Efficient Leverage Score Sampling for Tensor Train Decomposition

Tensor Train~(TT) decomposition is widely used in the machine learning and quantum physics communities as a popular tool to efficiently compress high-dimensional tensor data. In this paper, we propose an efficient algorithm to accelerate computing the TT decomposition with the Alternating Least Squares (ALS) algorithm relying on exact leverage scores sampling. For this purpose, we propose a data structure that allows us to efficiently sample from the tensor with time complexity logarithmic in the tensor size. Our contribution specifically leverages the canonical form of the TT decomposition. By maintaining the canonical form through each iteration of ALS, we can efficiently compute (and sample from) the leverage scores, thus achieving significant speed-up in solving each sketched least-square problem. Experiments on synthetic and real data on dense and sparse tensors demonstrate that our method outperforms SVD-based and ALS-based algorithms.

Paper

Similar papers

Peer review

Reviewer 8ErW7/10 · confidence 3/52024-07-07

Summary

This paper gives a better randomized alternating least squares algorithm for computing tensor factorizations. It's based on an exact characterization of leverage scores of the matrization of tensors via a suitable intermediate orthonormal representation. This is justified rigorously, and significant empirical gains were demonstrated.

Strengths

The approach taken is natural, and the bounds obtained are quite powerful. The experiments considered both synthetic and real data sets, and demonstrated clear gains in the parameter regimes considered.

Weaknesses

I'm a bit concerned about the setting of the experiments, which seem to be 3-dimensional dense tensors. My understanding is that a lot of the more complicated tensor instances are sparse and in higher dimensions. However, I'm not sure whether those have low rank representations.

Questions

Would it be possible to check how the orthogonality conditions are maintained in intermediate steps under the inexact arithmetic caused by round off errors? Aka. are the conditions for the characterizations of leverage scores preserved exactly? (this has been answered)

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

yes

Reviewer oeCx6/10 · confidence 4/52024-07-11

Summary

The authors proposed a leverage score sampling-based TT-ALS method to reduce the computational complexity of the traditional TT-ALS. Experimental results verify the performance of the proposed method.

Strengths

The paper is well written with good theoretical analysis and desired experimental performance. The method using leveraging score sampling is technically sound for TT decomposition.

Weaknesses

The contribution and novelty are not clearly stated compared with [Malik and Becker, 2021]. In [Malik and Becker, 2021], the leverage sampling was applied to TR. As TT can be treated as a special case of TR, the author does not clearly state that it is necessary to develop a new method for TT, or that there are some new theories or findings that are distinct from the previous TR structure.

Questions

1. My main concern about this work is its relation to the leverage score sampling-based TR-ALS [Malik and Becker, 2021]. As TT can be seen as a special case of TR, what is the main contribution of the current work compared with leverage score sampling-based TR-ALS? Please clarify that it is necessary to develop a new method beyond [Malik and Becker, 2021] (i.e. has a better theoretical guarantee or has better performance than leverage score sampling-based TR-ALS). 2. For real data experiments, how about the performance compared with other tensor decomposition structures, such as TR? As you can set the first and last TR-rank as 1 thus it reduces to a TT-like structure.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A

Reviewer BTzr4/10 · confidence 3/52024-07-12

Summary

This paper presents an efficient algorithm to use leverage-score sampling to solve least squares problems arising as subproblems in a larger alternating least squares (ALS) algorithm for building an approximate Tensor Train (TT) decomposition. The paper reports empirical evaluation of the proposed algorithm, showing runtime improvement compare to the baseline un-sampled problems.

Strengths

- Empirical evaluation show a runtime improvement with respect to unsketched TT-ALS, without reducing fit. - The paper presents an efficient algorithm for leverage score sampling of ALS subproblems in TT-ALS.

Weaknesses

Major comments (affecting recommendation): - The proposed method is compared empirically only to a small number of baselines. Seems that the most likely competatitor (Chen et al. 2023) is not compared against. - Runtime improvement in the experiments is modest. This is a sampling based method, so if sampling is very efficient (due to efficient leverage score approximation) I would expect a big improvement in runtime. This might not be the case if either more iterations are now needed for convergence or computing the sampling probabilities is very expensive. Unclear from then experiments which is the case. - The paper does not present an analysis end-to-end. It only analyzes how to subsample the ALS subproblems, and thus accelerate them. Not analyzed is how the solution of the subsampled system in lieu of the full subproblem affect ALS convergence. - Many details required for implementation of the algorithm are omitted. Many points regarding the algorithm are not clear. Minor comments (do not affect the recommendation): - Lines 125 - 126: add citation. - Figure 3 left: is this fit in the y-axis or misfit? - Line 108-109: I think you mean *omission* of multiplicative terms. - Lines 58-60: shouldn't runtime be O(jIR^3)? - Line 54: The term "left matricization" is used without defining it yet. - Line 55: I think it should be R_j - Eq (2): I think on the right it should be A[i,:]^T. Similar issue with Eq (3).

Questions

- Lines 96-97: Sketching is a good way to circumvent the need to compute leverage scores, and TensorSketch allows us to sketch matrices with Kronecker structure. And indeed, the authors point out that [Chen et al. 2023] do this. How do their results compare to the present paper? Why is what you are proposing a better idea than using TensorSketch? - Lines 64-66: What is the significance of this observation regarding the complexity of the algorithm? - Lines 125 - 126: One the canonical form has been computed, can we solve the ALS subproblem efficiently? - Did you compare empirically your method to the method of [Chen et al. 2023]. - Is your algorithm computing the canonical form anew in each ALS iteration, or are you somehow updating it? - Does your algorithm maintain/computes the leverage scores of each row (the number of such scores is exponential), or do you present only an efficient way to sample from it without forming it?

Rating

4

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Nothing to add.

Reviewer yjyo7/10 · confidence 4/52024-07-12

Summary

This paper considers the problem computing the tensor train TT decomposition of large tensors and proposes a novel randomized approach for efficiently solving the problem. In particular, the Alternating Least Squares (TT-ALS) algorithm is considered and exact leverage scores sampling approach is proposed to accelerate the approach. A data structure based approach is devised to efficiently compute the leverage scores of the appropriate matricization of the tensors and sample rows from them. Theoretical analysis show the sampling complexity. Numerical experiment results are presented which to illustrate the performance of the proposed algorithm.

Strengths

The strengths of the paper are: 1. A novel randomized approach is presented for computing the TT decomposition of large tensors. 2. New data structure and sampling approach are proposed. Theoretical results are presented. 3. Numerical results show that the proposed method is efficient.

Weaknesses

The weaknesses of the paper are: 1. The randomized SVD approach seems to be more efficient in terms cost compared the the prosed method. 2. Certain aspects of the paper can be improved (see below).

Questions

The paper presents an interesting approach to compute the TT decomposition of large tensors. However, I have the following comments that can further improve the paper: 1. r-TT-SVD seems better - Based on numerical results presented, it appears the randomized SVD based approach (r-TT-SVD) takes significantly lower time, yet achieves comparable performance (fit) as other methods. Are TT-ALS and rTT-ALS popular and relevant for TT decomposition. How do these methods differ in the quality of the decompositions computed? In some cases, rTT-SVD might have a slightly lower fit. But, since the runtime is so low, perhaps a larger sketch size would yield similar fit at lower cost. Details about the sketch sizes used are missing. 2. Presentation: Certain aspects of the presentation can be improved, and there are few minor typos. i. Are TT decompositions popular in tensor applications? Overall, they seem to have higher computational cost and are not very interpretable. Few alternate tensor decompositions achieve better compression with lower cost. A discussion on the motivation for the use of TT decomposition will benefit the paper (readers can better appreciate the results). ii. Introduction has tensor jargons which might not be known to general AI audience, such as 3D core chains, left-matricization, contraction of tensor, etc. It is better to avoid these or define them before use. iii. The computational cost of solving eq (1) can be added. iv. Theorem 1.1, point 1 has a $j$ missing in the runtime cost. v. Not sure what is the subscript (2) in Corollary 4.4. Many of the results in the paper is based on another Arxiv paper (which is not peer-reviewed). So, correctness of these results is not established.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See above comments ------ Post Rebuttal: I have read other reviews and authors' responses to all reviews. The responses have adequately addressed my concerns. The work has many merits. I have raised my score.

Reviewer oeCx2024-08-12

Thank the reviewer for their reply. I've increased the score.

Area Chair DqxZ2024-08-12

Reviewers: The author-reviewer discussions end tomorrow (August 13), so please read the authors' rebuttal carefully and respond if you haven't.

Authorsrebuttal2024-08-12

Dear reviewer, considering that the end of the discussion period is approaching, do you have any further questions? We think we have answered all your concerns in our rebuttal and hope you'll consider increasing your score.

Reviewer 8ErW2024-08-12

thank you

Thank you for the detailed responses. My concern w.r.t. roundoffs is that equality conditions like equation (6) might not be robust to perturbations to the original matrix. However, I now see that it's only for the initial matrix, not for the intermediate products of sampling. So I agree that it's inherently robust. I will raise my presentation score, and leave the overall unchanged.

Reviewer BTzr2024-08-13

[W1][Comparison against (Chen et al. 2023)]

Thank you for the detailed answers and clarifications. It will be beneficial to include many of the clarifications in revised versions of the manuscripts. I will consider how your answers affect my final score, which is "borderline" to begin with. However, let me followup the discussion on Chen et al. 2023. I see your point. However, note that Thm 4.4 in Chen et al. 2023 gives exact constants (no O() expression) and this is where the 3^q comes from. In contrast, your results is J=O(R^2/(\eps * \delta)) where the constant is unclear. For all I know, the O() hides a 3^q constant! So: if you flesh out the constant, do you really do not have any dependence on the tensor order? Re empirical comparison: even if we accept that there is a theoretical advantage in the form of a dependence of the order, this does not necessarily tell the whole story for a small q (where typically q=3 is the most common use-case). In fact, if q is fixed to 3, then 3^q is now a constant, and it really matters how it compares the constant of the your algorithm. I still hold that Chen et al. 2023 is the most appropriate empirical baseline, and without it the empirical evaluation is lacking.

Authorsrebuttal2024-08-13

>Thank you for the detailed answers and clarifications. It will be beneficial to include many of the clarifications in revised versions of the manuscripts. Thank you again for you review and questions. We will include all the points mentioned in our rebuttal in the revisions which will improve the clarity. > However, let me followup the discussion on Chen et al. 2023. I see your point. However, note that Thm 4.4 in Chen et al. 2023 gives exact constants (no O() expression) and this is where the 3^q comes from. In contrast, your results is J=O(R^2/(\eps * \delta)) where the constant is unclear. For all I know, the O() hides a 3^q constant! So: if you flesh out the constant, do you really do not have any dependence on the tensor order? There is no hidden dependency on the tensor order in the big O. In Corollary 4.4, the number of samples required for each least square problem, J=O(R^2/(\eps * \delta)), is independent on the tensor order N (N is not hidden in the constant). Thus, for one sweep of the ALS algorithm, the complexity is linear in the tensor order, as stated in the second part of the corollary (i.e. the overall ALS complexity is O(#it εδ R^4 N \sum_{j=1^N} (log I_j + I_j), where N is the order of the tensor). We will clarify this point in the revision. > even if we accept that there is a theoretical advantage in the form of a dependence of the order, this does not necessarily tell the whole story for a small q (where typically q=3 is the most common use-case). In fact, if q is fixed to 3, then 3^q is now a constant, and it really matters how it compares the constant of the your algorithm. We agree that if q is treated as a constant, then the asymptotic complexity is the same. However, we believe that designing an algorithm with only a linear dependency on the order, instead of the exponential one given in Chen et al, constitute a non-trivial contribution and advancement compared to Chen et al. In addition, the TT decomposition is particularly suited and relevant for (very) high order tensors where the tensor order cannot be treated as constant (as demonstrated by its uses in quantum physics for simulating many-body systems where the order corresponds to the number of particles in the system). > I still hold that Chen et al. 2023 is the most appropriate empirical baseline, and without it the empirical evaluation is lacking. Unfortunately, the lack of a publicly available code for the method presented in (Chen et al, 2023) did not allow us to include it in the synthetic experiment. Moreover, to the best of our understanding, the method proposed by Chen et. al could not scale to the size of the tensors that we consider in the large sparse tensors experiments.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC