Many-body Approximation for Non-negative Tensors

We present an alternative approach to decompose non-negative tensors, called many-body approximation. Traditional decomposition methods assume low-rankness in the representation, resulting in difficulties in global optimization and target rank selection. We avoid these problems by energy-based modeling of tensors, where a tensor and its mode correspond to a probability distribution and a random variable, respectively. Our model can be globally optimized in terms of the KL divergence minimization by taking the interaction between variables (that is, modes), into account that can be tuned more intuitively than ranks. Furthermore, we visualize interactions between modes as tensor networks and reveal a nontrivial relationship between many-body approximation and low-rank approximation. We demonstrate the effectiveness of our approach in tensor completion and approximation.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer wHyr6/10 · confidence 3/52023-07-02

Summary

Authors propose a tensor decomposition approach based on the Legendre decomposition and convex optimization with natural gradient-based algorithm for non-negative tensors which is inspired by many-body interactions in physics. The tensor is interpreted as a probability measure over a multidimensional discrete space given by its indices, and the complexity is regulated by the order and map of interactions between variables (this makes it possible to avoid the difficult problem of choosing the optimal rank of the decomposition). Experimental results demonstrating the good properties in terms of accuracy and computation time of the proposed method in comparison with a number of alternative methods of low-rank approximations.

Strengths

The proposed method looks novel and interesting, and it has benefits like a convex optimization formulation and a lack of rank parameters to tune. The work is well structured, written in a good style and easy to read.

Weaknesses

1. From the presented numerical experiments, the actual advantages of the proposed approach over baselines are not quite clear to me. For the case of a dataset with graphic images, what is the actual quality of the low-rank approximation (it would be useful to give examples of reconstructed images)? 2. As far as I know, there are effective rank-adaptive methods of low-rank approximations (rank-adaptive TT-cross method, rank-adaptive tensor ring, etc.), see, e.g., "Adaptive rank selection for tensor ring decomposition" (2021). Will the proposed approach be superior to such modern methods? 3. I did not see estimates of the computational complexity of the proposed method. What dimensions limit the applicability of the approach, i.e., can we use it, for example, for the 100-dim tensors?

Questions

1. [Line 23] "More recently, tensor networks [4] have been introduced...". I do not fully agree with this position. Tensor networks have been used for a very long time in the scientific community, including quantum applications (e.g., DMRG), and also the noted decompositions (CP, Tucker) represent a case of the tensor network. It seems to me that this paragraph should be substantially reformulated. It also seems appropriate to note such decompositions as MERA, PEPS, etc. 2. [Line 268] It might be worth refining the positioning of the graphics so that the section title (3.2) isn't at the end of the page. 3. [Line 330] "error[s] are..." 4. [Lines 349-350] "visualizing interactions between modes ... to see activated interactions between modes". Perhaps this sentence should be reformulated to avoid repeating the phrase.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The method can only be applied to tensors with non-negative elements. There are also likely to be problems associated with high computational complexity when trying to apply this approach to significantly multidimensional data arrays.

Reviewer gp7Q6/10 · confidence 4/52023-07-06

Summary

The paper proposes a new type of tensor decomposition, based on considering an interaction network among the different modes of a tensor. Nonnegative tensors are considered and factorized based on a Legendre decomposition, which may be viewed as representing the tensor elements by an underlying probability distribution. The paper then proposes to model the tensor in terms of interactions, each of which couples two or more modes. These interactions may be represented by a tensor network, which contains no rank modes, e.g., T_{ijkl}=U_{ij}V_{jk}W_{kl}. By contrast, typical tensor decompositions, include auxiliary/contracted modes. Using the theory of probability distributions, the authors argue that the resulting optimization problem to factorize the tensor according to interaction network, is convex. A gradient based method is used for optimization. Connections are made to other types of tensor networks, demonstrating that the interaction network arises by imposing additional structure on such networks. Experimental results show that, on the datasets considered, the interaction/many-body approximation method can be competitive in terms of accuracy.

Strengths

* The proposed tensor decomposition type appears to be new and is interesting / has connections to probability theory and physics. * The use of probability theory to demonstrate convexity of the optimization problem is innovative, however, it appears to me that the assertion of convexity is incorrect, see weaknesses. [After reviewing the rebuttal, I follow at least the high-level motivation for why the approach is convex and revised my evaluationp]

Weaknesses

* The Legendre decomposition is defined vaguely in the paper and supplementary material, and as far as I can tell this is not a broadly known concept (discussed only in a recent line of ML papers). This makes the paper not self-contained, and makes it difficult to ensure correctness of the arguments. The optimization method and proof of convexity of the optimization problem are also proven hastily / with lack of good definition/theorem/lemma/proof structure, in particular restate the theorems being used from prior work (e.g., when discussing results based on linear dependence of parameters in the distribution). * My main concern is that from a tensor point of view, it seems impossible that decomposing a tensor into a many-body approximation for an arbitrary interaction network leads to a convex optimization problem. Here is a counterexample. Consider a sixth order tensor with 3 interaction terms, each acting on two independent modes, T_{ijklmn}=U_{ij}V_{kl}W_{mn}. If the dimensions of modes j, l, and n are 1, we've obtained a rank-1 approximation of the tensor T'_{ikm}=T_{i1k1m1}=U_{i1}V_{k1}W_{m1}. Its well known that rank-1 approximation, which also corresponds to the tensor eigenvalue problem, is nonconvex (often has many local minima). [After reviewing the rebuttal, I see that KL-divergence enables convexity based on prior results.]

Questions

Please provide a counterargument to my assertion above that the optimization problem is in fact nonconvex. If I have somehow overlooked something on this, I would be much more supportive of the paper. If its indeed nonconvex, the paper needs to be corrected and re-evaluated.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

4 excellent

Limitations

None.

Reviewer oVis7/10 · confidence 3/52023-07-08

Summary

This work introduces a novel approach to non-negative tensor decomposition, termed "many-body approximation," which specifically addresses the relationship among modes of tensors. It is formulated as a variant of Legendre decomposition and realized through globally minimizes the Kullback-Leibler divergence. Additionally, this work explores the connections between the proposed method and existing methods, using interaction representation. A series of experimental results substantiate the superiority of the proposed method over other existing ones in dealing with tensor completion applications.

Strengths

1. This paper is comprehensive and includes all the essential components. The structure is logical, and the design principle is effectively and adequately illustrated. 2. The proposed method is a novel rank-free method, which is particularly advantageous as identifying an appropriate low-rank structure can be highly challenging in some practical applications. 3. This paper visualizes the presence or absence of interactions between modes and further demonstrates the interpretability of interactions among all the modes.

Weaknesses

1. The proposed decomposition is only applicable to non-negative tensors. This may restrict the method's effectiveness and generalizability when dealing with tensors that contain negative values. 2. Some theoretical analysis about the proposed "many-body approximation"method is missed.

Questions

Is it possible to derive some theoretical results to support the merits of the proposed "many-body approximation" method in dealing with tensor completion and approximation problems?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

See the Weakness\Questions above.

Reviewer 59t85/10 · confidence 2/52023-07-11

Summary

The authors introduce a new, energy-model based approach to the decomposition of non-negative tensors. They compare their method to mainstay techniques.

Strengths

The technique seems solid and reasonable, but how solid and reasonable is hard to assess (see below).

Weaknesses

As presented, it's hard for me to determine how novel and distinct the presented method is. I work extensively with tensors, but as practical tools, so I am not fully versed in the various techniques and advantages for the various decomposition technique. For me to determine this, I would need a much more extensive literature review (perhaps adding a table) and to compare the new technique much more thoroughly and extensively to existing techniques (to highly the similarities, differences, advantages, and disadvantages).

Questions

My questions all relate to the similarities and structures of the other techniques.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

2 fair

Contribution

2 fair

Limitations

This is not an issue. The work presented by the authors is a general tool.

Reviewer xQBv6/10 · confidence 3/52023-08-05

Summary

This paper proposes a novel many-body decomposition, based on Legendre decomposition, for non-negative tensor. Instead of specifying a rank parameter, the proposed method use energy-based model to the interactions among modes.

Strengths

1. The paper is overall well-written and easy to follow. Especially the figures that explain the interactions. 2. The work is technically sound and aims at an interesting angle of tensor decomposition which is worth more attention from the research community. 3. The proposed method doesn't require the rank parameter which can save a headache of parameter tuning in many existing tensor analyses.

Weaknesses

1. The novelty of the proposed method is limited given the existing Legendre decomposition. 2. The empirical advantages of the proposed methods are not sufficiently demonstrated. 3. The rank parameter is not necessarily required in traditional tensor methods, for example, tensor nuclear norm based approach. 4. The details of the proposed algorithm, LBTC, should be more emphasized in the main text. 5. How is the computational cost of the proposed approach, compared to the traditional tensor decompositions? There were many fast methods have been proposed for rank based decomposition, such as tensor CUR decompositions. 6. If I understand right, the proposed method is indeed a generalized rank-1 approximation. 7. "interactions between modes" -> "interactions among modes"

Questions

See weakness section

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

1. As the authors already stated in the paper, the proposed method can only be applied to non-negative tensors. 2. Compared to rank-based tensor decomposition, the proposed method seems to require more expert knowledge about the application itself.

Reviewer xQBv2023-08-10

Thank you for the authors' response. It has addressed most of my concerns. I will keep my positive score.

Reviewer 59t82023-08-10

Referee Response

The table is helpful, although it would be nice to have it compare other facets than just global optima and uniqueness of solution, and including additional discussion on the details would have been helpful as well. I'll keep my mildly positive review score.

Authorsrebuttal2023-08-17

We appreciate your additional feedback on our response and positive review. Here we provide further discussion of tensor decomposition methods. In general, it is difficult to find out the low-rank structure of a given data set. Therefore, how to choose a low-rank decomposition model is hard to answer. This is why it is nontrivial to state the general advantages and disadvantages of each model itself despite various studies on optimization methods and regularization. To alleviate this difficulty in model selection and provide intuitive modeling, we have introduced the concept of body and formulated a decomposition that does not rely on ranks. We will add the above discussion in Appendix in the camera-ready version.

Reviewer wHyr2023-08-13

Thank you for your detailed response and the accompanying graphic images. 1. I do not fully agree with your statement `the interaction is more interpretable than the traditional rank`. For example, in Lines 292-295 you repeatedly refer to the intuition of interactions' choice, but such intuition is possible in a significantly limited number of practical problems. 2. With regard to computational complexity, I meant the general form of the expected dependence on the problem's dimension (the order of the tensor in terms of your work) and the number of elements in each mode. For example, for tensor ring and tensor train decompositions, the complexity is usually positioned as linear in these quantities (subject to limited rank). Your approach, as far as I understand, has a significantly higher complexity, however, for the case of the dimensions considered in your work, this is not a problem. 3. I noticed that in section 3.3 in paragraph "Synthetic data" you missed the link to Figure from the application, I advise you to add it to the final text. However, these comments are not fundamental. The approach you suggested seems interesting to the scientific community, and you provided relevant clarifications to my initial comments in the review, so I'm raising my rating from 5 to 6.

Authorsrebuttal2023-08-17

We appreciate your careful reading of our response and acknowledging our contributions. We will revise our text according to your third suggestion in the camera-ready version.

Reviewer gp7Q2023-08-17

Thanks for the comments and clarifications. I was not aware of the results regarding convexity of low rank tensor product approximation with use of KL divergence. Thanks also for pointing out the subsections where some things are clarified, I had overlooked a coupled of things in my prior pass. I have updated my review to a weak accept as I now see why correctness of convexity should follow. I refrained from a full accept, as I am concerned that, even if convexity follows from KL divergence, why does that make the proposed method better than alternative tensor decompositions with a KL divergence objective? As far as I can tell this is not addressed in the experimental comparison (only tensor ring is considered, but based on prior work at least CP or tensor train with KL divergence seem like plausible alternatives). The paper should also describe the prior results regarding convexity of low-rank tensor decomposition with KL divergence as an objective and survey for what decompositions this is applied / is possible. This point seems to be a core motivation of the whole approach/analysis.

Authorsrebuttal2023-08-18

We appreciate your careful reading of our response and acknowledging our contributions. > Why does that make the proposed method better than alternative tensor decompositions with a KL divergence objective? We would like to clarify that low-rank tensor decompositions are non-convex except for the rank-1 case, even if the objective function is defined with the KL divergence. In contrast, our proposal is always convex, which makes our proposal more favorable than theirs in terms of the stability. > This is not addressed in the experimental comparison (only tensor ring is considered, but based on prior work at least CP or tensor train with KL divergence seem like plausible alternatives). In Figure 11 in Appendix, we have already provided comparisons of the proposal with non-negative Tucker decomposition (NTD) and non-negative tensor train decomposition (NTTF) when optimized with the Frobenius norm. We here further provide fit scores of NTD and Non-negative CP decomposition (CPAPR) using the same protocol in Figure 11(a), where these objective functions are defined with not the Frobenius norm but the KL-divergence from input tensor to the reconstructed tensor. #### NTD with the KL divergence | Rank | # Parameters | Fit score (worst) | Fit score (best) | | ---- | ---- | ---- | ---- | | [1,1,1,1] | 69 | 0.835 | 0.835 | | [5,5,5,5] | 965 | 0.835 | 0.894 | | [6,6,6,6] | 1704 | 0.835 | 0.901 | | [11,11,11,11] | 15389| 0.835 | 0.916| #### Non-negative CP decomposition (CPAPR) with the KL divergence | Rank | # Parameters | Fit score (worst) | Fit score (best) | | ---- | ---- | ---- | ---- | | 1 | 69 | 0.835 | 0.835 | | 3 | 204 | 0.835 | 0.877 | | 10 | 680 | 0.888 | 0.910 | (For larger ranks, the algorithm did not converge.) #### Many-body approximation | Body | # Parameters | Fit score | | ---- | ---- | ---- | | One | 64 | 0.835 | | Cyclic-two | 1052| 0.907 | | Two | 1418| 0.917 | | Three | 11762| 0.973 | Hyperparameters for CPAPR are set to default values of the TensorLy implementation, and those for NTD are described in Section E.2. Please note that NTD and CPAPR are convex only if their ranks are [1,1,1,1] and 1, respectively, and hence the worst and the best scores are the same in this case. > The paper should also describe the prior results regarding convexity of low-rank tensor decomposition with KL divergence as an objective and survey for what decompositions this is applied / is possible. The low-rank tensor decomposition with the KL divergence is non-convex except for the rank-1 case. We will add this description in the camera-ready version.

Reviewer gp7Q2023-08-18

In the KL divergence is not NP hard reference (10.1109/ACSSC.2017.8335432), Section IV introduces a convex objective for CP decomposition based on KL divergence. They do make use of inequalities to get it, so I still believe your statement regarding nonconvexity for other decompositions, but more mathematical justification or references seem needed to make/clarify this claim. Thanks for providing a comparison to CP/Tucker with KL divergence. The manybody approximation method seems to generally obtain reasonable results in experiment. Though in this comparison, its not clear to me what method CPAPR is. CP decomposition is perhaps the most natural choice for nonnegative tensors and there is a good bit of prior work on it focusing on nonnegativity and generalized loss functions. A comparison to library like PLANC would be more interesting than TensorLy. For strictly positive data, another effective method is to just take the logarithm of the data and minimize Frobenius norm relative to that. But I don't view addition of such experimental comparisons as necessary for publication.

Authorsrebuttal2023-08-20

We appreciate your constructive comments and valuable feedback. > more mathematical justification or references seem needed to make/clarify this claim. To support this claim, we would like to list known results about non-convexity with their references in the following. - The objective function with the KL divergence to approximate a given tensor by a CP formatted tensor is non-convex. This is described in the last paragraph in Section 1.1 in [c] and Equation (2.5) in [d] assuming $m_i$ has low-CP-rank structure. - The objective function with the KL divergence to approximate a given tensor by a Tucker formatted tensor is non-convex with respect to all factors. This is described in the first paragraph in Section 3.1 in [e]. The beta-divergence used in the literature is a generalization of the KL divergence. [c] DOI: 10.1080/10556788.2015.1009977 [d] DOI: 10.2172/1865529 [e] Axel Marmoret, Florian Voorwinden, Valentin Leplat, Jérémy E Cohen, Frédéric Bimbot. "Nonnegative Tucker Decomposition with Beta-divergence for Music Structure Analysis of audio signals." GRETSI 2022: XXVIIIe Colloque. 2022. > its not clear to me what method CPAPR is. We apologize for the missing reference of CPAPR. It is CP decomposition that minimizes the KL divergence instead of the Frobenius norm, which was introduced in [f]. Since both CPAPR and many-body approximation assumes non-negativity and objective functions with KL-divergence, CPAPR is a natural choice of a baseline. Please note that, in [f], although Poisson distribution for input data is discussed to motivate the choice of the KL divergence, technically we do not need such a distribution to apply CPAPR. [f] DOI: 10.1137/110859063 > A comparison to library like PLANC would be more interesting than TensorLy. For strictly positive data, another effective method is to just take the logarithm of the data and minimize Frobenius norm relative to that. But I don't view addition of such experimental comparisons as necessary for publication. Thank you for your valuable suggestions about heuristics and the useful library. We will consider them in our future work.

Reviewer oVis2023-08-19

Thanks so much for your clarifications about the issue of theoretical analysis. I shall raise my rating from 6 to 7.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC