acceptodds
Under review as a conference paper at ICLR 2027

LAVA: Explainability for Unsupervised Latent Embeddings

Abstract

Unsupervised black-box models are drivers of scientific discovery, but are difficult to interpret as their output is often a multidimensional embedding rather than a well-defined target. While explainability for supervised learning estimates how input determines model output, its unsupervised counterpart should uncover how input is captured by the learned latent structure. However, existing adaptations of supervised model explainability for unsupervised learning deliver explanations either at the single-sample or the dataset-summary level, too fine-grained or reductive to be meaningful, and cannot explain models without a mapping function. To address this, we propose LAVA, a model-agnostic method to explain embedding structure through feature covariation in the input data. Using prototype-like explanations, LAVA captures local patterns of input feature correlations that may reoccur globally across the embeddings. LAVA offers stable explanations at a chosen granularity, revealing domain-relevant patterns such as pixel motifs in images and disease signals in cellular processes otherwise missed by existing methods.

Primary area:
interpretability and explainable AI
Keywords:
explainability, interpretability, latent, embedding, unsupervised, representation, manifold, learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vGByOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method called LAVA to give model agnostic, prototype-based explanations for latent embeddings generated by unsupervised learning methods. The authors tested their method on MNIST, KPMP and BOW datasets. Their evaluations contain the stability of LAVA explanations, showing LAVA properties such as "revealing globally reoccurring local patterns", a qualitative comparison with LFXAI, and a case study relating a KPMP module to chronic kidney disease.

Critical strengths

This paper proposes a novel method for generating explanations for embeddings produced by unsupervised learning methods, which applies to non-parametric embeddings that existing methods cannot explain. The authors show a practical use case on scientific data, e.g., KPMP single-cell kidney gene expression, where a module is related to chronic kidney disease. The method design and the experimental designs are careful.

Critical weaknesses
  1. The current comparison lacks baselines. The authors compare to NMF in Appendix only, and to LFXAI in qualitative comparisons. No other explanation method is benchmarked with LAVA. Moreover, the quantitative evaluation is mainly about stability and internal consistency, which is not strong enough to show that LAVA gives a "good" explanation.

  2. The paper does not show that the explanations depend on the embedding being explained. The authors state that LAVA "detects feature covariation irrespective of feature usage by the model". But explanations should also give users the opportunity to recognize a "bad" learned embedding space, which LAVA cannot do if its output is embedding-independent. Or the authors should frame that LAVA is a "data-centric" explanation.

Questions for authors
  1. Is there more related work on similar methods beyond those mentioned in the Introduction? Benchmarking against these baselines is needed in the experimental results.

  2. Is it possible to evaluate the correctness of LAVA explanations with "ground-truth" explanations? The authors should consider datasets that contain human labels on feature attributions, or synthetic data with correlation patterns.

  3. Is it possible to conduct human evaluations and show practical uses, such as scientific discovery of "new concepts" or debugging a 'bad' embedding space with LAVA?

3 comments

We thank the reviewer and answer in turn.

Baseline, correctness, ground-truth, quantitative evaluation

We add new experiments on synthetic data (Section 3.1). These highlight properties and performance of the LAVA method, as well as loss behavior of the AMF method, based on 5 synthetic embedding datasets showing a simple starting example and 4 more complex cases of non-uniqueness and low-rank approximation. For detail on methodology and results, we kindly refer the reviewer to the added Section 3.1.

Comparison to related work, including feature attribution

We found three new references, which we added to the Introduction: ClusterShapley (Marcilio-Jr & Eler, 2021), a method adapting Shapley explanations for feature attribution to explain manually annotated clusters of manifold learning methods; FADEx (Meneses et al., 2026), a single-sample (local) feature attribution method for manifold learning; and a gradient-based feature attribution method for t-SNE (Corbugy et al., 2024). All three fall into previously defined categories of methods adapting feature importance to unsupervised embeddings, either single-sample or dataset-wide, or require manual annotation. We argue that LAVA presents a fundamentally different view of explainability, shifting the focus from “why” some structure is uncovered to “what” the structure uncovered, well suited for the unsupervised setting (Section 1). Specifically, LAVA explanations represent local patterns reoccurring globally across latent embeddings. This makes direct comparison difficult, as feature attribution answers a different question and it is not designed or easily extendable to show commonality or reoccurrence of patterns across the entire structure of a latent space like LAVA (Section 3.4). We further note that feature attribution ground truth cannot be used to assess LAVA explanations, as LAVA examines features that covary together in neighborhoods of embeddings, rather than why an embedding method groups those samples into neighborhoods. Generally, we think repurposing existing XAI methods towards a fundamentally different goal, to enable direct or quantitative comparison with LAVA, is non-trivial and beyond the scope (as it would require substantial methodology contributions). We also note that there is no widely accepted consensus on the definition of an explanation, where distinct types of XAI can offer reasonable and complementary explanations from different perspectives on the same model. We believe that the comparison of LAVA to the LFXAI method, as a representative of feature-attribution XAI adapted for unsupervised embedding methods, illustrates these challenges well and also makes a general case about the differences and complementarity between LAVA and existing methods (Section 3.4).

LAVA embedding-dependence and diagnosing embedding “quality”

Conceptually, LAVA defines locality neighborhoods based on the embeddings, which inform pattern discovery across localities. As a result, LAVA explanations are influenced by the structure of the embeddings and do not merely recapitulate original data. We now include an embedding-awareness experiment in Section 3.3, assessing neighborhood and LAVA explanation (dis)similarity based on the original data and each of 3 different embeddings of MNIST: 2 differently parameterized UMAPs and a convolutional autoencoder. Briefly, we see differences in neighborhoods and in LAVA explanations, showing LAVA is indeed embedding-dependent. We also confirm that correlation patterns of the original space are poorly preserved when using a badly parametrized UMAP. Practically, comparison of LAVA explanations can be used complementarily to neighborhood retention (Han et al., 2022), to assess preservation of covariance structure between embeddings and the original data, and to diagnose “bad” embeddings (Section 3.3).

LAVA “detects feature covariation irrespective of feature usage by the model”

Clarification: explanations generated by LAVA depend on the embedding, since they capture patterns based on localities defined as neighborhoods in the latent space. At the same time, localities are summarized based on correlations of the original features, enabling LAVA to recover correlations that support the embedding structure rather than only those that might have been used by the method. In that sense, LAVA explanations are both data-oriented and embedding dependent.

Scientific discovery

Our results show LAVA extracts reoccurring correlation structure across MNIST digits, topics in documents of the BOW dataset, and kidney disease signal across cell types (Section 3), which are not apparent in the clustering (in the case of MNIST, also not apparent using LFXAI feature attribution; Section 3.4). For KPMP, we show LAVA discovers signal related to chronic kidney disease (CKD) specific to epithelial cells without supervision, identifying new genes for further investigation as potential disease biomarkers or therapy targets.

Reviewer_vGByOpenReview ↗

Thank you for your reply.

I thank the authors for their reply and additional experiments. However, the current reply does not address all my concerns. Regarding the baselines, the authors argue that the current feature-attributions methods are not comparable, but they do not include other options. Moreover, the human evaluation would be an interesting experiment which is not discussed either. Moreover, it would be great if the revised text in the pdf are highlighted in another color for a better view.

AuthorsOpenReview ↗

Addressing remaining comments

We thank the reviewer for their reply and apologize for not fully answering their concerns related to baseline comparisons and human evaluation more directly. We do so below.

Comparisons

We argue that LAVA addresses a fundamentally different question than existing explainability methods (Sec. 1), from “why” an embedding is generated, to what patterns in the data are captured by the latent structure. To quote: "explaining black-box embeddings requires new strategies for revealing what patterns in the data are captured by a learned multivariate latent structure". Existing methods for unsupervised embeddings yield “single-sample or dataset-summary explanations that are generally too fine-grained or reductive to describe the latent structure”. This holds for XAI methods in general, including feature attribution and example-centric XAI.

The methods we added to the revised paper (FADEx, ClusterShapley, Corbugy et al.) produce single-sample, cluster-level, or dataset-level feature attributions. To our knowledge, no existing method extracts locally reoccurring (correlation) patterns across embeddings, making direct comparison difficult. Repurposing methods towards this goal would require new methodology. Given these constraints, we compare LAVA at different levels for which alternatives exist:

  • XAI: LAVA vs. LFXAI (Sec. 3.4): for the same MNIST embedding, LFXAI aggregates feature importances vs. LAVA reveals reoccurring patterns, offering complementary views.
  • XAI: LAVA vs. neighborhood retention XAI (Sec. 3.3): a new embedding-awareness experiment compares LAVA explanations between original data and 3 embeddings, showing that LAVA adds a new view of covariance structure preservation beyond existing neighborhood retention XAI (Han et al., 2022).
  • Module extraction: AMF vs. NMF (Sec 3.2, App. B.11): replacing AMF with NMF, a standard factorization, leads to 50% overestimation error ratio and denser components, representing average correlations vs. LAVA’s desired reoccurring patterns.
  • Ground truth (Sec. 3.1): a new experiment evaluates LAVA on 5 distinct synthetic embeddings with ground-truth, including cases of non-uniqueness.

Beyond feature attribution, other XAI also answers different questions than LAVA. For instance, example importance identifies individual examples that influence an embedding, whereas LAVA focuses on local neighborhoods (sets of examples) and describing patterns that reoccur across embedding neighborhoods. Similarly, prototype XAI selects representative samples (e.g., of a class or cluster) for parts of an embedding space, unlike LAVA modules which reoccur across the embeddings. Furthermore, concept-based methods such as TCAV (Kim et al., 2018) require predefined concepts, whereas LAVA discovers patterns without prior knowledge. Finally, counterfactuals and anchors are defined with respect to a predicted target, which unsupervised methods lack (Introduction, lines 46-49).

If the reviewer has a specific method in mind, we would be happy to incorporate further discussion on how it relates to LAVA.

Human evaluation

We agree with the reviewer that a human evaluation would be interesting. Unfortunately, a user study was not feasible within the revision period. Instead, we aimed to show the utility of LAVA through an extensive range of applications across domains and settings:

  • LAVA on MNIST digits (Sec. 3.2–3.4): LAVA reveals correlation patterns that reoccur across digits (e.g., shape similarities of digits 3, 5, and 8), which are not apparent from the clustering or from LFXAI explanations.
  • LAVA on Bag-of-words (Sec. 3.2): LAVA discovers topics shared across documents.
  • LAVA on KPMP kidney single-cell data (Sec. 3.5): without supervision, LAVA finds a module related to chronic kidney disease specific to epithelial cells, pointing to candidate gene markers for further investigation.
  • LAVA-enabled diagnosis of embeddings (Sec. 3.3): comparing LAVA explanations shows that a badly parameterized UMAP fails to preserve the correlation structure of the original data and quantifies differences of correlation structure between different embeddings more generally.
  • LAVA recovery of ground truth (Sec. 3.1): correctness against ground truth, and the effect of non-uniqueness, showing LAVA results on 5 different synthetic embedding datasets.

We see LAVA as a proof of concept for a different view of explainability for unsupervised ML, one that explains the structure of embeddings rather than individual examples or their summary (Sec. 4). We believe we demonstrate that LAVA brings added value to the field of XAI, and think that our contribution will be of interest to researchers and future work in XAI. We hope that this can be recognized.


In the new version of the PDF, we now also highlight all added text and figures in yellow, as requested. We thank the reviewer once more for their time and effort.

Reviewer_3VjDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for explaining the latent embedding produced by a non-supervised learning model. The goal of the method is to capture local patterns of input feature correlations that recur globally.

The method consists of a weighted k-means clustering to define locality, representing locality as vector of pairwise feature correlations, and then extracting modules (proto-types) of feature correlations through a proposed AMF factorization method similar to NMF but using maximums rather than sums.

The proposed method is applied to three domains - MNIST with UMAP embeddings, KPMP kidney gene expression with UMAP embeddings, and Bag-of-Word text documents with t-SNE. Some qualitative analysis is given of generated explanations, and a comparison against an existing label-free method is also presented.

Critical strengths

The problem formulation - explaining latent embeddings produced by unsupervised/self-supervised methods - as a what question of understanding the learned latent structure, is interesting and important. The general approach of focusing on capturing local neighborhood patterns of feature correlation also seems sensible.

Critical weaknesses

I would have liked to see more quantitative evaluations, beyond measuring just stability, but measuring correctness in the sense that the proposed method is able to uncover true structure, such as by designing a synthetic dataset with underlying correlation modules.

A bigger concern is conceptually, is the proposed method uncovering covariance structure in the embedding space, or just in the original feature space itself? And closely related to this, does it make sense to be applying the method to UMAP and t-SNE embeddings, which are designed to maintain the same original local neighborhood structure in the original high dimensional space in the projected low dimensional space? How different would the results be if the method was simply run in the original feature space? I would have liked to see the method applied to a learned latent space designed to capture structure/manifold differing from the original feature space, such as from auto-encoder/self-supervised learning/contrastive learning, where there should be a clear difference between the results of running the method on the latent vs original feature space.

Questions for authors

Are there any computational or optimization hurdles to running LAVA on something like the latent space from an auto-encoder/self-supervised method, ie extending to latent representations with higher dimensionality?

1 comment

We thank the reviewer for their comments and questions. We answer all of them in turn.

Quantitative evaluation of correctness

We followed the reviewer’s suggestion to show the ability of LAVA to uncover true structure in synthetic datasets, and have now added a new section with experiments on synthetic data (Section 3.1). We generate 4 ground-truth correlation modules and 5 synthetic embedding datasets containing different combinations of clusters of samples based on those modules. Each case is designed to show specific properties and assess performance of LAVA explanations in the corresponding scenario, including non-uniqueness and low-rank approximation. For further details on methodology and results, we kindly refer the reviewer to Section 3.1.

LAVA embedding-dependence

Conceptually, LAVA defines locality neighborhoods based on the embeddings, which it then uses to identify modules of correlation across localities. We therefore expect that the discovered modules are influenced by and reflect the structure of the embeddings rather than merely recapitulate the structure of the original data. Following the reviewer’s question, we now include an embedding-awareness experiment in Section 3.3. In this experiment, we look both at neighborhood and LAVA explanation (dis)similarity when applying LAVA to the original feature space, and to each of 3 different latent embeddings of the MNIST dataset: 2 differently parameterized UMAPs and a convolutional autoencoder. In short, we see differences in neighborhoods (measured by Jaccard similarity) and in LAVA explanations (using presence-weighted module averages as described in Section 3.2), showing that LAVA is indeed embedding-dependent. We also highlight how comparison of LAVA explanations can be used complementarily to neighborhood retention analyses (Han et al., 2022), to assess to what extent covariance structure is preserved between latent neighborhoods and original neighborhoods, or more generally between latent neighborhoods of different embedding methods.

Computational scalability: running LAVA with higher dimensionality representations

We kindly refer the reviewer to our complexity and runtime experiments in Section 2.4 and Appendix B.12. In terms of latent dimensionality specifically, locality placement complexity scales linearly with latent dimensionality, and is therefore reasonably scalable to large latent spaces (e.g., the raw 784-dimensional feature space of MNIST in the newly added experiment above). Subsequent steps (calculation of correlations and AMF module extraction) are not affected by the dimensionality of the embedding space, as they utilize only localities and their representation (correlation patterns). In general, we think the largest computational hurdle might be the number of input features D, which influences the number of pairwise correlations and thereby the complexity of module extraction (AMF) quadratically.

Reviewer_x1tnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes LAVA, a model-agnostic approach to explain unsupervised black-box embedding models. LAVA produces prototype explanations for a local neighborhood in the embedding space, capturing common patterns that occur in the data as modules. The main innovation is the proposed Association Matrix Factorization (AMF), which separates the correlation/locality matrix into modules and their occurrences. Unlike NMP, AMP allows modules to be present in the same neighborhood and organizes patterns by their presence rather than averaging.

Critical strengths

AMF is an interesting formulation, and it seems well-motivated. By capturing common patterns, the prototype explanations found by LAVA effectively capture common patterns, and the experiments show convincingly that factoring into modules (which serve as the prototype explanations) and their occurrences can generate insight from an embedding model.

Critical weaknesses

While AMF makes intuitive sense, and the loss is well-motivated in context, I would like to see a more complete picture of AMF's theoretical properties. This would help confirm if e.g. Adam produces optimal solutions, or if the AMF factorizations are unique.

Questions for authors
  • As mentioned, I would like to see some theoretical analysis of AMF's objective. In particular, are the module and presence matrices unique?

  • Appendix B.7.4 mentions that one possible source of uncertainty for LFXAI is the shift-equivariance of a convolutional embedding model. Does this suggest that LAVA is more robust to these sorts of transformations? Or could AMF be modified to produce equivariant modules?

3 comments

We thank the reviewer for their comments and questions. We answer all of them in turn.

Theoretical analysis of AMF

We looked further into literature of related matrix factorization methods such as NMF. Most relevantly, Karaev and Miettinen (2019) conclude that, like NMF, subtropical matrix factorization is NP-hard: this result pertains to both determining the subtropical rank of the original matrix, as well as approximating a minimum-error solution (“within any polynomially computable multiplicative factor”). In the revised paper, we relate AMF to subtropical matrix factorization via their shared max formulation, and then analyze AMF’s custom loss to extend the result, finally concluding that AMF is also NP-hard. This is briefly mentioned in Section 2.3, and further expanded in App. A.4.2 as follows:

  • “The max formulation of AMF corresponds to subtropical matrix factorization, which uses a “max-times” product to decompose a non-negative matrix. Karaev & Miettinen (2019) show that, under conditions in which the loss function is 0 when a subtropical factorization is exact, and positive otherwise, both the computation of the subtropical rank of a matrix and approximation of a minimal-error solution (“within any polynomially computable multiplicative factor”) are NP-hard. Trivially, AMF loss is 0 when C^=C\hat{C} = C. While the main loss term of AMF is scale invariant, the added scale regularization term also makes the loss > 0 in all cases where C^≠C\hat{C} \neq C. We can therefore conclude that AMF is also NP-hard.”,

Furthermore, ADAM, as an adaptive stochastic gradient descent method, is not guaranteed to find optimal solutions of non-convex loss functions. To mitigate issues of poor local minima, we run module extraction several times with random initialization, analyzing loss and stability, and choosing the lowest loss solutions for a given number of modules (Sections 2.3 and 3.2).

Non-uniqueness of module and presence matrices

We found theoretical results for NMF (Rickard and Cichocki, 2008), which in our opinion are not trivially extendable to AMF and are therefore beyond the scope of this paper. Nevertheless, we note that their key observation is that “if each component has at least one point for which it is the only active component, the solution is unique”. Similarly, in Section 3.1 of the revised paper, we now explore several synthetic cases, where we also highlight a non-uniqueness scenario and its impact on LAVA explanations. Specifically, we show a non-uniqueness case where the underlying modules always co-occur in combinations (‘chain’ example), and how the resulting inseparability is resolved when a single module additionally appears in isolation (‘anchored chain’ example). We think these combined results, in addition to previous empirical results on real-world data (Sections 3.2-3.3), both characterize the non-uniqueness of AMF solutions and support quality and stability of solutions in real-world cases.

Robustness of LAVA to shift-equivariance of convolutional embeddings

On the question about Appendix B.7.4: we did not mean to suggest that LAVA is more robust to affine or other transformations, but rather to suggest a possible reason why we could expect SHAP-based feature importances of a convolutional model to have high variance across samples in the same latent neighborhood. We removed this comment from the revised paper to avoid confusion, as it is not relevant for our claims about LAVA or its comparison to LFXAI. In general, if a shift transformation is applied to all images of a dataset, and the embeddings are unaltered, LAVA explanations should be approximately equivariant in the sense that pixels and their correlations in localities will be transformed but preserved at different locations. On the other hand, the related case of producing invariant modules using AMF is also an interesting but challenging proposition, as it would require developing a notion of invariant correlations. Perhaps this could be done by using a kernel pooled across the image, like in convolutional networks. However, the correlation values (and even what are considered to be features) would then be strongly dependent on the design of the kernel, making results difficult to reason about. In the last paragraph of Appendix A.3, we comment on the possibility of altering LAVA’s locality representation to be that of concept activations rather than correlations. If an activation of a concept in an image or a part of an image covered by a kernel is considered to be ground-truth, these activations (max-pooled across the image if defined by a kernel) would ultimately make AMF explanations as invariant as the activation of these concepts. While outside of the scope of our work, we can include this brief discussion in Appendix A.3 if the reviewer thinks it is a relevant point of interest for future research.

Thank you for the response. My questions have been adequately answered and I maintain a positive opinion of this work. Reading the responses to other reviewers, I appreciate the additional results as well. I personally would like to see the shift-equivariance discussion in Appendix A.3; although I recognize it is not entirely in the scope of this work, it could be an interesting direction for future work to build on AMP, and it deserves at least some discussion, in my view.

We have now added the discussion about equivariance from our previous reply to Appendix A.3 as requested, and uploaded the revised manuscript.

We thank the reviewer once more for their time, effort, and input!

Comments on the paper

1 comment

We thank all the reviewers for their time and valuable input. In addition to individual responses, we provide an overview of the 3 reviews and how we address them in the revised manuscript. Update (November 19): we also add an overview of all of our experimental results at the bottom of the comment.

Critical strengths

  • Problem setting

    • The formulation of explainability for unsupervised embeddings as a goal of explaining latent structure is interesting, and important (reviewers 3VjD, vGBy)
  • Solution

    • The method (LAVA) is novel and applies to non-parametric embeddings that existing methods cannot explain (reviewer vGBy)
    • The method design is careful (reviewer vGBy)
    • The focus of LAVA on neighborhoods and correlations is sensible (reviewer 3VjD)
    • The proposed AMF method is interesting and well-motivated to capture reoccurring patterns (reviewer x1tn)
  • Results

    • Experiments are well thought-out/convincing (reviewers x1tn, vGBy)
    • Results show that insights can be generated from embeddings (reviewer x1tn)
    • Experiments show a practical, biological use-case (reviewer vGBy)

Critical weaknesses and how they are addressed

  • Request for theoretical analysis of LAVA’s AMF method (reviewer x1tn)
    • we comment on theoretical results of NP-hardness from related literature and show how they extend to AMF (Section 2.3; Appendix A.4.2)
    • we add synthetic dataset experiments (see below)
  • Request for evaluation of correctness in cases with ground truth (reviewers 3VjD, vGBy)
    • we add synthetic embedding experiments to examine LAVA explanations with known ground-truth modules, showing behavior across several study cases, including non-uniqueness and low-rank approximation (Section 3.1; Figure 3)
  • Request to show dependence of LAVA explanations on embeddings (reviewers 3VjD, vGBy)
    • we add an experiment quantifying (dis)similarity of LAVA explanations obtained for 4 different representations of the MNIST dataset (including the original feature space and 3 other embeddings), and we indicate that comparing LAVA explanations can be used as another view of embedding similarity, complementary to neighborhood preservation analyses (Section 3.3; Figure 6)
  • Request for comparisons to related work (reviewer vGBy)
    • we cite three additional XAI methods in the Introduction, discuss that LAVA differs fundamentally from existing methods, and argue that comparison with LFXAI (Section 3.4) exemplifies both challenges of comparison and complementarity of their explanations
  • Request for examples of scientific discovery and or practical use cases (reviewer vGBy)
    • scientific discovery: we discuss how existing experiments show the potential of LAVA across three different domains, including a biological example of finding new leads for disease biomarkers
    • practical use case: we comment on leveraging comparison of LAVA explanations to identify “bad” embeddings, based on an example included in Section 3.3

⁠Summary of results, analyses, and experiments after revision (update November 19)

  • Correctness: LAVA explanations on 5 synthetic embeddings with ground-truth modules (Sec. 3.1)
  • Theory and method analysis: NP-hardness (Sec. 2.3), runtime (Sec. 2.4), non-uniqueness (Sec. 3.1)
  • Real-world data: results on 3 datasets: MNIST digits, KPMP gene expression, and BOW text (Secs. 3.2–3.5)
  • Robustness: stability evaluation, sensitivity to hyperparameters, and ablations (Sec. 3.2)
  • Pattern properties: reoccurrence, presence, and co-presence of LAVA modules (Sec. 3.3)
  • Embedding-awareness: LAVA explanations on the original data and 3 embeddings (Sec. 3.3)
  • Comparisons: AMF vs. NMF for module extraction (Sec. 3.3); LAVA explanations vs. neighborhood retention (Sec. 3.3); LAVA vs. LFXAI explanations (Sec. 3.4)
  • Scientific discovery: LAVA reveals a kidney disease signal and candidate genes in KPMP data (Sec. 3.5)
  • Practical use: evaluating embeddings and diagnosing "bad" ones (Sec. 3.3)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.