Fast Exact Leverage Score Sampling from Khatri-Rao Products with Applications to Tensor Decomposition

We present a data structure to randomly sample rows from the Khatri-Rao product of several matrices according to the exact distribution of its leverage scores. Our proposed sampler draws each row in time logarithmic in the height of the Khatri-Rao product and quadratic in its column count, with persistent space overhead at most the size of the input matrices. As a result, it tractably draws samples even when the matrices forming the Khatri-Rao product have tens of millions of rows each. When used to sketch the linear least squares problems arising in CANDECOMP / PARAFAC tensor decomposition, our method achieves lower asymptotic complexity per solve than recent state-of-the-art methods. Experiments on billion-scale sparse tensors validate our claims, with our algorithm achieving higher accuracy than competing methods as the decomposition rank grows.

Paper

References (36)

Scroll for more · 24 remaining

Similar papers

Peer review

Reviewer LxVj5/10 · confidence 4/52023-06-17

Summary

This work studies fast, exact leverage score sampling for Khatri-Rao product matrices (given access to its factor matrices) and an application to CP tensor decomposition via alternating least squares with row sampling. The main theoretical contribution is a data structure for sampling from the leverage score distribution of a Khatri-Rao product by iteratively sampling from the correct conditional marginal distribution as it constructs the row sample (one factor at a time). To apply this most effectively to CP decompositions, the authors propose a two-step version of this sampler that first samples a rank-1 eigenspace of the partial Gram factor matrices. The authors provide a good comparison of their work to the leverage score-based CP decomposition algorithm of Larsen and Kolda (Journal of Matrix Analysis and Applications, 2022).

Strengths

- The sparse factor matrix version of Theorem 1.1 complements the paper. - The comparison with Woodruff-Zandieh [27] is valuable and discusses the shortcomings of the theory-heavy results (i.e., hidden second-order constants that blow up). - The paper is very nicely organized. - The two-stage eigenspace sampling idea in Section 3.2 is a solid research contribution.

Weaknesses

- Theorem 3.1 is one of the main theory components of the paper, but largely comes from [Malik, ICML 2022]. - Section 3.1 could benefit from a formal description of what the leaf nodes represent. In particular, what is $S_{0}(v)$ and how does this partitioning get decided (i.e., where to draw the cut points)? It seems reasonable that the result is correct, just not immediately implementable by the reader. - In Corollary 3.3, it would be better to use the original expression $R=O(R \max\{\log(R/\delta), 1/(\varepsilon \delta)\})$ samples so as to not cause confusion by the implicit case assumption. - Re experiments: It would be good to compare the running time of this sampling-based method with both ALS-ARLS-LEV and an out-of-the-box ALS implementation (e.g., MATLAB or Tensorly). Analyzing the fit of the decomposition in isolation doesn't tell the full story.

Questions

**Questions** - Re Theorem 1.1: If a single entry of $U_j$ changes, is there a faster update time? - Should $I$ in Table 1 be $I_j$ to represent the $j$-th factor update? - In Line 81, we have $S \in \mathbb{R}^{J \times I}$, so shouldn't we have $S[j,i]$ in the following line? - In Line 95, what is the tilde in the lower bound for $J$? Related: Line 99 should be $\Omega(R/(\varepsilon \delta))$ since this is a lower bound. - In the discussion about Kronecker regression, Cheng et al. [4] and Larsen-Kolda [12] sample from the leverage score distribution of the Kronecker product for their CP decomposition algorithms (i.e., the product of the leverage scores). This connection seems worth pointing out. - In Line 337, shouldn't the fixed sample count increase with the rank? This is needed for the leverage score sampling guarantees. **Typos and suggestions** - [line 14] "denoted" --> "denoted by" - [line 43] Suggestion: Can remove the parenthesis around the summand in the second part of Theorem 1.1. The sentence "The structure can also draw samples..." can probably be generalized to say "... for any subset of factor matrices"? - [line 49] You say that "our applications deal with dense inputs," but this is immediately followed by mention of the sparse Amazon tensor experiments. The writing here could be improved. - [line 55] Suggestion: Add a citation for CP-ARLS-LEV after you first mention it. - [line 58] Table 1 could benefit from $\tilde{O}$ notation to account for missing constant factors. The description of what the complexity is for could be made more explicit too: "Complexity of factor matrix $U_j$ for $N$-dimensional dense CP decomposition..." - [line 64] suggestion: $j$'th --> $j$-th, same for later occurrences - [line 79] suggestion: "sampling operators" --> "row sampling operators" - [line 131] "Kronecker regression is distinct" --> "is a distinct" - [line 141] suggestion: "autoregressive fashion" is an indirect way to explain the procedure. Consider dropping this phrase and merging the two sentences. - [line 143] $I_n$ have been scalars so far. Therefore, it is better to use $(i_1, \dots, i_N) \in [I_1] \times \dots \times [I_N]$ to denote the set of indices. - [line 144] suggestion: consider restating the dimension of $G_k$ and $G$ to help the reader remember what some of these operators mean, e.g., $G := (\text{expression}) \in \mathbb{R}^{R \times R}$. - [line 153] Please include the citation to Malik 2022 too, so that the reader can easily click through to the references. - [line 162] Typo: "theorem 1.1" --> "Theorem 1.1" - [line 165] Suggestion: Rewrite the sentence as: Let $h \in \mathbb{R}^{R}$ be a vector and let $Y \in \mathbb{R}^{R \times R}$ be a ... [delete "respectively"].

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer qocR7/10 · confidence 3/52023-06-29

Summary

The paper studies algorithms to efficiently sample a row from a matrix $A = U_1 \odot U_2 \odot ... \odot U_N$ with probability proportional to the leverage scores of $A$. Here, $B \odot C$ denotes the _Khatri-Rao_ product of $B$ and $C$, or the column-wise kronecker-product of the columns of $B$ and $C$. In particular, this means that $A$ has a small number of columns $R$, but has an exponential number of rows $\prod_{j=1}^N |I_j|$ (where $U_j \in \mathbb{R}^{|I_j| \times R}$). This exponential blowup in the number of rows of $A$ in the core computational focus on this paper, as well as several related papers in the area. This paper's technical contribution is in proposing a data structure that can preprocess $U_1, ..., U_N$ such that a row of $A$ can be sampled with probability proportional to its leverage score. This data structure's space complexity matches the sizes of the input, and at query time it takes only $O(NR^2 \log \max \\{I_j, R\\})$ amortized time per row sampled. This new data structure is based on "binary-tree inversion sampling", a binary search algorithm that can help sample from distributions efficiently, and is used to speed up a computational bottleneck in a known approach to sample leverage scores from Khatri-Rao products. Theoretical results demonstrate the correctness of this algorithm, and experiments demonstrate the effectiveness of this algorithm.

Strengths

The paper is well written, nicely motivated, pretty easy to understand, has a clear presentation of the subproblem it speeds up, and provides good context relative to existing work in the area (afaik; I'm no expert in tensor algorithms). **Originality and Quality:** A fine line needs to be drawn here, and this is the subtlest point of my review. Per the authors' account, existing work in leverage score sampling from Khatri-Rao products exists and provides nice results. A good example is Theorem 3.1 from this paper, which is an adaptation of an observation made by prior work about a particular way to compute the leverage scores of $U_1 \odot ... \odot U_N$. The core of this paper is taking Theorem 3.1 and finding a faster way to compute it's right hand side. This means that the prior work gave much of the framework for around paper's proposed algorithm, and we can sorta think of this paper's contribution as finding a novel way to compute a subroutine more quickly. That said, the approach to computing the subroutine is novel and interesting. I've not seen prior work in this area use a binary-tree in such a way. There's clearly a good deal of effort put into the math (though I only verified bits and pieces of the appendix). The paper very much stands as sufficiently original, but I want to be clear that the originality lands firmly in section 3.1 of the paper -- how to efficiently sample following this framework that existed in prior work. **Clarity:** The paper is well written. I have only a couple mild gripes about writing and presentation (listed later in the "Questions" section), and I feel that I understood all the big ideas of this paper. **Significance:** This paper feels well motivated, and the improvement to sample a row with runtime that depends logarithmically on the number of rows in $A$ is quiet strong and interesting. The experiments suggest that this truly is a state-of-the-art algorithm for a meaningful suite of metrics. I do have some gripes with the presentation of the experiments, which undercuts my confidence in the experimental evidence a smidge, but I think this both not essential and can be easily corrected.

Weaknesses

The weaknesses are few and far between in this paper. A couple section lack a bit of clarity, like how the big-Oh rates for some of the downstream tensor algorithms are insufficiently explained. These are issues that can easily be fixed for a camera-ready version of the paper. Out of these mild gripes, there are three that do stand out a bit: 1. Several experiments lack confidence intervals. This muddies the story of the algorithm's empirical efficacy, especially in Figure 3, where it feels very confusing to see the error of the green curve in the left figure jump up significantly. 2. In a few place, discussion of big-Oh runtime isn't fully explained. For instance, I don't understand how the complexity of STS-CP was derived on Table 1 (appendix A.1). More details are in the "Questions" section. 3. The last paragraph from section 3.2 takes the core technical ideas and is supposed to summarize the algorithm. However, it instead introduces some new notation and makes me wonder if it's actually summarizing an algorithm or if it's instead trying to briefly explain some of the deeper technical edgecases the algorithm has to handle. Either way, it's pretty confusing of a paragraph for me to read, and undercuts my confidence in understanding the full algorithm. None of these are serious gripes, and call all be fixed with mild updates to the presentation. So I'm still very positive on this paper!

Questions

I don't really have deep concerns to ask about here. I'll just enumerate a long list of minor typos and confusions. 1. Why use (afaik) non-standard notation, like having $d$ be the dimension (number of cols) of A, or $p_i$ be the probability of sampling row $i$, or $m$ or $c$ or $s$ being the number of sampled rows? 1. [Line 30] As someone familiar with leverage scores, I was surprised not to see a log appear in the sample complexity of leverage score sampling. I'd make it into a $\tilde O$. 1. [Lines 49 and 52] Line 49 says that the applications mainly deal with dense inputs. Line 52 says that the most practical benefit is on sparse matrices. These aren't necessarily statements that are at-odds, but it does feel like a bit of whiplash to read these statements back-to-back. The argument on lines 279-280 would be good to add here. 1. [Line 55] Do you have any understanding of why your algorithm which boasts a new and much smaller big-Oh term has 2% slower runtime? 1. [Table 1] I dunno how you came to the complexity-per-iteration here. As a reader, I felt like I should be able to pattern-match between Theorem 1.1 and Table 1 to understand the complexity of STS-CP, but I really couldn't make them line up. 1. [Line 131] "is **a** distinct" 1. [Line 132] "There" not "Here" 1. [Line 194] Get rid of the square on $[0,1]^2$ 1. [Line 132] Explicitly argue why $F=1$ is affordable space-wise now, but was too expensive in the setting at the start of section 3.2 1. [Line 241] What is $Z_j$? 1. [Figures 2,3,5] Add confidence intervals 1. [Figure 4] Swap the x and y axis. It's hard to read with fit being on the x-axis. Add some space between the subplots -- they're real strange shoved together like that, making it harder to read. 1. [Line 344] Well... STS-CP does have an oscillating error pattern on Amazon too. Maybe mention it? Eh, I'm sorta torn here. 1. [Algo 6, line 5] Missing a period in $1,..,N$

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer Njzt7/10 · confidence 4/52023-07-01

Summary

This work proposes a new algorithm to efficiently perform exact leverage score sampling on Khatri-Rao products. This work is built on top of the TNS-CP algorithm (Malik et al., 2022), and a more efficient data structure is used to achieve better sampling computational cost. This algorithm can be used to accelerate sketching based alternating least squares (ALS) for CP decomposition, and it has been shown that state-of-the-art complexity per ALS sweep has been achieved. Experimental results show that this sampler is efficient and yields better accuracy than previous leverage score sampling based CP-ALS algorithms for large real sparse tensors.

Strengths

1. Accelerating leverage score sampling and large scale CP decomposition is an important topic, and this work achieves state-of-the-art results in both the theoretical analysis and the experimental results. 2. The data structure used to accelerate leverage score sampling of Khatri-Rao product is novel.

Weaknesses

Overall I believe this is a good contribution. I only have one minor comment: it takes me a while to figure out the logics in Sections 3.1 and 3.2, and I think adding a figure to summarize the two sampling process would be good.

Questions

n/a

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

n/a

Reviewer u8zc5/10 · confidence 3/52023-07-05

Summary

This paper designs a method for efficiently minimizing least squares regression, when the design matrix is equal to the Khatri-Rao product of multiple matrices. Because this Khatri-Rao product can have a huge dimension, the paper is motivated to study sampling techniques to reduce the dimension of the product matrix. This least squares problem is motivated by applications to tensor decomposition, in particular when alternating least squares is used for finding CP tensor dimension. Each step of the alternating least squares corresponds to a least squares regression, where the design matrix is the Khatri-Rao product of the other tensor factors in the CP dimension. To solve this sketching problem, the paper adapts the leverage score sampling technique to this Khatri-Rao product design matrix. The algorithm scales in a linear order proportionally to the sum of the dimensions of each matrix in the product. Experiments are provided, comparing the approach to methods from a prior work by Larsen and Kolda (SIAM J. Matrix Analysis and Applications (2022)). The results show moderate improvement in terms of sparse tensor decomposition.

Strengths

S1) The main is mainly of a technical nature: the algorithm, which is based on detailed calculations of the Khatri-Rao product, appears sound, and this is validated in the experiments. S2) The sampling from Khatri-Rao product relies on a binary-tree inversion sampling technique, and this tree construction has time cost $O(I R^2)$ and storage cost $O(R^2 I / F)$. S3) The experiments compared against the method of Larsen and Kolda appear significant, especially at higher target rank values, for several sparse tensors with as many as $10^9$ nonzeros.

Weaknesses

W1) From my reading, the paper does not do a good job of motivating the problem. For instance, the main motivating example is tensor decomposition, but the connection is spelled explicitly only until Section 3.3. Are there any other plausible examples for motivating the proposed problem? W2) I find the derivation steps to be quite laborious but also quite tedious. I find the illustration of Figure 1 to be very useful, and I think that a better presentation that conveys the key steps would be very helpful, for instance, in a special case of the Khatri-Rao product. W3) The significance of the results is quite limited; in my opinion, I think the paper could benefit from stronger statements about how significant the contributions are. Solely judged based on the experiments, the improvement is quite weak since only one baseline is considered in the experiments. From a theoretical standpoint, the comparison stated in the related work could be made more clear as well.

Questions

- It would be nice to give some description of the tensor used in the experiments (without having to refer to the appendix). - I understand this question may be out of scope, but it would be nice to compare fit against other tensor decompositions, e.g., the Tucker decomposition. Would the newly developed techniques apply to Tucker decomposition? - It is mentioned in the related work that this paper is very closely related to a method by Woodruff and Zandieh (2022). I wonder how the methods would compare against each other in the experiments.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, the paper discussed limitations in Section 5. Due to the technical nature, I do not see any potential societal impact of their work.

Reviewer LxVj2023-08-13

Thank you for the detailed response, especially the runtime comparison table. I also read all the other reviews and author responses, and am updating my rating to 5 (borderline accept).

Reviewer Njzt2023-08-19

Thank you for the detailed response, I will keep my score.

Reviewer qocR2023-08-21

Thanks for the response!

Sorry for the delay in my response. I enjoyed the authors' message, and happily maintain my score. The paper should be accepted. I'm glad to see the change in the figures, and the detailed responses in the bullet points above are well written. Some quick notes: 1 . If this is standard in parts of the tensor literature, then feel no need to change that notation. I'm not used to it, but that's a me problem. 4 . I really like that paragraph you inserted it. Maybe say ~65000 instead of $2^16$, just because $2^16$ feels like it should be much larger of a number than it actually is? Eh, kinda torn on this. 11 . I enjoy the updated figures, and they contain nice trend data. It's perfectly sufficient for a rebuttal, but for a camera-ready draft I think I'd recommend swapping the stars out -- they're hard to see unless I zoom in quite a bit. 13 . Strange, a good point from the frequency plot perspective. It might be worth acknowledging the distinction on line 345, but it's really up to you.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC