Semi-Supervised Sparse Gaussian Classification: Provable Benefits of Unlabeled Data

The premise of semi-supervised learning (SSL) is that combining labeled and unlabeled data yields significantly more accurate models. Despite empirical successes, the theoretical understanding of SSL is still far from complete. In this work, we study SSL for high dimensional sparse Gaussian classification. To construct an accurate classifier a key task is feature selection, detecting the few variables that separate the two classes. % For this SSL setting, we analyze information theoretic lower bounds for accurate feature selection as well as computational lower bounds, assuming the low-degree likelihood hardness conjecture. % Our key contribution is the identification of a regime in the problem parameters (dimension, sparsity, number of labeled and unlabeled samples) where SSL is guaranteed to be advantageous for classification. Specifically, there is a regime where it is possible to construct in polynomial time an accurate SSL classifier. However, % any computationally efficient supervised or unsupervised learning schemes, that separately use only the labeled or unlabeled data would fail. Our work highlights the provable benefits of combining labeled and unlabeled data for {classification and} feature selection in high dimensions. We present simulations that complement our theoretical analysis.

Paper

Similar papers

Peer review

Reviewer 6RTo7/10 · confidence 4/52024-06-30

Summary

In this paper, the authors studied the problem of semi-supervised learning in a 2-class classification problem with a special distribution setting. They assumed that the samples in each class came from an isotropic Gaussian distribution with unknown mean vectors $\mu_1$ and $\mu_{-1}$. They also assumed that the vector $\Delta \mu = \mu_1 - \mu_{-1}$ is sparse, with $k << p$ non-zero elements, where $p$ is the dimension of the feature space, and $\lambda = \|\Delta\mu\|_2^2/4 = \mathcal{O}(1)$. The claims of the paper are as follows: * Derive information-theoretic lower bounds for exact support recovery in the semi-supervised learning (SSL) setting. * Establish computational lower bounds for classification and support recovery in SSL as $(k, L, n, p) \to \infty$, where $L$ and $n$ are the number of labeled and unlabeled samples, respectively. * Identify a region where SSL is computationally advantageous for classification and feature selection. The authors derive lower bounds on the number of labeled and unlabeled samples needed to efficiently recover the support set of $\Delta\mu$ and classify data. They also propose an SSL algorithm that can efficiently recover the support set of $\Delta\mu$ and classify data, outperforming any supervised or unsupervised learning algorithm that relies solely on labeled or unlabeled samples.

Strengths

* The theoretical results are interesting, and the proofs are rigorous. * The writing is very good and easy to follow. * The messages of the paper are very clear.

Weaknesses

* It seems that the algorithm named LSPCA cannot be used as a semi-supervised algorithm on real-world data and only has theoretical value in the special setting of the problem. It would be more interesting if the authors proposed an algorithm that could be applied to real-world data, yielding good results in practice while also having strong theoretical guarantees.

Questions

*1*- In line 133, it is stated that the most difficult $k$-sparse vector with a lower bound on the absolute value of the non-zero terms is a $k$-sparse vector where its non-zero elements belong to $\{\pm\sqrt{\frac{\lambda}{k}}\}$. Is there any proof for this? Is it obvious? *2*- I think in line 823, $\Theta^S(j)$ should be $\sqrt{\frac{\lambda}{k}}1(j\in S)$. Is that correct? If it is, I think it changes the result in some of the theorems. *3*- I believe the lower bound in Corollary 2.4 is not tight, or at least it is not proven to be tight in the paper. One reason is the assumption that given $S$, labeled and unlabeled samples are independent. *4*- Is there any proof that the bound in Theorem 2.3 is tight? *5*- The proof for Theorem 2.6 is indexed as the proof of Theorem 4 in the supplementary material. *6*- In the proof of Theorem 2.6, the distributions $\mathbb{P}$ and $\mathbb{Q}$ are different from those in equations (8) and (9). Can you please explain this? Also, why should we test between the distributions in equations (8) and (9)? It seems that the distributions in the proof make more sense. *7*- In Theorem 3.2, is it sufficient to solve the problem just for the case where non-zero elements are $\pm\sqrt{\frac{\lambda}{k}}$?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Please check the questions.

Reviewer 3M9B9/10 · confidence 5/52024-07-05

Summary

The authors identify a regime where semi-supervised learning has computational advantage over (purely) supervised or unsupervised learning. They propose an algorithm that achieves it and demonstrates the efficacy of the algorithms in simulations.

Strengths

- Extremely well-written paper - The theoretical results I believe will become landmark results in the area of semi-supervised learning - The literature review is thorough

Weaknesses

- What is the answer to the question on L59, namely "On the computational side, is there a computational-statistical gap in SSL?" For the white "hole" in Figure 1, the computational-statistical gap remains, is this correct? It would be good to state this explicitly. - Also, it would be nice to add references to the results in Section 2 to the paragraph starting on L86. - It would be nice to rigorously explain the reduction on Eqn (3), even if in the appendix.

Questions

- Do you have a conjecture for the information theory lower bound in the white "hole" region? - Do you have any insights as to how one might analyze LS^2PCA?

Rating

9

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

Yes

Reviewer uoYb7/10 · confidence 4/52024-07-12

Summary

This paper studies the classical problem of clustering Gaussians with sparse means. While this problem was previously considered under the unsupervised (Sparse PCA) and supervised settings, the main innovation of the authors is to identify a phase where labeled and unlabeled data can be used together to estimate the cluster means, even when this task would be impossible using only either of them. This semi-supervised algorithm is complemented by various informational and computational lower bounds; in particular, the authors establish the existence of an information-computation gap for the detection task in the semi-supervised setting.They also provide numerical simulations for their algorithm, showing how to improve it by using out-of-the-box sparse-PCA algorithms in its second step.

Strengths

I found this paper very interesting to read; it considers a very interesting and natural question, and fills most of the blank parts of the phase diagram with novel results. The main achievement of the paper, the SSL algorithm, is both completely new and very easy to understand.

Weaknesses

I think the exposition of the paper could be slightly improved. In particular: - it would be nice to mention which part of Figure 1 comes from which theorem, and which are novel to the paper; - Theorem 2.3 and Corollary 2.4 are presented very dryly, and it is hard to understand exactly what they aim to prove (in particular because I'm not fully sure when $\alpha$ refers to an arbitrary constant or the sparsity level $\log_p(k)$). - I didn't exactly get the point of Theorem 2.1: it seems to give a lower bound at $\beta = 1/2$ instead of $1 - \alpha$, which is weaker in the regime $\alpha \in (0, 1/2)$. What does it bring compared to the results of Donoho and Jin ?

Questions

- You conjecture in the paper that the white region is actually computationally hard; do you have any (informal) arguments to support this conjecture ?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer NrxF6/10 · confidence 3/52024-07-12

Summary

The paper studies semi-supervised learning in a simple mixture of two Gauss(ians setting. Specifically, there is a uniform mixture of two Gaussians in $p$ dimensions with unknown means $N(\mu_1, I_p), N(\mu_{-1}, I_p)$. We assume that the difference between the means $\Delta\mu = \mu_1 - \mu_{-1}$ is a $k$-sparse vector and for simplicity, we also assume all nonzero entries are $\pm \sqrt{\lambda/k}$ for some parameter $\lambda$ that controls the signal-to-noise ratio. Now the learner receives both unlabeled samples from the mixture (which are just points in $\mathbb{R}^p$) and some labeled samples which also include the label of $+1$ or $-1$. The goal of the learner is to learn the support of $\Delta \mu$ (say with at most $o(k)$ errors) , which then gives a classifier that separates the two Gaussians. Information theoretically, with only labeled samples or only unlabeled samples, the optimal rates are well-understood (in a bit of an over-simplification and ignoring log factors, they are $\sim k/\lambda$ and $\sim k/\lambda^2$ respectively). If we let $L$ and $n$ be the number of labeled and unlabeled samples, the paper shows that interpolating between the two rates, using a mixed number of labeled vs unlabeled samples cannot help. Thus, up to constant factors, there is no benefit, information-theoretically, from combining labeled and unlabeled data. This portion of the results are very similar to [Tifrea et. al. 2023], which proves the same type of result in the setting with no sparsity. THe main contribution of the paper is identifying a regime where there are potential computational benefits to combining labeled and unlabeled data. There is evidence (namely low-degree/SQ lower bounds) that solving the unsupervised variant of the problem is computationally hard with fewer than $k^2/\lambda^2$ samples. Now consider a regime where we have a slightly subquadratic number, say $k^\gamma/\lambda^2$ for $\gamma < 2$, unlabeled samples. This paper shows that in this regime, even if we combine with a small amount of labeled data, the problem remains computationally hard. On the other hand, the paper also shows that if we combine with a larger amount of labeled data (but still not enough to just solve the supervised setting by itself), then the problem becomes computationally tractable. There is a constant factor gap between the amount of labeled data in the upper and lower bounds, but nevertheless, this identifies a region where combining labeled and unlabeled data makes the problem computationally tractable whereas using either only the labeled or unlabeled data individually would be intractable.

Strengths

The paper makes advances in an important research direction of theory for understanding semi-supervised learning The paper has a nice conceptual message in identifying a setting where there is a provable benefit to combining labeled and unlabeled data, proving both upper and lower bounds (under commonly believed computational assumptions). It is also nice that the main algorithm gives a clean way to combine two different types of information. The algorithm in the paper is simple enough to implement and the authors are able to support their conclusions with numerical experiments

Weaknesses

The paper only studies a limited setting of a mixture of two Gaussians The benefit from the unlabeled data, (in terms of reducing the number of labeled examples required), is only a fairly small (less than $2$) constant factor It is a bit technical to actually describe/understand the setting where the algorithm in the paper has provable gains

Questions

.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer 6RTo2024-08-10

I thank the authors for their answers. They addressed my questions, so I increased my score accordingly.

Reviewer NrxF2024-08-11

Thank you for the response and addressing my concerns/questions. Is it correct that if the parameters are bounded away from the boundaries of $\alpha = 1/2, \gamma = 2$ then the gain is only a constant factor?

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC