Pith. sign in

REVIEW 4 major objections 4 minor 53 references

Are Representation Disentanglement and Interpretability Linked in Recommendation Models? A Critical Review and Reproducibility Study

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that in recommender models, representation disentanglement is strongly correlated with representation interpretability but not consistently with recommendation effectiveness.

desk verdict Careful reproducibility study with a useful null result, but the headline disentanglement–interpretability correlation is inflated by shared classifier construction. read the letter →

arxiv 2501.18805 v1 pith:5PSC4K6N submitted 2025-01-30 cs.IR

Ervin Dervishaj , Tuukka Ruotsalo , Maria Maistro , Christina Lioma This is my paper · ORCID
classification cs.IR
keywords disentangledrepresentationlearningrecommendersystemsinterpretabilityfeatureattributionLIMESHAPDCIreproducibility
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper reproduces the effectiveness, disentanglement, and interpretability of five well-known recommendation models, plus a popularity baseline, on four datasets, and asks whether the widely claimed benefits of disentangled user representations actually show up in measurements. It finds that disentanglement scores are not consistently or significantly related to recommendation effectiveness, so the alleged trade-off between disentanglement and accuracy does not appear in these models and datasets. It does find a strong positive correlation between disentanglement and two new interpretability measures built from LIME and SHAP feature attributions. The upshot is that disentanglement in recommender systems behaves as an interpretability property, not an accuracy property, at least under the ground-truth factors this study constructs.

What carries the argument

The argument runs through three measurement devices. First, the DCI framework's disentanglement and completeness scores: binary classifiers predict each ground-truth factor from the latent dimensions, and the resulting feature-importance matrix is reduced to entropy-based scores that reward concentrated, non-redundant encodings. Second, the paper's own interpretability measures, LIME-global and SHAP-global: they average per-dimension LIME and SHAP importance for each factor classifier, concatenate these into a matrix, and take the mean Jensen-Shannon divergence between its columns, rewarding sparse and non-redundant feature attributions. Third, repeated-measures correlation, which accounts for the grouped structure of runs within a model or dataset. The ground-truth factors themselves are proxies built from item tags or bookshelves, with users assigned to a factor when at least half of their interacted items fall in it. These devices together let the paper convert qualitative claims about interpretability into numbers and correlate them.

What would settle it

Recompute the same correlation using an alternative definition of ground-truth factors, for example explicit item genres or category splits instead of tag clusters, different cluster counts, or a different user-assignment threshold, and check whether the repeated-measures correlation between disentanglement and LIME-global or SHAP-global stays in [0.51, 0.95]. A second check is to train models with random latent dimensions or shuffled factor labels: if LIME-global and SHAP-global still correlate with disentanglement, the measures are tagging shared structure rather than semantic separation.

Watch

Extended reading notes

Core claim

The central finding is a positive correlation between representation disentanglement and representation interpretability, measured across a set of matrix-factorization, autoencoder, and variational-autoencoder models on Amazon-CD, ML1M, Yelp, and GoodReads-Children. Using the DCI disentanglement and completeness scores as the disentanglement measures, and new LIME-global and SHAP-global scores as interpretability measures, the repeated-measures correlation between them lies in [0.51, 0.95] across datasets and models. In contrast, the correlation between disentanglement and effectiveness (NDCG, recall, MRR, coverage) is inconsistent and often not statistically significant, which contradicts the common assumption of a trade-off. The paper also reports that reproducing previously published disentanglement scores is hard: effectiveness reproduces within 10% on ML1M but within 43% on GoodReads-Children, and the one prior disentanglement score set they attempted could not be reproduced despite following the description.

Load-bearing premise

The load-bearing premise is that the hand-built proxies, namely the 100 most popular tags clustered into 20 groups or merged bookshelves with users assigned when at least half their items match, capture the real factors of variation behind user choices; if they do not, every disentanglement and interpretability score, and the correlation between them, could be an artifact of the proxy construction.

Editorial extensions

If this is right

  • If disentanglement is primarily an interpretability property, then optimizing a recommender for disentanglement alone should not be expected to improve ranking accuracy; the two goals are decoupled in these experiments.
  • The absence of a consistent disentanglement-effectiveness trade-off means regularization for disentanglement does not necessarily cost accuracy in recommender settings, weakening a common argument against it.
  • Quantitative interpretability scores built from LIME and SHAP can be used to compare models on interpretability without relying on visual inspection of two-dimensional projections.
  • MacridVAE's claimed advantage in encoding user intents is supported by the interpretability measures even though its disentanglement scores are often lower than simpler models like PureSVD.
  • Reproducing disentanglement results requires access to the exact ground-truth factor construction; small choices in thresholds or bookshelf merging can change scores substantially.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the two interpretability scores and the two disentanglement scores are estimated from the same fitted factor classifiers, a natural robustness extension is to recompute the correlation with held-out factor classifiers or with different classifier families; if the [0.51, 0.95] range persists, the linkage is not an artifact of the estimator.
  • A testable next step is to replace the tag- and bookshelf-based factors with explicit user-side factors such as genre preferences or demographic segments; if the positive correlation survives, it would show the disentanglement-interpretability link is not an accident of item-content proxies.
  • The LIME-global and SHAP-global construction is model-agnostic and could be transplanted to latent spaces outside recommender systems, giving vision or language models a quantitative interpretability score tied to factor classifiers.
  • The finding that PureSVD scores high on disentanglement yet low on effectiveness suggests interpretability scores should be reported alongside effectiveness, not as a proxy for it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents a reproducibility study of five recommender models (Top-Popular, PureSVD, MultiDAE, MultiVAE, β-VAE, MacridVAE) on four datasets (Amazon-CD, ML1M, Yelp, GoodReads-Children). It measures recommendation effectiveness, disentanglement (DCI disentanglement and completeness), and proposes two new interpretability metrics, LIME-global and SHAP-global, computed from LIME/SHAP attributions of classifiers trained to predict hand-constructed ground-truth factors. The central claims are that (i) reported effectiveness of prior disentangled recommender models is only partially reproducible, with discrepancies up to 43% on GoodReads-Children, (ii) disentanglement is not consistently correlated with effectiveness, and (iii) disentanglement is strongly positively correlated with LIME-/SHAP-global interpretability (RMCORR in [0.51, 0.95]). The paper includes public code and is framed as the first quantitative study linking disentanglement to representation interpretability in recommender systems.

Significance. If the results hold, the paper would provide a useful quantitative check on a commonly asserted but rarely measured link between disentanglement and interpretability in recommender systems, and it would contribute a much-needed reproducibility dataset and codebase for disentangled recommendation models. The study is also valuable for its negative finding on the disentanglement-effectiveness trade-off, which challenges a widely repeated assumption. However, the central RQ3 claim is weakened by the fact that both disentanglement and the proposed interpretability measures are computed from the same classifiers trained on the same hand-constructed proxy factors, so the positive correlation may be partly structural rather than an independent empirical discovery. The paper's strengths include its open code, careful documentation of hyperparameter tuning, and honest reporting of failed reproduction attempts.

major comments (4)
  1. [§3.1, Eq. (1) and LIME-/SHAP-global definition] The disentanglement measure D and the proposed LIME-global/SHAP-global interpretability measures are not independent: both are computed from the same K gradient-boosting classifiers trained to predict the same K hand-constructed proxy factors. D is a function of the impurity-based importance matrix F of these classifiers, while LIME-/SHAP-global are functions of the LIME/SHAP importance matrices S of the same classifiers. A latent space in which each proxy factor is predicted by a distinct dimension will yield both high D (low-entropy rows of F) and high JS divergence (distinct, concentrated columns of S); a diffuse representation will yield low scores on both. The strong RMCORR reported in Section 4.3 may therefore be partly or wholly a mathematical consequence of this shared construction rather than evidence of an empirical link between disentanglement and interpretability. The authors should provide a control or sensitivity analysis, for example computing LIME/SHAP-global from classifiers trained on shuffled factor labels, using held-out factor labels, or regressing out completeness, to show that the correlation is not structural.
  2. [§3.2, Ground truth factors] The ground-truth factors of variation are defined by a chain of hand-set choices: the 100 most popular tags, k-means clustering into 20 clusters, the ML1M relevance threshold M=0.4, the 50% user-membership threshold, and manual merging/dropping of bookshelves. Every disentanglement and interpretability score in the paper uses these proxies, and both sides of the RQ3 correlation share the same proxies. If these proxies do not match the true factors driving user choices, the mismatch would affect D and LIME-/SHAP-global in the same direction, inflating the observed correlation. Section 4.4 acknowledges the proxy limitation but does not test its impact. Please add sensitivity analyses varying k, the number of tags, and the membership threshold, or a comparison with an alternative factor definition, to establish that the correlation is robust to the proxy construction.
  3. [§3.1, LIME-/SHAP-global definition] The definition of the interpretability measures as 'the mean of the Jensen-Shannon (JS) divergence computed between every pair of columns of S' is incomplete and potentially invalid. JS divergence is defined for probability distributions, but the text only says the columns of S are 'normalize[d] into [0,1]'. If this is min-max scaling, the columns are not probability distributions and JS divergence is not well-defined; if it is sum-to-one normalization, the text should say so explicitly. This is a load-bearing issue because LIME-global and SHAP-global are the paper's proposed measures and are used in the central RQ3 correlation. Please specify the exact normalization and confirm that the inputs to JS divergence are valid probability distributions.
  4. [Table 3] The paper reports that it cannot reproduce the only prior quantitative disentanglement results (Nema et al. [35]), with relative differences in D up to 78% and in C up to 56%, despite trying multiple classifier families. Because the same DCI pipeline is used to produce the disentanglement scores that enter the RQ3 correlation, the authors should provide additional evidence that their D and C measurements are reliable, for example by reporting classifier accuracy on held-out factors or by showing that the qualitative conclusions are stable under different classifier choices. Without such evidence, the external validity of the central correlation claim remains uncertain. This is not a fatal flaw, but it needs to be addressed in a revision.
minor comments (4)
  1. [§3.2 vs §4.1 and §3.3] The binarization threshold is inconsistent: Section 3.2 says 'set to 1 all ratings ≥ 1', while Section 4.1 and Section 3.3 say 'all ratings > 1'. This difference changes the data and should be reconciled in the final version.
  2. [Tables 4 and 5] Top-Popular is listed in Table 4 but has no reported disentanglement or interpretability values in Table 5; the reason (no user representation) should be stated explicitly in the text or table footnote.
  3. [Figures 1 and 2] The captions describe the two figures as 'intra-model RMCORR per dataset' and 'intra-dataset RMCORR per model', but both figures appear to display the same correlation matrices arranged differently. The terminology is confusing and should be clarified.
  4. [§4.3] The claim that the correlation 'holds across all the datasets and models' is stronger than what the reported RMCORR values show: several entries in Figure 2 are not statistically significant (e.g., some LIME-global versus disentanglement correlations in individual models). Please qualify the claim with the actual ranges and significance levels.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported disentanglement–interpretability correlation in RQ3 is largely built into the measurement: both variables are computed from the same K classifiers and the same hand-constructed factor proxies.

  1. self definitional [Section 3.1 (Interpretability measures), Eq. (1); Section 4.3]
    "To quantify the interpretability of the disentangled representations, we utilise the K binary classifiers used for disentanglement and completeness and adapt LIME and SHAP into global measures, called LIME-global and SHAP-global."

    Disentanglement D and completeness C in Eq. (1) are computed from the importance matrix F of K gradient-boosting classifiers trained to predict the hand-constructed ground-truth factors. LIME-global and SHAP-global are then computed from the mean absolute LIME/SHAP importance vectors of the same K classifiers, collected into matrix S and scored by pairwise Jensen-Shannon divergence. Both measures therefore quantify the concentration and distinctness of feature importance of the same fitted classifiers with respect to the same factor labels. A representation whose latent dimensions align with the proxy factors will tend to score high on both, while a misaligned representation will tend to score low on both.

full rationale

The paper contains a substantial independent component: RQ1 reproduces effectiveness and disentanglement scores against existing external results, and RQ2 studies the disentanglement-effectiveness correlation using standard metrics; those parts are not circular. However, the central RQ3 finding, that disentanglement is closely related to representation interpretability, is compromised by construction. The paper's own text states that LIME-global and SHAP-global use the same K binary classifiers used for disentanglement and completeness, and the same hand-constructed factors enter both sides of the correlation. D is a row-wise entropy concentration measure over the importance matrix F, while LIME/SHAP-global is a column-wise divergence measure over a closely related importance matrix S derived from the same classifiers. High values of both are promoted by the same property: each factor's classifier attends to a distinct, concentrated subset of latent dimensions. This makes the observed strong positive correlation partly an artifact of the shared measurement framework, not a fully independent validation. The paper acknowledges limitations of the proxy factors and of LIME/SHAP, but it does not acknowledge or test the non-independence between the two measures. The score of 6 reflects partial circularity: one key claimed result reduces substantially to the construction of the measures, while other results in the paper remain independently meaningful.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central analysis rests on proxy ground-truth factors built from item metadata, and on the assumption that the proposed LIME-global and SHAP-global scores capture interpretability. These are domain assumptions inherited from prior work, not standard math. The hand-chosen thresholds are free parameters that materially affect all reported metrics.

free parameters (5)
  • Number of ground-truth clusters k = 20
    Hand-chosen number of k-means clusters for grouping the 100 most popular item tags/categories into ground-truth factors (Section 3.2).
  • Top tags/categories kept = 100
    Hand-chosen cutoff retaining the most frequent tags/categories before clustering; influences which factors are available (Section 3.2).
  • ML1M relevance threshold M = 0.4
    Relevance score threshold for assigning an ML1M item to a tag cluster, copied from Nema et al. [35]; affects the derived factors (Section 3.2).
  • User-cluster membership threshold = 50%
    A user is assigned to a factor cluster only if at least half of their interacted items fall in that cluster; a hand-chosen binarization of the factor labels (Section 3.2).
  • Bookshelf merge/drop criteria = manual
    Bookshelves manually merged if names look similar and dropped if semantically uninformative, following [35]; this shapes the GoodReads ground-truth factors (Section 3.2).
assumptions (4)
  • domain assumption Item tags, categories, and bookshelves are valid proxies for the true factors of variation in user preferences.
    All disentanglement and interpretability scores depend on these factors (Section 3.2); the paper acknowledges this only as a limitation in Section 4.4.
  • standard math The DCI metrics (disentanglement and completeness) correctly quantify representation disentanglement in the recommender setting.
    Uses the Eastwood-Williams framework [12] with gradient-boosting classifiers on latent representations (Section 3.1).
  • standard math Repeated-measures correlation provides valid inference despite non-i.i.d. resamples.
    Used to account for intra-model and intra-dataset correlations (Section 4.2, [1]).
  • domain assumption LIME and SHAP feature importance on the factor classifiers measures human-perceptible interpretability.
    The new LIME-global and SHAP-global scores assume importance concentration and sparsity equal interpretability; limitations of LIME and SHAP are inherited (Section 4.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are Representation Disentanglement and Interpretability Linked in Recommendation Models? A Critical Review and Reproducibility Study." pith.science (2026). https://pith.science/paper/5PSC4K6N

@misc{pith2026250118805,
  author       = {Pith},
  title        = {Pith review of: Are Representation Disentanglement and Interpretability Linked in Recommendation Models? A Critical Review and Reproducibility Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PSC4K6N}},
  note         = {Machine review of arXiv:2501.18805}
}
read the original abstract

Unsupervised learning of disentangled representations has been closely tied to enhancing the representation intepretability of Recommender Systems (RSs). This has been achieved by making the representation of individual features more distinctly separated, so that it is easier to attribute the contribution of features to the model's predictions. However, such advantages in interpretability and feature attribution have mainly been explored qualitatively. Moreover, the effect of disentanglement on the model's recommendation performance has been largely overlooked. In this work, we reproduce the recommendation performance, representation disentanglement and representation interpretability of five well-known recommendation models on four RS datasets. We quantify disentanglement and investigate the link of disentanglement with recommendation effectiveness and representation interpretability. While several existing work in RSs have proposed disentangled representations as a gateway to improved effectiveness and interpretability, our findings show that disentanglement is not necessarily related to effectiveness but is closely related to representation interpretability. Our code and results are publicly available at https://github.com/edervishaj/disentanglement-interpretability-recsys.

Figures

Figures reproduced from arXiv: 2501.18805 by the authors.

Figure 1
Figure 1. Repeated measurements correlation of effectiveness [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Repeated measurements correlation of effectiveness [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 50 canonical work pages

  1. [35]

    In: Proceedings of the 30th ACM International Conference on Information & Knowledge Manag ement

    Nema, P., Karatzoglou, A., Radlinski, F.: Disentanglin g preference representations for recommendation critiquing with β -V AE. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Manag ement. pp. 1356– 1365 (2021)

    work page 2021

  2. [1]

    Frontiers in psy- chology 8, 456 (2017)

    Bakdash, J.Z., Marusich, L.R.: Repeated measures correl ation. Frontiers in psy- chology 8, 456 (2017)

    work page 2017

  3. [2]

    IEEE transactions on pattern analysis an d machine intelligence 35(8), 1798–1828 (2013)

    Bengio, Y., Courville, A., Vincent, P.: Representation l earning: A review and new perspectives. IEEE transactions on pattern analysis an d machine intelligence 35(8), 1798–1828 (2013)

    work page 2013

  4. [3]

    In: Interna- tional conference on machine learning

    Bergstra, J., Yamins, D., Cox, D.: Making a science of mode l search: Hyperparam- eter optimization in hundreds of dimensions for vision arch itectures. In: Interna- tional conference on machine learning. pp. 115–123. PMLR (2 013)

    work page

  5. [4]

    Machine Learning 15, 223–250 (1994)

    Bohanec, M., Bratko, I.: Trading accuracy for simplicity in decision trees. Machine Learning 15, 223–250 (1994)

    work page 1994

  6. [5]

    arXiv preprint arXiv:1804.03599 (2018)

    Burgess, C.P., Higgins, I., Pal, A., Matthey, L., Watters , N., Desjardins, G., Lerch- ner, A.: Understanding disentangling in β -V AE. arXiv preprint arXiv:1804.03599 (2018)

    arXiv 2018

  7. [6]

    Advances in neural in formation processing systems 31 (2018)

    Chen, R.T., Li, X., Grosse, R.B., Duvenaud, D.K.: Isolati ng sources of disentan- glement in variational autoencoders. Advances in neural in formation processing systems 31 (2018)

    work page 2018

  8. [7]

    Advances in neural information processi ng systems 29 (2016)

    Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskeve r, I., Abbeel, P.: Info- gan: Interpretable representation learning by informatio n maximizing generative adversarial nets. Advances in neural information processi ng systems 29 (2016)

    work page 2016

Show all 53 references
  1. [8]

    In: Proceedings of the fourth A CM conference on Recommender systems

    Cremonesi, P., Koren, Y., Turrin, R.: Performance of reco mmender algorithms on top-n recommendation tasks. In: Proceedings of the fourth A CM conference on Recommender systems. pp. 39–46 (2010)

    2010

  2. [9]

    : Disentangling features for fashion recommendation

    De Divitiis, L., Becattini, F., Baecchi, C., Del Bimbo, A. : Disentangling features for fashion recommendation. ACM Transactions on Multimedi a Computing, Com- munications and Applications 19(1s), 1–21 (2023)

    2023

  3. [10]

    arXiv preprint arXiv:1702.08608 (2017)

    Doshi-Velez, F., Kim, B.: Towards a rigorous science of i nterpretable machine learn- ing. arXiv preprint arXiv:1702.08608 (2017)

    2017 arXiv

  4. [11]

    Advances in neural information processing systems 31 (2018)

    Dupont, E.: Learning disentangled joint continuous and discrete representations. Advances in neural information processing systems 31 (2018)

    2018

  5. [12]

    In: International conference on learning representations (2018) 14 Dervishaj et al

    Eastwood, C., Williams, C.K.: A framework for the quanti tative evaluation of dis- entangled representations. In: International conference on learning representations (2018) 14 Dervishaj et al

    2018

  6. [13]

    In: Proceedings of the 13th ACM conference on recommender syste ms

    Ferrari Dacrema, M., Cremonesi, P., Jannach, D.: Are we r eally making much progress? a worrying analysis of recent neural recommendat ion approaches. In: Proceedings of the 13th ACM conference on recommender syste ms. pp. 101–109 (2019)

    2019

  7. [14]

    ACM Transactions on Information Systems 42(1), 1–27 (2023)

    Gao, C., Wang, S., Li, S., Chen, J., He, X., Lei, W., Li, B., Zhang, Y., Jiang, P.: Cirs: Bursting filter bubbles by counterfactual interactiv e recommender system. ACM Transactions on Information Systems 42(1), 1–27 (2023)

    2023

  8. [15]

    In: Proceeding s of the fourth ACM conference on Recommender systems

    Ge, M., Delgado-Battenfeld, C., Jannach, D.: Beyond acc uracy: evaluating recom- mender systems by coverage and serendipity. In: Proceeding s of the fourth ACM conference on Recommender systems. pp. 257–260 (2010)

    2010

  9. [16]

    Ad vances in neural infor- mation processing systems 27 (2014)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Wa rde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Ad vances in neural infor- mation processing systems 27 (2014)

    2014

  10. [17]

    ACM comp uting surveys (CSUR) 51(5), 1–42 (2018)

    Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Gi annotti, F., Pedreschi, D.: A survey of methods for explaining black box models. ACM comp uting surveys (CSUR) 51(5), 1–42 (2018)

    2018

  11. [18]

    In: Proce edings of the 30th ACM International Conference on Multimedia

    Guo, Z., Li, G., Li, J., Chen, H.: TopicV AE: Topic-aware d isentanglement repre- sentation learning for enhanced recommendation. In: Proce edings of the 30th ACM International Conference on Multimedia. pp. 511–520 (2022 )

    2022

  12. [19]

    ACM transactions on interactive intelligent systems (tiis) 5(4), 1–19 (2015)

    Harper, F.M., Konstan, J.A.: The movielens datasets: Hi story and context. ACM transactions on interactive intelligent systems (tiis) 5(4), 1–19 (2015)

    2015

  13. [20]

    In: Proceedings of the 26th international confer ence on world wide web

    He, X., Liao, L., Zhang, H., Nie, L., Hu, X., Chua, T.S.: Ne ural collaborative filtering. In: Proceedings of the 26th international confer ence on world wide web. pp. 173–182 (2017)

    2017

  14. [21]

    In: International conference on learnin g representations (2016)

    Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X ., Botvinick, M., Mohamed, S., Lerchner, A.: β -V AE: Learning basic visual concepts with a constrained var ia- tional framework. In: International conference on learnin g representations (2016)

    2016

  15. [22]

    ACM Transactions on Information Systems (TOIS) 20(4), 422–446 (2002)

    Järvelin, K., Kekäläinen, J.: Cumulated gain-based eva luation of ir techniques. ACM Transactions on Information Systems (TOIS) 20(4), 422–446 (2002)

    2002

  16. [23]

    , Wortman Vaughan, J.: Interpreting interpretability: understanding data scien tists’ use of interpretability tools for machine learning

    Kaur, H., Nori, H., Jenkins, S., Caruana, R., Wallach, H. , Wortman Vaughan, J.: Interpreting interpretability: understanding data scien tists’ use of interpretability tools for machine learning. In: Proceedings of the 2020 CHI c onference on human factors in computing systems...

    2020

  17. [24]

    In: Int ernational Conference on Machine Learning

    Kim, H., Mnih, A.: Disentangling by factorising. In: Int ernational Conference on Machine Learning. pp. 2649–2658. PMLR (2018)

    2018

  18. [25]

    In: International Con- ference on Learning Representations (ICLR) (2014)

    Kingma, D.P., Welling, M.: Auto-encoding variational b ayes. In: International Con- ference on Learning Representations (ICLR) (2014)

    2014

  19. [26]

    J.: Building machines that learn and think like people

    Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman, S. J.: Building machines that learn and think like people. Behavioral and brain scien ces 40, e253 (2017)

    2017

  20. [27]

    In: International Conference on Learnin g Representations (2019)

    Lezama, J.: Overcoming the disentanglement vs reconstr uction trade-off via jaco- bian supervision. In: International Conference on Learnin g Representations (2019)

    2019

  21. [28]

    In: Interna- tional Conference on Database Systems for Advanced Applica tions

    Li, Y., Zhao, P., Wang, D., Xian, X., Liu, Y., Sheng, V.S.: Learning disentangled user representation based on controllable V AE for recommen dation. In: Interna- tional Conference on Database Systems for Advanced Applica tions. pp. 179–194. Springer (2021)

    2021

  22. [29]

    In: Proceedings of the 2018 wor ld wide web conference

    Liang, D., Krishnan, R.G., Hoffman, M.D., Jebara, T.: Var iational autoencoders for collaborative filtering. In: Proceedings of the 2018 wor ld wide web conference. pp. 689–698 (2018)

    2018

  23. [30]

    In: international conference on machine learning

    Locatello, F., Bauer, S., Lucic, M., Raetsch, G., Gelly, S., Schölkopf, B., Bachem, O.: Challenging common assumptions in the unsupervised lea rning of disentangled Are Disentanglement and Interpretability Linked in RSs? 15 representations. In: international conference on mach...

    2019

  24. [31]

    Advances in neural information processing systems 30 (2017)

    Lundberg, S.M., Lee, S.I.: A unified approach to interpre ting model predictions. Advances in neural information processing systems 30 (2017)

    2017

  25. [32]

    Advances in neural information proces sing systems 32 (2019)

    Ma, J., Zhou, C., Cui, P., Yang, H., Zhu, W.: Learning dise ntangled representations for recommendation. Advances in neural information proces sing systems 32 (2019)

    2019

  26. [33]

    In: The Twelfth Inter national Conference on Learning Representations (2024)

    Meo, C., Mahon, L., Goyal, A., Dauwels, J.: α -TCV AE: On the relationship be- tween disentanglement and diversity. In: The Twelfth Inter national Conference on Learning Representations (2024)

    2024

  27. [34]

    In: Proceedings of the 16th ACM Conference on Recommender Systems

    Michiels, L., Verachtert, R., Goethals, B.: Recpack: An (other) experimentation toolkit for top-n recommendation using implicit feedback d ata. In: Proceedings of the 16th ACM Conference on Recommender Systems. pp. 648–651 (2022)

    2022

  28. [36]

    Ni, J., Li, J., McAuley, J.: Justifying recommendations using distantly-labeled re- views and fine-grained aspects. In: Proceedings of the 2019 c onference on empirical methods in natural language processing and the 9th internat ional joint conference on natural language proces...

    2019

  29. [37]

    In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Ren, X., Xia, L., Zhao, J., Yin, D., Huang, C.: Disentangl ed contrastive collabo- rative filtering. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. p. 11 37–1146. SIGIR ’23 (2023)

    2023

  30. [38]

    why should i tru st you?

    Ribeiro, M.T., Singh, S., Guestrin, C.: “why should i tru st you?” explaining the pre- dictions of any classifier. In: Proceedings of the 22nd ACM SI GKDD international conference on knowledge discovery and data mining. pp. 1135 –1144 (2016)

    2016

  31. [39]

    Statistic Sur- veys 16, 1–85 (2022)

    Rudin, C., Chen, C., Chen, Z., Huang, H., Semenova, L., Zh ong, C.: Interpretable machine learning: Fundamental principles and 10 grand chal lenges. Statistic Sur- veys 16, 1–85 (2022)

    2022

  32. [40]

    Neural computation 4(6), 863–879 (1992)

    Schmidhuber, J.: Learning factorial codes by predictab ility minimization. Neural computation 4(6), 863–879 (1992)

    1992

  33. [41]

    Advances in neural inf ormation processing systems 32 (2019)

    Shi, Y., Paige, B., Torr, P., et al.: Variational mixture -of-experts autoencoders for multi-modal deep generative models. Advances in neural inf ormation processing systems 32 (2019)

    2019

  34. [42]

    In: International conference on machine learning

    Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribut ion for deep networks. In: International conference on machine learning. pp. 3319–33 28. PMLR (2017)

    2017

  35. [43]

    TREC-8 Question Answering Track Report pp

    Voorhees, E.: Proceedings of the 8th text retrieval conf erence. TREC-8 Question Answering Track Report pp. 77–82 (1999)

    1999

  36. [44]

    In: Proceedings of the 12th ACM conference on recommender syste ms

    Wan, M., McAuley, J.: Item recommendation on monotonic b ehavior chains. In: Proceedings of the 12th ACM conference on recommender syste ms. pp. 86–94 (2018)

    2018

  37. [45]

    arXiv preprint arXiv:190 5.13416 (2019)

    Wan, M., Misra, R., Nakashole, N., McAuley, J.: Fine-gra ined spoiler detection from large-scale review corpora. arXiv preprint arXiv:190 5.13416 (2019)

    2019

  38. [46]

    In: Proceedings of the 43rd intern ational ACM SIGIR con- ference on research and development in information retriev al

    Wang, X., Jin, H., Zhang, A., He, X., Xu, T., Chua, T.S.: Di sentangled graph collaborative filtering. In: Proceedings of the 43rd intern ational ACM SIGIR con- ference on research and development in information retriev al. pp. 1001–1010 (2020)

    2020

  39. [47]

    IEEE T ransactions on Knowl- edge and Data Engineering (2022) 16 Dervishaj et al

    Wang, X., Li, Q., Yu, D., Cui, P., Wang, Z., Xu, G.: Causal d isentanglement for semantics-aware intent learning in recommendation. IEEE T ransactions on Knowl- edge and Data Engineering (2022) 16 Dervishaj et al

    2022

  40. [48]

    arXiv preprint arXiv:2211.11695 (2022)

    Wang, X., Chen, H., Tang, S., Wu, Z., Zhu, W.: Disentangle d representation learn- ing. arXiv preprint arXiv:2211.11695 (2022)

    2022 arXiv

  41. [49]

    In: Proceedings o f the ACM Web Confer- ence 2023

    Yang, Y., Huang, C., Xia, L., Huang, C., Luo, D., Lin, K.: D ebiased contrastive learning for sequential recommendation. In: Proceedings o f the ACM Web Confer- ence 2023. pp. 1063–1073 (2023)

    2023

  42. [50]

    IEEE Transactions on Knowledge and Data Engineering (2024)

    Zhang, L., Liu, G., Liu, X., Wu, J.: Denoising item graph w ith disentangled learn- ing for recommendation. IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  43. [51]

    IEEE Transactions on Knowledge and Data Engineering (2022)

    Zhao, Z., Chen, J., Zhou, S., He, X., Cao, X., Zhang, F., Wu , W.: Popularity bias is not always evil: Disentangling benign and harmful bias for r ecommendation. IEEE Transactions on Knowledge and Data Engineering (2022)

    2022

  44. [52]

    In: Proc eedings of the ACM Web Conference 2022

    Zheng, Y., Gao, C., Chang, J., Niu, Y., Song, Y., Jin, D., L i, Y.: Disentangling long and short-term interests for recommendation. In: Proc eedings of the ACM Web Conference 2022. pp. 2256–2267 (2022)

    2022

  45. [53]

    In: P roceedings of the Web Conference 2021

    Zheng, Y., Gao, C., Li, X., He, X., Li, Y., Jin, D.: Disenta ngling user interest and conformity for recommendation with causal embedding. In: P roceedings of the Web Conference 2021. pp. 2980–2991 (2021)

    2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.