Disentangling Interpretable Factors with Supervised Independent Subspace Principal Component Analysis

The success of machine learning models relies heavily on effectively representing high-dimensional data. However, ensuring data representations capture human-understandable concepts remains difficult, often requiring the incorporation of prior knowledge and decomposition of data into multiple subspaces. Traditional linear methods fall short in modeling more than one space, while more expressive deep learning approaches lack interpretability. Here, we introduce Supervised Independent Subspace Principal Component Analysis ($\texttt{sisPCA}$), a PCA extension designed for multi-subspace learning. Leveraging the Hilbert-Schmidt Independence Criterion (HSIC), $\texttt{sisPCA}$ incorporates supervision and simultaneously ensures subspace disentanglement. We demonstrate $\texttt{sisPCA}$'s connections with autoencoders and regularized linear regression and showcase its ability to identify and separate hidden data structures through extensive applications, including breast cancer diagnosis from image features, learning aging-associated DNA methylation changes, and single-cell analysis of malaria infection. Our results reveal distinct functional pathways associated with malaria colonization, underscoring the essentiality of explainable representation in high-dimensional data analysis.

Paper

References (30)

Scroll for more · 18 remaining

Similar papers

Peer review

Reviewer debZ7/10 · confidence 4/52024-07-03

Summary

Here the authors propose an extension of PCA that decomposes datasets into multiple independent subspaces that are encouraged to reflect provided covariates of interest. To enforce these dependencies, the authors leverage the Hilbert-Schmidt Independence Criterion (HSIC).

Strengths

In short, my views of the manuscript are generally positive. Some specific thoughts below: * **Clarity**: I found the manuscript well written and easy to follow. Well done! * **Motivation**: A recent line of work [1,2] has observed that disentangling separate sources of variation in case-control datasets (e.g. in scRNA-seq studies [3-4]) can assist with data exploration tasks. This work extends similar ideas to settings where no explicit case versus control datasets exist, but samples can still be associated with different covariates of interest. I can imagine this work will similarly prove useful in a number of data exploration tasks. * **Novelty**: To my knowledge, the authors proposed method is indeed novel, and seems like a sensible extension of the previous supervised PCA method and recent works on supervised disentanglement (e.g. [1-2]). [1] Abid et al., "Contrastive principal component analysis" [2] Abid et al., "Contrastive variational autoencoder enhances salient features" [3] Weinberger et al., "Isolating salient variations of interest in single-cell data with contrastiveVI" [4] Jones et al., "Contrastive latent variable modeling with application to case-control sequencing experiments"

Weaknesses

While my views of the manuscript are generally positive, I do have a couple of minor concerns. If the authors are able to address my concerns I would be happy to raise my score. * **Impact of $\lambda$**: The authors' method attempts to find a balance between encouraging dependence between subspaces and their corresponding covariates while also minimizing the dependence between individual subspaces, and this balance is tuned via $\lambda$ in Equation (1). I imagine this parameter could have a significant impact on the embeddings/loadings produced by sisPCA, though this isn't explored in the manuscript. It would be great if the authors could perform additional experiments exploring these changes. For example, do the GO results presented in Figure 5 vary for different values of $\lambda$? How different do the UMAPs look for the scRNA-seq data for different values of $\lambda$? * **Automatic selection of $\lambda**: Related to my previous point, the authors briefly mention that $\lambda$ could be selected using a procedure similar to [1]. It would be great to see an example of this in practice: forcing the user to select $\lambda$ manually seems like it could present a big obstacle in terms of usability, so I think it would be useful to see how the automated method performs. * **Re: connections with self-supervised learning**: In the abstract the authors note "We elucidate the mathematical connections of sisPCA with self-supervised learning and regularized linear regression". While I was able to locate the regularized linear regression results in Appendix B.1, I couldn't find any additional mathematical details on the self-supervision connection (the only other mention of self-supervision in the main text was in line 83 in the related works section). Could the authors expand upon this connection? (Please let me know if I missed anything here). Minor points (these did not affect my score): * It appears that the authors may have used \citet accidentally instead of \citep throughout the manuscript. * Figure quality was relatively low-resolution when I printed out the manuscript- perhaps the authors could try increasing the DPI? * In line 91 the authors mention "contrastive learning"- it may be worth putting explicit citations to e.g. Abid et al. [1] to clarify for the reader that "contrastive learning" here isn't referring to contrastive in the sense of e.g. SimCLR. * I couldn't find any details on how the real-world biological datasets were preprocessed (e.g. I assume library size normalization was applied to the scRNA-seq data?) [1] Abid et al., "Contrastive principal component analysis"

Questions

Did the authors try applying their method with different values of $\lambda$ to the real-world datasets?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The greatest limitations I see are (1) the method relying on linear transformations (as opposed to e.g. deep neural networks) and (2) potential difficulties in setting $\lambda$. For (1) the authors specifically note that they are intentionally trading off expressivity for increased interpretability

Reviewer QSK96/10 · confidence 4/52024-07-13

Summary

This paper proposed a linear dimensionality reduction and subspace extraction method based on Hilbert-Schmidt Independence Criterion (HSIC) and Supervised PCA. In specific, several interpretable subspaces, which are independent from each other, are disentangled from the data observations and the leaned subspaces are enforced to be maximally dependent on the supervision variables. The proposed approach is shown to be effective on both synthetic data and two real datasets.

Strengths

1. A new method for subspace learning is proposed using supervised PCA and dependence maximization and minimization with HSIC for extracting independent latent low-dimensional subspaces. The framework is corroborated effective in extracting linearly mixed simulation data, and the learned representations are more interpretable than that of supervised PCA. 2. The problem formulation and optimization are clearly derived and delivered. The overall manuscript is written in an organized way and the notations are well-defined and relatively easy to follow.

Weaknesses

1. It is not clear if the proposed method is guaranteed to extract the latent subspaces and/or under what conditions/assumptions the method would work or fail. For example, in the synthetic experiment, the mixing matrix is uniformly drawn from [0, 1]. Does it still work if the matrix is changed to Gaussian? Overall, the theoretical analysis is somewhat lacking, e.g., identifiability analysis for linearly mixed subspaces. 2. Baselines seem missing in the real data applications. Only PCA and supervised PCA are considered here. The authors are encouraged to include more state-of-the-art methods to showcase the superiority of the proposed method, e.g., how does PCA work under such a setting? Are other HSIC-based nonlinear methods, e.g., [1], capable of extracting the independent subspace with interpretability? [1] Lopez, Romain, Jeffrey Regier, Michael I. Jordan, and Nir Yosef. "Information constraints on auto-encoding variational bayes." Advances in neural information processing systems 31 (2018).

Questions

1. In Sec. 3.2, the linear kernel is used to target variables. Does it make a difference if a nonlinear kernel function is used? 2. What are A and B in Fig. 2? Are they the target variables? In the same figure, why does it maximize dependence (HSIC) with the components of a subspace? I thought the only two objectives are 1) maximizing dependence between the subspace and its corresponding target variable; 2) minimizing the dependence between the subspaces. How is this maximization reflected in Eq. (2)? 3. What does balanced/unbalanced supervision mean mathematically in Conjecture 3.1?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors are encouraged to discuss limitation of the work in the main paper.

Reviewer UeQu6/10 · confidence 4/52024-07-13

Summary

The work proposes a new method that determines the factors of variation bonded to different labels in a supervised way, obtaining a method akin to supervised PCA but in several independent subspaces thorough the Hilbert-Schmidt independent criterion (that it’s employed to maximize both the independence between different subspaces and the dependence between the data subspaces and the labels). The paper focus on a linear kernel application of HSIC, but other kernels are also considered. Then it’s applied to some simulated data and to DNA/RNA sequencing data.

Strengths

The theoretical connections are interesting. The paper is technically sound, with a well-derived mathematical background. The results of genetic sequence analysis are meaningful and interesting.

Weaknesses

The novelty is limited, it appears to be an extension of previously described methods. The comparison is provided only with PCA or sPCA, which are not by any means SOTA methods. The explanation of the simulated data is not clear, what should it be the expected result? Are you learning the dimensions of the subspaces? A better description of the experiments is needed. How do you choose the hyperparameters (for instance, the dimension of the subspaces is 10 in both sections 4.2 and 4.3, why?). In section 4.2, the use of HSIC as a quality metric is not fair, since it’s already included in your loss. The significance of the paper would be increased if the authors prove that the method is useful for the analysis of other types of real-world datasets. The discussion is too short and not too clear. I acknowledge the limitations of space, but the discussion is one of the most important sections and should be properly done (even by moving some mathematical details to SI).

Questions

Could you please provide a comparison of your method with other SOTA methods? Would you prove that the method is useful in other kinds of data types? Which are the hyperparameters of the method? How do you set them?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

I’m not satisfied with the way the limitations are addressed. One should explain them clearly and provide examples in which the method is not working.

Authorsrebuttal2024-08-12

New Application on Breast Cancer Diagnostic Data

We appreciate your feedback and apologize for our previous insufficient response. Due to space constraints, we couldn't include additional figures and results initially. To further illustrate sisPCA's versatility as a plug-in replacement of PCA for disentangling quantities, we provide below a new application, which will be included in the appendix of the final paper, and as a tutorial upon package release. ### **Problem and Dataset** We used the Kaggle Breast Cancer Wisconsin Data Set (569 samples, 30 summary features, uciml/breast-cancer-wisconsin-data, CC BY-NC-SA 4.0 license). The 30 real-valued features are computed from imaging data of breast mass, which include the mean, standard error and the extremum of quantities like cell nuclear radius, texture, perimeter etc. Here, our goals are to: 1. Learn a representation for predicting sample disease status (Malignant or Benign, not used during training). 2. Understand how original features contribute to data variability, the learned representation, and diagnosis potential. ### **Experiments and Results** We focus our comparison on three linear models: PCA, sPCA, and sisPCA, using zero-centered and variance-standardized data as inputs. The diagnosis label used to measure subspace quality remains invisible to all models. Below we summarize the quantitative results, and will include subspace visualization in the final paper. #### **PCA** We projected all features (dim = 30) onto one PCA subspace (dim = 6, determined by the elbow rule), explaining 61.6% of total variance. Malignant and benign samples appear well separated in PC1 and PC2. 'symmetry_mean' (loading = -0.223) and 'radius_mean' (loading = -0.219) are the top 2 features that negatively contribute to PC1. That is, the higher the two feature scores, the lower PC1 score and the greater the possibility of the sample being malignant. #### **sPCA** From the PC1 loadings, we sought to construct two subspaces to separately reflect nuclei size (‘radius_mean’ ) and shape ('symmetry_mean'). We set $Y_{radius}$ ('radius_mean', 'radius_sd') of $569 \times 2$ as the target variable for the radius subspace, and $Y_{symmetry}$ ('symmetry_mean', 'symmetry_sd') of $569 \times 2$ as the target for the symmetry subspace. The remaining 26 features were projected onto the two subspaces using sPCA (dim = 3, effective dim = 2). *Both subspaces better explained diagnosis status than PCA (**Table S3**) but remained highly entangled.* Specifically, the PC2 loadings of the two spaces have a Pearson correlation of 0.716, and 'perimeter_worst', which is highly correlated to 'radius_mean' (corr = 0.965), also strongly contributes to the PC2 of the symmetry subspace (loading = 0.238). #### **sisPCA** We next applied sisPCA ($\lambda$ = 10) to further disentangle the two subspaces, increasing their separation (Grassmann distance increased from 1.593 to 2.058, PC2 loadings correlation decreased to 0.257). Here ‘perimeter’ features no longer contribute to the symmetry subspace. *As a result, the radius subspace remained predictive of diagnosis, while the symmetry subspace became less relevant (**Table S3**).* We confirmed the finding by directly measuring the predictive potential of the target variables $Y_{radius}$ and $Y_{symmetry}$ (Silhouette score = 0.457 and 0.092, respectively). **Table S3: Predictability of diagnosis status, measured by Silhouette score** | | Radius subspace (dim = 3) | Symmetry subspace (dim = 3) | Overall subspace (dim = 6) | |:-|:-|:-|:-| | PCA (one subspace) | 0.294 (PC 1-3) | 0.013 (PC 4-6) | 0.160 (PC 1-6) | | sPCA | 0.534 | 0.374 | 0.464 | | sisPCA | 0.553 | 0.027 | 0.511 | ### **Interpretation** 1. Our results suggest that nuclear size (radius subspace) is more informative for breast cancer diagnosis than nuclear shape (symmetry subspace), aligning with clinical observations [1]. 2. Unsupervised PCA and sPCA captured both radius and symmetry aspects as they contribute most to data variability. *Without disentanglement, the two aspects remain intertwined, potentially leading to incorrect conclusions about symmetry's predictive power for diagnosis.* 3. *sisPCA successfully separated the two representations, revealing their distinct relationships to diagnosis.* The results are interpretable: the radius subspace is constructed using features like 'area' and 'perimeter', while the symmetry subspace uses features like 'compactness' and 'smoothness'. This example demonstrates sisPCA's ability to disentangle different aspects of data variation and uncover underlying relationships. **Importantly, sisPCA improves upon the diagnosis potential of PCA's representation, even through the diagnosis labels were never used during training.** [1] Kashyap, Anamika, et al. "Role of nuclear morphometry in breast cancer and its correlation with cytomorphological grading of breast cancer: A study of 64 cases." Journal of cytology 35.1 (2018): 41-45.

Reviewer UeQu2024-08-13

I thank the authors for their reply. This new application makes both the application and the method much clearer, and I would put it in the main text (just before 4.2). I will raise my score accordingly.

Authorsrebuttal2024-08-13

Thank you

Thank you for the feedback and suggestion. We will add this application as a new subsection and adjust the main text accordingly.

Reviewer PWKM6/10 · confidence 3/52024-07-21

Summary

This paper presents a method for distentangling multiple independent linear latent subspaces that align with a set of response variables. This uses the Hilbert-Schmidt Independence Criterion (HSIC) to measure the dependence each subspace and the targe variable providing supervision for that subspace, a concept which was previously applied to infer a single subspace that was connected to a set of response variables. HSIC is additionally used to enforce independence between the subspaces, to encourage the method to find independent subspaces that correlate with each of the provided dependent variables. This expands on Supervised PCA by enabling inference of an indendent subspace for each response variable, rather than a single subspace which aligns with the vector of response variables, reflecting the fact that different responses may be interacting with different subspaces in the data, e.g., methylation changes reflecting natural human aging and changes related to a disease like cancer. The method is evaluated on synthetic data and shown to do a good job at extracting the different subspaces. In particular, Supervised PCA is dominated by one of the two supervised subspaces. The method is then evaluated on real methylation data and single-cell gene expression data. In the cancer data, the new method is shown to provide subspaces with greater separatio

Strengths

The paper tackles an important problem in the analysis of high-dimensional omics data, where there is a need to identify interpretable subspaces that can be mapped onto higher-level biological processes but where different phenotypes (e.g., aging vs cancer) may manifest differently, requiring independent subspaces rather than supervision of a single subspace as per sPCA. The paper is well-written and combines simulated data experiments, which help understand the method is behaving as expected, with two applications on real data.

Weaknesses

One could argue that the work is somewhat incremental. The use of the HSIC is not new and nor is the idea of providing supervision to a latent variable model to ensure the latent subspaces reflected some additional variables. However, I feel the contribution is novel enough to warrant presentation at NeurIPS

Questions

No questions

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors describe the well-known limitations of unsupervised PCA, i.e., identifiability, and other limitations implied by their choice of kernel.

Reviewer UeQu2024-08-11

Follow up

I thank the authors for their reply and acknowledge their usefulness. I will, accordingly, raise my score. However, I have to say that I'm not satisfied with the reply to "other kinds of data sets" reply and it refrains me from further raising the score.

Reviewer QSK92024-08-11

Thank you for the clarification and for including the non-linear VAE-based baselines, which should make this work more solid. I agree that linear methods have more benefits in terms of interpretability and simplicity.

Authorsrebuttal2024-08-13

Thank you

Thank you for your feedback. We appreciate your recognition and the suggestion on adding the non-linear baselines which has helped improve the quality of our work. We will include relevant benchmark results in the final version.

Reviewer debZ2024-08-11

Dear authors, Thank you for your thoughtful responses! I've accordingly raised my score. Based on the rebuttal I have a few remaining minor comments: * I would encourage the authors to avoid the name "sVAE" when describing the supervised VAE baseline, as this name has previously been used for a different class of models [1,2] * Two recent works [3,4] have investigated using HSIC-based penalties with VAEs for similar supervised disentanglement tasks; for completeness it would be great if the authors could include citations to these works * Please do include your final results re: the automatic selection procedure for $\lambda$, as I believe this will make a big impact in terms of usability of the method. [1]: Lachapelle et al. "Disentanglement via mechanism sparsity regularization: A new principle for nonlinear ICA" [2]: Lopez et al. "Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling" [3]: Tu et al. "A Supervised Contrastive Framework for Learning Disentangled Representations of Cell Perturbation Data" [4]: Qiu et al. "Isolating structured salient variations in single-cell transcriptomic data with StrastiveVI"

Authorsrebuttal2024-08-13

Thank you

Thank you for your positive feedback. We very much appreciate your recognitions and all your helpful suggestions for improving our paper. We will include the automatic selection procedure as part of the package release. To further enhance usability, we will also release the re-implemented HCV (hsVAE-sc) models for single-cell data and provide corresponding tutorials on example datsets and general documentations.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC