Propensity Score Alignment of Unpaired Multimodal Data

Multimodal representation learning techniques typically rely on paired samples to learn common representations, but paired samples are challenging to collect in fields such as biology where measurement devices often destroy the samples. This paper presents an approach to address the challenge of aligning unpaired samples across disparate modalities in multimodal representation learning. We draw an analogy between potential outcomes in causal inference and potential views in multimodal observations, which allows us to use Rubin's framework to estimate a common space in which to match samples. Our approach assumes we collect samples that are experimentally perturbed by treatments, and uses this to estimate a propensity score from each modality, which encapsulates all shared information between a latent state and treatment and can be used to define a distance between samples. We experiment with two alignment techniques that leverage this distance -- shared nearest neighbours (SNN) and optimal transport (OT) matching -- and find that OT matching results in significant improvements over state-of-the-art alignment approaches in both a synthetic multi-modal setting and in real-world data from NeurIPS Multimodal Single-Cell Integration Challenge.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Peer review

Reviewer qEyT6/10 · confidence 4/52024-07-01

Summary

This work addresses the problem of aligning unpaired samples from multimodal data, whereby the problem is to find sample $x_i$ from modality $i$ that is best “related” to sample $x_j$ from modality $j$, when those two samples come from the observation of the same phenomenon according to two different modalities (e.g. sensors). This is an important problem that arise in many endeavors including biology, and that has been somehow overlooked in the literature of multimodal representation learning, whereby large quantities of paired data (images and their corresponding captions) are available. In a nutshell, the idea proposed in this work is as follows. Provided the following assumptions are valid, that is there exists a common latent space $Z$, which is entangled with a “perturbation signal” or label $t$, but that it is independent of modality specific noise $U_{i,j}$, then it is possible to estimate a common space for matching samples through the “bridge” across modalities offered by the perturbation signal $t$. In practice, this amounts to training a classifier to predict $t$ given one of the available modality, which allows the creation of a transformed dataset endowed with a Euclidean distance. Such distance can be used to compute a matching matrix $M$, whereby entry $(i,j)$ represents the likelihood of $x_i$ being matched to $x_j$. Matching matrix $M$ can be obtained efficiently through the lenses of optimal transport matching with entropic regularization. It should be noted that $M$ can then be used to construct pseudo-samples for any learning task that requires paired samples $(x_i, x_j)$ by replacing them with $(x_i, \hat{x}_j)$, where the second term is obtained through the application of the matching matrix to a structural equation $f_\theta$, which is a parametric function with learnable parameters. Several experiments on synthetic data, and a dataset whereby ground truth pairing is available, as well as experiments where no ground truth is available, indicate that the proposed method, in combination with optimal transport matching, provides an improvement over state of the art matching methods.

Strengths

* This work presents a simple method to produce a matching matrix that can be used to pair multimodal data. The algorithm only requires training of two classifiers and the execution of the Sinkhorn’s algorithm to compute the optimal transport matching. * Experiments show that the proposed method using propensity scores in combination with optimal transport matching outperform the literature in terms of MSE and other specific metrics. * Experimental result indicate, surprisingly, that the proposed method implicitly enforces invariant to modality specific information, which endows propensity score with an improved generalization, although this specific phenomenon is only hinted at and requires additional studies

Weaknesses

* The proposed method relies on strong assumptions, which are only superficially discussed. Relaxation of the conditional independence assumption A1 is discussed in appendix A, but only for the case of a small perturbation set $t \in \{ 0,1 \}$, which allows the authors conclude that under exact optimal transport, this is equivalent to operate on order preserving effects of $t$. The validity of the method also relies on A2, which allows computing propensity score from data alone, and not the shared latent space $Z$. * The overall perturbation method, as well as the existence of a shared latent representation $Z$ is only vaguely described. I have looked up the key reference [Yang et al, 2021], but the lack of detail is uncomfortable. It is possible to get an intuition of how the method works, but I would prefer such important technical details to be spelled out clearly. * There is no mention of the scalability of the proposed approach. Building a matching matrix can be a daunting task for some domains whereby a large amount of unpaired training data is available. In this case, the matching matrix could be either a big square matrix, or a very skinny and tall matrix that could pose some computational challenges. It would also be important to discuss about what would be the best operating scenario for the proposed method in terms of the number of samples per modality. In very unbalanced scenarios, would the classifiers using one or the other modality really behave similarly as per Equation (5)?

Questions

* Would it be possible to explain (and eventually add to the appendix) in more detail the setup from [Yang et al, 2021] which is used in this work? More precisely, it would be great to have the definition of the multimodal autoencoder, the existence of a shared latent space, and the details of the perturbations $t$ and their effects on $Z$. This would really help formalizing better the use of $t$ as a bridge between modalities. * Can you provide a thorough discussion about the scalability of the proposed approach, the impact of imbalance in the size of the multimodal data (e.g. $n_1 >> n_2$), and the possible application to other domains with more abundant data? * Can you elaborate more on the baselines you compare your method to? Lines 235-239 are not detailed enough to fully appreciate the differences w.r.t. the proposed method. ======= POST REBUTTAL UPDATE ======= Dear authors, thank you for the rebuttal. I have raised my score.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

Yes, the authors acknowledge that their work relies on strong assumptions, but suggest that empirical results indicate their method is robust to failures. Nevertheless, I could not find the discussion on the experimental results that clearly indicates such robustness.

Reviewer XqRN5/10 · confidence 3/52024-07-03

Summary

The paper proposes a novel matching method for unpaired multimodal (bimodal) data. It aligns unpaired samples across modalities using a propensity score. Based on additional treatment information, the propensity score is learned and used to align multimodal samples. The method is evaluated on three datasets: one imaging dataset and one single-cell dataset (both with available matching ground truth); and one additional single-cell microscopy dataset without available ground truth.

Strengths

- Interesting idea: using a propensity score that leverages (often) available labels/treatments to learn a classifier is interesting. - Important Problem: A long line of multimodal methods describes how to leverage best-paired samples, but little work has been done on aligning unpaired samples. Besides biology, there is (probably) a lot of unpaired multimodal data. - well written and easy to read

Weaknesses

- Final purpose of the method: To me, it was not fully clear what the final goal/purpose of the method is. Is it (1) the aligning of unpaired multimodal samples? Or (2), learning from unpaired, multimodal samples while aligning the samples? If (1) is the goal, what can we conclude from the improved results? If (2) is the goal, it would be interesting to see if the improved alignment/matching results in better downstream task/learning from multimodal data compared to other works (e.g., Yang et al., 2021) or just unimodal methods without any pairing of data. - Evaluation: It is not straightforward to understand the used metrics (maybe also related to the point above) and relate the performance metrics to the quality of the method (besides a relative performance between different alignment methods). Maybe reporting some classification-based performances for alignment (is the assigned sample from the other modality the correct one) or some downstream task performance (what task can we solve using the trained method?) would help. - Although the method is advertised as multimodal, no experiments or proofs use more than two modalities. Maybe the term bimodal would be more accurate? Otherwise, an experiment or derivation for more than two modalities would be appreciated.

Questions

- what are the labels/perturbations in the image-based dataset? I did not see the information in section 6 (but maybe I missed it. Apologies in that case) - section 6.2: what is the reason for using the first 200 principal components for the gene expression modality (and not for the protein data as well)? - does assumption A1 hold in the datasets used? Is there a way to check whether the assumption holds? - there are some typos: - line 90: the our instead of the or our - line 168: form instead of from - line 227: against on instead of against or on

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors address some of the limitations.

Reviewer iHqN7/10 · confidence 3/52024-07-15

Summary

This paper presents a new method for pairing unpaired examples across different modalities using the labels of the examples. The method essentially trains a classifier for predicting labels for examples in each modality, then uses the classifier's logits across two modalities to calculate a similarity matrix. Empirical results show that the method is promising in aligning data from different modalities.

Strengths

- The paper is written clearly. - The problem of aligning unpaird data in multimodal learning is quite important since in many cases there's limited or no access to paired data for representation learning. This paper proposes a new method for it, paving the way to future work in this area. - The method is relatively straightforward and easy to follow. - Empirical results show that the method perform well compared to the baseline.

Weaknesses

The method uses the information in the labels in order to match labels via classifiers. There are two drawbacks with this approach: (1) the quality of matching depends directly on the choice of the classifier and its capacity/complexity. (2) more importantly, in many multimodal representation learning settings the label signal is unavailable.

Questions

Is it possible that the use of labels as the sole signal to match examples from two modalities results in example clustering based on labels? i.e. examples from modality 1 that belong to class t are matched strongly against those from modality 2 that also belong to class t? In that case, is the method still beneficial compared to not doing any matching (i.e. just using data from individual modalities)? Having a discussion on this in the paper will be useful.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

There's a short section on limitations of the method, explaining (1) the reliance of the method on labels, and (2) the strong assumptions made in the method.

Authorsrebuttal2024-08-07

Likewise, thank you so much for the time you put into your review! Please let us know if there's anything else you need to know, and if you have any other concerns that are preventing you from raising your score.

Reviewer qEyT2024-08-11

Thank you for your rebuttal

Dear authors, thank you for the thorough rebuttal to my observations and questions. I appreciate the explanation on the assumptions made in this work, and how do they compare to the ones done in the state of the art. I also appreciate the discussion on the obtained results, which surpass currently available methods. I will raise my score to reflect these observations.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC