∣)F_1=2|\mathcal{G}\cap\mathcal{P}|/(|\mathcal{G}|+|\mathcal{P}|) without separately reporting precision and recall. Xu et al. report all three metrics for subspace evaluation methods and distinguish ATON's feature-weight output from its thresholded subspace version ATON$'$. The identity precision == recall =F1=F_1 holds for the same pair of sets when ∣P∣=∣G∣|\mathcal{P}|=|\mathcal{G}|. Applying it across Tables 3 and 4 additionally requires identical predictions, queries and aggregation. The original query-level predictions and exact selection configuration for Table 4 are unavailable, so the Type-I/Type-II distinction does not by itself verify the reported 0.616/0.571 difference. The supplementary TMQ-S evaluation checks the identity on saved predictions but does not reconstruct those historical entries. The revisions in Secs. 3.4, 4.1 and 4.2, the table captions, and Appendices G and K.1 now distinguish these protocols.

Response to Reviewer nFNU: Q2

TMQ and WANDER differ in representation learning and optimization. WANDER (Guan et al., Information Fusion 2025) builds a weighted digraph to encode each quadruplet's geometry, learns a multi-kernelized embedding, and optimizes a weighted-digraph-based quadruplet loss. TMQ learns an explicit linear projection W∈Rs×dW\in\mathbb R^{s\times d} with latent-coordinate weights a∈Rsa\in\mathbb R^s. Its three-hinge loss L(q)L(q) directly constrains DiiD_{ii}, DqoD_{qo} and DqiD_{qi}. Following Proposition 1, L(q)=0L(q)=0 requires Dqi≥Dii+2mD_{qi}\ge D_{ii}+2m and Dqo∈[Dii+m, Dqi−m]D_{qo}\in[D_{ii}+m,\,D_{qi}-m], where m>0m>0 is the margin. Thus TMQ specifies the desired distance ordering in the learned projection, while WANDER uses weighted-digraph guidance in a kernelized representation (Secs. 3.2--3.3).

The attention design differs in how the representation is weighted. WANDER integrates customized self-attention with its multi-kernelized embedding. TMQ estimates aa through an explicit global/local decomposition of q=(x,xo,xi,xi′)q=(x,x_o,x_i,x'_i): one global multi-head branch receives all four embedded positions, and four local multi-head branches each omit a different position. Omitting xox_o exposes the query--normal contrasts; omitting xx isolates relations among the reference samples; omitting either inlier changes the normal context. The branch scores are normalized and averaged across the sampled quadruplets to obtain aa, so the latent-coordinate weights reflect the complete quadruplet and the separate reference contexts (Sec. 3.2).

The explanation output is also different. WANDER extracts an outlying subspace, namely a subset of the original features. TMQ returns the query-conditioned operator Bx=diag⁡(ax∗)Wx∗B_x=\operatorname{diag}(a_x^*)W_x^*, with ϕx(z)=Bxz\phi_x(z)=B_xz (Sec. 3.4). Each row of BxB_x describes a signed linear combination of input features. For a normal reference bb, the contrast ϕx(x)−ϕx(b)=Bx(x−b)\phi_x(x)-\phi_x(b)=B_x(x-b) shows how feature deviations reinforce or cancel one another in the learned projection. A coordinate subset identifies selected features but does not retain these coefficients or their signed combinations. Appendix K.1 provides the complete learned operator and a worked query--reference contrast to make this output distinction concrete.

The matched TMQ-S attention controls in Sec. 4.5 do not show a combined-branch advantage; replacing TMQ's attention with WANDER's in the same backbone remains untested. The architectural differences therefore do not establish an isolated performance advantage.

Thank you for your careful review and detailed feedback, which helped us clarify the evaluation protocols and distinguish architectural differences from evidence of their contribution. If these clarifications and the matched controls address your concerns, we would be grateful if you would reconsider your assessment and revise your score upward. We sincerely appreciate your time and consideration.

I thank the authors for their detailed responses. The authors have resloved most of my concerns. My remaining concerns are as follows:

  • Validity of the reference explanations. The reference subset is selected by maximizing the query’s min–max-normalized anomaly score under separately fitted subspace detectors. However, this criterion does not account for how many normal observations receive scores above the query. Consequently, a subset may be selected even when it poorly distinguishes the query from normal behavior. This issue can occur with a single anomalous query and therefore cannot be explained solely by heterogeneous anomaly subtypes. Please validate the selected reference subsets using query-versus-normal ranks or held-out false-positive rates, accounting for subset selection. More fundamentally, agreement with these binary labels measures agreement with a particular annotation procedure, rather than independently validated explanation accuracy. The distinction is especially relevant when evaluating SHAP/LIME, whose native objective concerns the behavior of a specified predictive model.
  • The objective function is too similar to WANDER, which leads to limited novelty.

Since feature shapley and outlier interpretation is not my direct research field, i would also like to see other reviewers' feedback.

AuthorsOpenReview ↗

Beyond Feature Subspaces: Query-Conditioned Projection Learning for Outlier Interpretation

Response to Reviewer nFNU: Q1

We evaluated the HBOS reference generator used by \texttt{eval/eva_main.py} across all twelve datasets. The original exhaustive search reproduces all 3,863 current local reference subsets; Appendix K.5 reports the validation protocol and results.

The query-versus-normal rank check confirms your concern: 1,092/3,863 queries (28.27%) have more than 5% of normal observations scoring above them in their selected subsets. We report both strict and tie-inclusive normal ranks. Maximizing the normalized score therefore does not guarantee strong separation.

We then tested the selection rule with five splits per dataset, each separating fitting, calibration and test observations. At nominal 5%, the dataset-average held-out false-positive rate is 20.75% when thresholding the selected subset alone and 3.88% when calibrating the maximum over the entire search. Anomaly detection also falls from 55.46% to 23.71%. Thus accounting for subset selection materially changes the results.

This validation covers HBOS; the other reference generators remain unvalidated by these tests. Appendix K.5 also separates the archived and current PageBlocks labels, which differ for 508/510 queries. Tables 3 and 4 retain their interpretation as agreement with detector-derived proxies, without establishing independent explanation accuracy or superior SHAP/LIME fidelity.

Thank you for helping us clarify reference validity and model fidelity. If these clarifications and the supplementary evidence address your concerns, we would be grateful if you would reconsider your assessment and revise your score upward. We sincerely appreciate your time and consideration.

Response to Reviewer nFNU: Q2

Thank you for raising this point. TMQ differs from WANDER in the constraints imposed by its objective, the explanation it seeks, and the computation used to obtain that explanation.

First, the objective functions impose different distance constraints. In common squared-distance notation, the per-quadruplet terms of \href{https://pure.ulster.ac.uk/ws/portalfiles/portal/218009379/InformationFusion_2025PreProof.pdf}{WANDER's Eq. (8)} are

LW(q)=g1[Dqo−Dqi′+m]++g2[Dii−Dqi+m]++g3[Dii−Dio+m]+.L_{W}(q)=g_{1}[D_{qo}-D_{qi'}+m]_{+}+g_{2}[D_{ii}-D_{qi}+m]_{+}+g_{3}[D_{ii}-D_{io}+m]_{+}.

Here distances are measured in each method's learned embedding, DiiD_{ii} is the distance between the two inliers, g1,g2,g3g_1,g_2,g_3 are WANDER's digraph-derived weights, Dqi′D_{qi'} compares the query with the second inlier, and DioD_{io} compares the first inlier with the other outlier. Writing both margins as m>0m>0 for comparison, TMQ instead uses

LT(q)=[Dii−Dqo+m]++[Dii−Dqi+m]++[Dqo−Dqi+m]+.L_{T}(q)=[D_{ii}-D_{qo}+m]_{+}+[D_{ii}-D_{qi}+m]_{+}+[D_{qo}-D_{qi}+m]_{+}.

TMQ therefore explicitly bounds the query--other-outlier distance: zero loss requires Dqo∈[Dii+m,Dqi−m]D_{qo}\in[D_{ii}+m,D_{qi}-m]. WANDER's objective does not impose this lower bound. For example, with one-dimensional embedded positions (x,xo,xi,xi′)=(0,0,2,3)(x,x_o,x_i,x_i')=(0,0,2,3) and m=1m=1, all three WANDER hinges vanish, whereas TMQ's first hinge equals 2. Thus the losses have different zero-loss sets. Proposition 1 and Appendix K.2 also identify and test when this additional constraint conflicts with same-subtype explanations.

Second, the explanation targets differ. WANDER optimizes a kernelized embedding to extract an outlying subset of the original features. TMQ learns and returns the query-conditioned operator Bx=diag⁡(ax∗)Wx∗B_x = \operatorname{diag}(a_x^{\ast}) W_x^{\ast}, with an explicit signed contrast Bx(x−b)B_x(x-b) relative to a normal reference bb. The target is a weighted relation among features that characterizes the query's departure from normal behavior. Retaining the operator exposes reinforcing and compensating contributions that a feature subset does not specify. Appendix K.1 provides a complete learned operator and a worked contrast; feature rankings are an evaluation readout of this explanation.

Third, TMQ avoids WANDER's weighted-digraph construction and multi-kernel feature computation. It operates directly on an s×ds\times d projection and sampled quadruplets, removing those kernel computations when processing higher-dimensional inputs. Projection work is O(qds)O(qds) for qq quadruplets; attention adds O(qs2+q2s)O(qs^2+q^2s) under the formulation in Appendix F, so the full cost also depends on the latent width and sampling/batching budget. The reported benchmark provides a concrete scalability result: TMQ completes fraud (n=284,807n=284{,}807, d=29d=29), while WANDER exceeds the two-week budget (Table 4). This supports a practical difference in the tested setting; a dimension-specific advantage requires a matched high-dimensional runtime comparison.

These distinctions concern the mathematical constraints, the information returned to the analyst, and the computational formulation. We welcome the other reviewers' assessments of these points. We would be grateful for your kind consideration of our revised work.

Reviewer_1jq6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that existing outlier interpretation methods are limited because they typically explain anomalies by identifying subsets of features. However, some anomalies may only become apparent through combinations of features that cannot be captured by such subsets. To address this limitation, the authors propose TMQ, which learns a query-specific projection to better distinguish an anomalous observation from normal ones and identify the features contributing to its abnormality.

Critical strengths

S1. Important Research Problem. -- Explaining why a particular case counts as an anomaly is of key importance in many domains. This paper addresses a practical problem that goes further than the accuracy of anomaly detection.

S2. Promising direction for outlier interpretation. -- The idea of looking at combinations of features and how their interactions could explain normal and abnormal behavior could be interesting to explore further, even though the proposed method still has some limitations.

Critical weaknesses

W1. Lack of detector fidelity. -- TMQ uses the labels generated by an unsupervised detector to explain in the next step why a data point is considered abnormal. However, it applies its own separation criterion to understand the reasoning behind this decision. Since it does not consider how the detector actually makes its decision, the explanation provided by TMQ and the detector's decision may not agree. So the connection between both is missing. Although the authors evaluate the agreement with feature reference proxies derived from the detector, this still does not show to what degree the explanations actually reflect the detector's scoring function.

W2. Missing justification for the separation criterion. -- The paper uses a nearest-neighbor separation rule to check whether feature subsets can distinguish between normal and abnormal data points. A feature subset can still be meaningful even if it doesn't pass the test. The result could be an artifact of the chosen threshold or the chosen criterion. The paper lacks a clear justification for why this specific criterion is enough to determine whether a feature is useful or not for explaining a nearest-neighbor anomaly. To my understanding, nearest-neighbor is one of the weaker approaches to providing an explanation, and there are methods that capture much richer structures in the data and could explain why a feature is important, while the nearest-neighbor approach would classify it as unimportant or not informative enough.

W3. Unclear behavior with conflicting detectors. -- Different anomaly detectors can classify different data points as normal or abnormal. The question here is what happens if two detectors produce completely different labels. For example, one detector could classify a point as normal, while another considers the same point as abnormal. This means that TMQ could provide completely different or even contradictory explanations for the same data point. How to deal with such contradictions is not clearly explained in the paper. For example, if a practitioner uses two different detectors that disagree, which explanation should they trust?

W4. Missing justification for design choices. -- In the ablation study, the paper shows that one of its core contributions, the attention branches, provides only a small performance gain, while the quadruplet objective has a much larger impact on the overall performance. To my understanding, the quadruplet objective is already established in the literature. This makes it unclear to what degree we actually need the Global/Local Attention architecture. Stronger ablation studies would help to better motivate this design choice and make the contribution more novel than just applying already established objectives.

Questions for authors

Please explain how your answers will be reflected in the paper (just educating reviewers in this private discussion room is not sufficient):

Q1. TMQ does not directly use the assumptions or the scoring function of the underlying anomaly detector. This raises the question of how both can actually contribute to the same explanation. For example, Isolation Forest might detect anomalies based on completely different patterns than TMQ uses for its explanations. How can we then be sure that both are actually explaining the same thing? This becomes even more important when two anomaly detectors completely disagree like LOF and KNN. In this case, a user could get two different or even contradictory explanations for the same data point. How can we then justify which explanation is meaningful and which one is not?

Q2. I'm currently not convinced by the nearest-neighbor criterion since, to my understanding, it provides a weak contrast, especially in high-dimensional spaces. By evaluating whether a subspace is useful or not based on only one criterion, we may miss important information that is still present in the subspace. Have the authors also considered other tests to understand whether a subspace can provide a meaningful explanation or not? It would also be interesting to see how dimensionality (d > 1000), different thresholds, or the size of the subspace affect the results. This would help to better understand when the criterion is useful and when it might fail. Such a discussion is currently missing in the paper.

6 comments
AuthorsOpenReview ↗

Beyond Feature Subspaces: Query-Conditioned Projection Learning for Outlier Interpretation

Response to Reviewer 1jq6: Q1

TMQ performs detector-agnostic, post-hoc outlier interpretation. Its input is a dataset whose samples are already labeled as normal or anomalous. Given a flagged query, TMQ learns a projection that characterizes its deviation relative to normal and other anomalous reference samples. The detector's internal assumptions, architecture and scoring function are therefore not required inputs: TMQ operates on the data and supplied labels after detection. This allows the same interpretation procedure to be applied across detectors.

The explanation remains conditional on the supplied labels and reference population. The original benchmark uses supplied dataset labels; the supplementary fidelity test uses each fitted detector's labels, as clarified in Sec. 3.1. When detectors disagree, we interpret each flagged query under the corresponding labeling and retain the disagreement. Among held-out rows, 65 are flagged by all three detectors and 321 by a nonempty proper subset. Selecting among conflicting interpretations requires domain or subtype evidence. Fidelity to a particular detector's score is a separate empirical question.

We additionally tested score alignment on six datasets using frozen iForest, LOF and 20-NN detectors fitted on a shared 80% training split. TMQ-S explains 110 distinct held-out queries (208 detector-query pairs; five seeds) using training-only reference pools. Eight common held-out normal donors replace feature groups without refitting detectors or thresholds. Deletion gains over random ordering are +0.044, +0.061 and +0.052, with positive paired dataset-bootstrap 95% intervals; local Spearman correlations with frozen scores are 0.621, 0.681 and 0.675. Sec. 4.5 and Appendix K.3 report these results and the complete protocol. They demonstrate empirical alignment under the tested normal-replacement perturbations, while the TMQ objective itself remains label-conditioned.

Response to Reviewer 1jq6: Q2

The original submission already evaluated high-dimensional data: Table 24 (Appendix H) includes InternetAds with d=1555d=1555, using its original input coordinates. It reports 1-NN separability failure over 30,000 sampled subsets with ∣S∣≤3|S|\le3: 16.8% at the maximum-inlier threshold and 13.6% at both q99 and q95. Thus high-dimensional coverage and multiple thresholds were present before discussion.

During discussion, we added controlled studies, summarized in Sec. 4.5 and detailed in Appendix K.4: a synthetic dimension sweep at d=10,100,500,1000,2000d=10,100,500,1000,2000 over two anomaly mechanisms (five seeds each); comparisons of 1-NN, mean 5-NN and LOF-5; and calibration of the maximum score over searched subspaces using independent normals. The new InternetAds study uses five splits and 9,555 candidates of sizes 1, 2, 3, 5 and 10. At q99, unadjusted/calibrated detection is 94.5%/29.4%, with normal false-positive rates of 72.7%/0.16%. This is a separate protocol from Table 24.

We also added search-budget and maximum-subspace-size comparisons. On InternetAds at unadjusted 1-NN q99, increasing the maximum size from 1 to 10 changes detection/normal false-positive rates from 89.5%/35.2% to 94.5%/72.7%, while eligible candidates increase from 1,555 to 9,555. This measures the joint effect of size and search breadth, not size alone at fixed candidate count. Both Table 24 and these new studies evaluate isolation criteria and finite subspace searches; they do not measure TMQ's high-dimensional interpretation accuracy. Their conclusions remain conditional on the tested criterion and candidate family.

Thank you for your careful review and constructive questions, which helped us clarify the relationship between detector labels and explanations and examine the isolation criterion across dimensions, thresholds, and search settings. If the revised explanations and additional experiments address your concerns, we would be grateful if you would reconsider your assessment and revise your score upward. We sincerely appreciate your time and consideration.

Reviewer_1jq6OpenReview ↗

Reply to Authors' Rebuttal

I would like to thank the authors for their detailed response and the additional experiments. However, several of my main concerns remain unsatisfied.

Regarding Q1 -- I understand that TMQ is detector-independent. However, TMQ uses its own target and does not take the detector’s decision-making process into account. While the additional experiments indicate a certain degree of agreement with the detector values, But this agreement is not strong enough to confirm that TMQ actually explains the detector’s decision. In particular, KNN and LOF are based on different assumptions. And so, they can yield conflicting results. If this happens, TMQ cannot determine which interpretation is more reliable. The reference to expert knowledge does not explain what specific added value TMQ offers in such circumstances.

Regarding Q2 -- The revision still leaves me without a fundamental justification for why the “nearest-neighbor” approach is appropriate for judging whether a feature subspace provides a valid explanation. Because the comparisons with 5-NN and LOF still rely on neighborhood information, they do not fully address this concern. In addition, a feature subspace may contain important information even if it does not satisfy the chosen criterion. This limitation is now acknowledged by the authors themselves. Therefore, I remain unconvinced for Q2.

Regarding W4 -- The additional ablation experiments actually reinforce my original concerns. In the controlled TMQ-S experiments, the combined global/local attention architecture shows no clear advantage over the individual branches, while a simpler MLP performs slightly better. In addition, performance improves when the first term of the quadruplet objective function expression is omitted. While these experiments are based on a separate implementation, they raise further questions regarding the necessity of the proposed architecture, which is a central contribution of the work.

Overall -- I welcome the additional experiments and clarifications. I still believe, that some fundamental issues of design have not been sufficiently addressed. In my opinion, these points require more comprehensive revisions and explanations. Therefore, I maintain my original evaluation.

AuthorsOpenReview ↗

Reply to Reviewer 1jq6's comment

Dear Reviewer 1jq6:

Thank you for the follow-up.

W4: What the controlled ablations support.

We agree with your interpretation of the matched TMQ-S controls. The full-minus-first-hinge-removal AUPR difference is −0.034, with a paired 95% interval of [−0.060, −0.011]. The matched MLP is also slightly better: the full-minus-MLP difference is −0.004 [−0.007, −0.001]. The branch comparisons do not establish an advantage for combining global and local attention. These results do not support claiming that the complete architecture or the first hinge is necessary under the tested protocol.

The separate implementation limits extrapolation to the historical runs, but it does not invalidate these internally matched comparisons. Similarly, the original triplet replacement in Table 5 changes more than the first hinge and cannot establish that hinge’s individual benefit.

Proposition 1 provides a conditional explanation of the objective’s behavior, rather than a superiority guarantee. The first hinge requires Dqo≥Dii+mD_{qo}\geq D_{ii}+m, which can conflict with keeping anomalies from the same mechanism close. The known-subtype experiment supports this concern: with same-subtype partners, full/removal support F1 is 0.694/1.000. The different-subtype result, 1.000/0.997, does not establish a substantial advantage for the full objective. We therefore regard these controls as evidence about when the proposed constraints are inappropriate, and do not use them to justify architectural necessity.

Q1: The explanation target and its concrete added information.

TMQ does not establish why a particular detector produced its score. Its target is an observational contrast between a flagged query and the supplied reference populations. The frozen-detector experiments in Appendix K.3 demonstrate alignment under the tested substitutions; they do not establish general fidelity to the detector’s decision process.

The additional information TMQ provides is an explicit signed feature combination. Appendix K.1 illustrates this on a glass query: although A4 has the largest magnitude-based importance, A0 contributes −0.163 along the displayed direction, while A4 contributes +0.024 and partially offsets that deviation. The complete signed contrast along that direction is −0.102. This exposes the direction of departure and compensating contributions that the feature ranking alone does not report.

This example illustrates the explanation object; it does not demonstrate detector arbitration or improved analyst utility. When KNN and LOF disagree, TMQ cannot determine which detector or interpretation is semantically correct. Our earlier reference to expert knowledge described the additional evidence needed for that judgment, rather than a capability of TMQ. The supplied labels, normal reference, and resulting contrast must remain explicit.

Q2: Isolation failure is not explanation invalidity.

We agree that a subspace can retain useful information while failing the chosen nearest-neighbor test. Exhaustive enumeration removes uncertainty about the search under that test; it does not validate the test as a necessary or sufficient condition for a meaningful explanation. Comparisons with 5-NN and LOF assess sensitivity within neighborhood-based criteria and do not resolve this conceptual issue.

The narrower motivation is that coordinate selection with Euclidean distance and a learned projection impose different geometries. For a pairwise difference Δ\Delta, their squared distances are respectively Δ⊤diag(1S)Δ\Delta^\top\mathrm{diag}(\mathbf{1}_S)\Delta and Δ⊤Bx⊤BxΔ\Delta^\top B_x^\top B_x\Delta; the latter permits cross-feature terms. This motivates learning and displaying a weighted contrast. It does not imply that the original coordinates lack the information, or that a richer scoring function within a feature subspace could not use it.

Accordingly, Table 1 supports claims about isolation under its specified criterion and searched family. It does not establish the absence of meaningful explanations in failed subspaces. Appendix K.5 additionally audits a non-neighborhood, histogram-based score, but that audit concerns reference quality and selection effects; we do not present it as validation of a universal explanation criterion.

To make these boundaries consistent throughout the manuscript, we will tighten the Abstract and Introduction, remove the inference that failed isolation means no useful explanation exists, and state beside Table 5 that the original ablations do not establish component necessity. The contribution we seek to retain is the explicit query-conditioned contrast, together with the analysis and empirical characterization of its assumptions and failure regimes. We recognize that this contribution must be assessed on its own evidence, without relying on broader claims about detector fidelity, subspace inadequacy, or architectural necessity.

Reviewer_yqFHOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a limitation of feature-subset explanations for outliers: a query may not be isolated from the normal reference in any coordinate subspace. The authors examine this on 47 ADBench datasets, distinguishing exhaustive search from restricted search. They then propose TMQ, which learns a linear projection separately for each query using quadruplets containing the query, another outlier, and two inliers. Global and leave-one-position-out attention branches weight the latent coordinates, and a three-hinge objective orders the relevant distances. The projection also produces an input-feature ranking. On twelve datasets, TMQ has the best reported pooled mean on six of seven metrics and the best mean F1 on eight datasets.

Critical strengths
  1. The separability study gives the paper a useful starting point. In particular, the low-dimensional cases where every coordinate subset can be enumerated show a limitation that cannot be fixed by a better search heuristic. I also appreciate that Table 1 reports the percentile-based criterion and the cases where exhaustive search succeeds. The evidence supports a specific limitation of isolation-based subspace explanations, and that is a worthwhile reason to consider weighted feature combinations.
  2. The empirical comparison covers a reasonably broad range of datasets and detector-derived references. TMQ improves pooled AUPR from ATON's 0.649 to 0.692 and pooled F1 from 0.550 to 0.571. The gains are not uniform, but they extend beyond a single favorable dataset. The loss analysis is also helpful: Proposition 1 states exactly which distance ordering is imposed and identifies the case in which the extra constraint can conflict with the desired explanation. This makes the method easier to assess than an attention architecture justified only by visualizations.
Critical weaknesses
  1. The connection between the proposed explanation and the main evaluation is still incomplete. Section 3.4 gives a column-norm feature readout, but explicitly leaves open whether this is the readout used for the reported scores. Section 4.1 similarly leaves the high-dimensional annotation procedure unresolved. These are central details: agreement with a detector-selected subset does not directly establish that a learned feature combination provides a better explanation. If PCA was used for the reference annotations, the coordinates being ranked also need to be identified. I would find the evidence more convincing with one worked example that shows the learned coefficients, query and reference values, and the resulting contrast that coordinate selection misses.
  2. The outlier sampling rule does not fully match the motivation for the first hinge. The method selects a nearby outlier, while the analysis says that separating the pair is useful when they represent different anomaly mechanisms. Nearby outliers could instead share the same mechanism, especially with the main setting k1 = 1. The paper acknowledges this issue, but the triplet ablation removes both additional relations and therefore does not isolate the effect of the lower bound on query-outlier distance. Its large performance drop is not enough to conclude that this particular repulsion term is beneficial.
Questions for authors
  1. Which feature readout and annotation search procedure produced Tables 3 and 4? Please specify whether the rankings are in the original coordinates or PCA coordinates, particularly for MNIST, and whether Eq. (4) is the actual evaluated readout. A worked example of the full projection-based explanation would help establish what the additional expressiveness buys the analyst.
  2. What happens if only the first hinge in Eq. (2) is removed, keeping the same quadruplets, attention network, and remaining two hinges? This would directly test whether forcing the query away from its nearby outlier helps, rather than attributing the full triplet-to-quadruplet gain to that constraint.
4 comments
AuthorsOpenReview ↗

Beyond Feature Subspaces: Query-Conditioned Projection Learning for Outlier Interpretation

Response to Reviewer yqFH: Q1 For the feature-weight comparison in Table 3, we use real_len: each query's predicted feature set contains the top-ranked coordinates matching its annotation size. Table 4 reports F1F_1 for the subspace comparison; its predicted explanation sets must be distinguished from this top-∣G∣|\mathcal{G}| evaluation readout. Annotations were generated using the root-level evaluation_od.py, which enumerates nonempty coordinate subsets, fits a detector separately on each subset, min-max normalizes its scores, and selects the subset maximizing the query's normalized score.

For MNIST, exhaustive search over its 100 input dimensions would require 2100−12^{100}-1 subsets. Following ATON's preprocessing protocol, we reduced MNIST to ten principal components before annotation and evaluation for Tables 3 and 4. Both rankings and annotations therefore refer to these PCA coordinates. For the separate high-dimensional separability studies (Appendices H and K.4), including InternetAds (d=1555d=1555), we retain the original input coordinates and search a finite candidate family, without claiming exhaustive coverage.

The explanation output is the operator BxB_x; the feature ranking is its evaluation readout. The implementation uses attention-weighted column ℓ1\ell_1 magnitudes, while Eq. (4) defines the alternative ℓ2\ell_2 readout. Appendix K.1 now includes a worked projection-based explanation, with the complete operator matrix, query and reference values, and signed projected contrast.

Response to Reviewer yqFH: Q2 We directly tested the query-outlier repulsion term by comparing LL with L−1=L−[Dii−Dqo+m]+L_{-1}=L-[D_{ii}-D_{qo}+m]_+. Quadruplets, network, initialization, optimizer, stopping budget and real_len readout are fixed; the other two hinges remain intact. This isolates the first constraint from the broader triplet-to-quadruplet change.

In the supplementary TMQ-S implementation, all 3,863 anomalous queries across twelve datasets, 81 detector-reference cells and five seeds give full/removal AUPR of 0.455/0.488 and F1F_1 of 0.365/0.403. The paired AUPR difference is -0.034 (95% dataset-bootstrap interval [-0.060, -0.011]). The first hinge does not improve agreement under this recipe.

When nearby outliers share a subtype, full/removal support F1F_1 is 0.694/1.000 while their projected squared distance DqoD_{qo} is 11.519/2.455. Thus stronger separation can damage recovery of their shared anomalous features. Different-subtype partners give F1F_1 of 1.000/0.997. These 1,280 synthetic fits support the conditional conflict in Proposition 1, not a general repulsion advantage. Sec. 4.5 summarizes the results, and Appendix K.2 includes this controlled example and its full experimental protocol. We therefore do not attribute the full triplet-to-quadruplet gain to the first hinge.

Thank you for your positive assessment and constructive questions, which helped us clarify the evaluation protocol and directly test the first-hinge constraint. If the revised explanations, worked example, and controlled results address your concerns, we would be grateful if you would consider increasing your confidence in the assessment and revising your score upward. We sincerely appreciate your time and consideration.

Thank you for the detailed response. The clarifications regarding the feature readout and PCA coordinates help address my questions about the evaluation protocol. I also appreciate the controlled ablation of the first hinge, which directly addresses my second concern and clarifies the limitations of the repulsion term. However, the evidence that the learned feature combinations provide better explanations remains limited. On balance, I maintain my original rating of 3

AuthorsOpenReview ↗

Reply to Reviewer yqFH's comment

Dear Reviewer yqFH:

Thank you for the follow-up and for explaining your remaining reservation. We appreciate that the readout/PCA clarifications helped address the evaluation questions and that the controlled first-hinge ablation directly addressed your second concern.

Regarding whether learned feature combinations provide better explanations, we would like to distinguish their additional descriptive content from evidence of improved explanation quality. In the worked glass example in Appendix K.1, A4 has the highest magnitude-based importance, but along the displayed projection direction, A0 contributes −0.163 while A4 contributes +0.024, partially offsetting the deviation. The total signed contrast is −0.102. This makes the direction of departure and compensating contributions explicit, whereas the feature ranking alone does not report them. We agree that this illustration does not by itself establish superior explanation quality or improved analyst utility.

A complementary evaluation, not discussed in our previous reply to you, is reported in Appendix K.3. On held-out queries from six datasets, with detectors fixed before evaluation, TMQ-S’s feature ordering yields deletion gains over random ordering of +0.044, +0.061, and +0.052 for iForest, LOF, and 20-NN, respectively, with positive paired dataset-bootstrap 95% intervals. These results provide evidence of alignment with detector scores under the specified normal-replacement perturbations. They evaluate the feature ordering, however, and do not directly establish that the signed combination is more useful than a feature-subset explanation.

We therefore recognize the remaining limitation you identify. The supported contribution combines a study of coordinate-subspace isolation under a specified criterion, an explicit query-conditioned signed contrast, and an analysis of when the proposed relational constraint conflicts with the intended explanation. We do not treat the worked example or the controlled ablations as establishing a general superiority claim.

We appreciate your continued consideration. If these clarifications and the complementary evidence increase your confidence in assessing the paper, we would be grateful if you could reflect that in your final confidence assessment, while retaining any reservations you consider warranted.

AuthorsOpenReview ↗

Looking Forward to Reviewer yqFH’s Response

Dear Reviewer yqFH,

Thank you again for your detailed review. As the discussion period is still ongoing, we would like to kindly ask whether our response addresses your concerns.

If any questions remain or further clarification would be helpful, please do not hesitate to let us know. We would be happy to continue he discussion

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.