Matrix completion is the task of filling in the missing entries of a partially observed matrix, which is equivalent to performing data imputation in statistics. A wide range of datasets are naturally organized in matrix form. One example is the movie-ratings matrix, as appears in the Netflix problem: Given a ratings matrix in which each entry represents the rating of movie by customer , if customer has watched movie and is otherwise missing, we would like to predict the remaining entries in order to make good recommendations to customers on what to watch next. Another example is the document-term matrix: The frequencies of words used in a collection of documents can be represented as a matrix, where each entry corresponds to the number of times the associated term appears in the indicated document.

Matrix completion of a partially revealed 5 by 5 matrix with rank-1. Left: observed incomplete matrix; Right: matrix completion result.

Without any restrictions on the number of degrees of freedom in the completed matrix this problem is underdetermined since the hidden entries could be assigned arbitrary values. Thus we require some assumption on the matrix to create a well-posed problem, such as assuming it has maximal determinant, is positive definite, or is low-rank.[1][2]

For example, one may assume the matrix has low-rank structure, and then seek to find the lowest rank matrix or, if the rank of the completed matrix is known, a matrix of rank that matches the known entries. The illustration shows that a partially revealed rank-1 matrix (on the left) can be completed with zero-error (on the right) since all the rows with missing entries should be the same as the third row. In the case of the Netflix problem the ratings matrix is expected to be low-rank since user preferences can often be described by a few factors, such as the movie genre and time of release. Other applications include computer vision, where missing pixels in images need to be reconstructed, detecting the global positioning of sensors in a network from partial distance information, and multiclass learning. The matrix completion problem is in general NP-hard, but under additional assumptions there are efficient algorithms that achieve exact reconstruction with high probability.

In statistical learning point of view, the matrix completion problem is an application of matrix regularization which is a generalization of vector regularization. For example, in the low-rank matrix completion problem one may apply the regularization penalty taking the form of a nuclear norm

Low rank matrix completion

edit

One of the variants of the matrix completion problem is to find the lowest rank matrix   which matches the matrix  , which we wish to recover, for all entries in the set   of observed entries. The mathematical formulation of this problem is as follows:

 

Candès and Recht[3] proved that with assumptions on the sampling of the observed entries and sufficiently many sampled entries this problem has a unique solution with high probability.

An equivalent formulation, given that the matrix   to be recovered is known to be of rank  , is to solve for   where  

Assumptions

edit

A number of assumptions on the sampling of the observed entries and the number of sampled entries are frequently made to simplify the analysis and to ensure the problem is not underdetermined.

Uniform sampling of observed entries

edit

To make the analysis tractable, it is often assumed that the set   of observed entries and fixed cardinality is sampled uniformly at random from the collection of all subsets of entries of cardinality  . To further simplify the analysis, it is instead assumed that   is constructed by Bernoulli sampling, i.e. that each entry is observed with probability  . If   is set to   where   is the desired expected cardinality of  , and   are the dimensions of the matrix (let   without loss of generality),   is within   of   with high probability, thus Bernoulli sampling is a good approximation for uniform sampling.[3] Another simplification is to assume that entries are sampled independently and with replacement.[4]

Lower bound on number of observed entries

edit

Suppose the   by   matrix   (with  ) we are trying to recover has rank  . There is an information theoretic lower bound on how many entries must be observed before   can be uniquely reconstructed. The set of   by   matrices with rank less than or equal to   is an algebraic variety in  with dimension  . Using this result, one can show that at least   entries must be observed for matrix completion in   to have a unique solution when   .[5]

Secondly, there must be at least one observed entry per row and column of  . The singular value decomposition of   is given by  . If row   is unobserved, it is easy to see the   right singular vector of  ,  , can be changed to some arbitrary value and still yield a matrix matching   over the set of observed entries. Similarly, if column   is unobserved, the   left singular vector of  ,   can be arbitrary. If we assume Bernoulli sampling of the set of observed entries, the Coupon collector effect implies that entries on the order of   must be observed to ensure that there is an observation from each row and column with high probability.[6]

Combining the necessary conditions and assuming that   (a valid assumption for many practical applications), the lower bound on the number of observed entries required to prevent the problem of matrix completion from being underdetermined is on the order of  .

Incoherence

edit

The concept of incoherence arose in compressed sensing. It is introduced in the context of matrix completion to ensure the singular vectors of   are not too "sparse" in the sense that all coordinates of each singular vector are of comparable magnitude instead of just a few coordinates having significantly larger magnitudes.[7][8] The standard basis vectors are then undesirable as singular vectors, and the vector   in   is desirable. As an example of what could go wrong if the singular vectors are sufficiently "sparse", consider the   by   matrix   with singular value decomposition  . Almost all the entries of   must be sampled before it can be reconstructed.

Candès and Recht[3] define the coherence of a matrix   with column space an  dimensional subspace of   as  , where   is the orthogonal projection onto  . Incoherence then asserts that given the singular value decomposition   of the   by   matrix  ,

  1.  
  2. The entries of   have magnitudes upper bounded by  

for some  .

Low rank matrix completion with noise

edit

In real world application, one often observe only a few entries corrupted at least by a small amount of noise. For example, in the Netflix problem, the ratings are uncertain. Candès and Plan [9] showed that it is possible to fill in the many missing entries of large low-rank matrices from just a few noisy samples by nuclear norm minimization. The noisy model assumes that we observe

 

where   is a noise term. Note that the noise can be either stochastic or deterministic. Alternatively the model can be expressed as

 

where   is an   matrix with entries   for   assuming that   for some   .To recover the incomplete matrix, we try to solve the following optimization problem:

 

Among all matrices consistent with the data, find the one with minimum nuclear norm. Candès and Plan [9] have shown that this reconstruction is accurate. They have proved that when perfect noiseless recovery occurs, then matrix completion is stable vis a vis perturbations. The error is proportional to the noise level  . Therefore, when the noise level is small, the error is small. Here the matrix completion problem does not obey the restricted isometry property (RIP). For matrices, the RIP would assume that the sampling operator obeys

 

for all matrices   with sufficiently small rank and   sufficiently small. The methods are also applicable to sparse signal recovery problems in which the RIP does not hold.

High rank matrix completion

edit

The high rank matrix completion in general is NP-Hard. However, with certain assumptions, some incomplete high rank matrix or even full rank matrix can be completed.

Eriksson, Balzano and Nowak [10] have considered the problem of completing a matrix with the assumption that the columns of the matrix belong to a union of multiple low-rank subspaces. Since the columns belong to a union of subspaces, the problem may be viewed as a missing-data version of the subspace clustering problem. Let   be an   matrix whose (complete) columns lie in a union of at most   subspaces, each of  , and assume  . Eriksson, Balzano and Nowak [10] showed that under mild assumptions each column of   can be perfectly recovered with high probability from an incomplete version so long as at least   entries of   are observed uniformly at random, with   a constant depending on the usual incoherence conditions, the geometrical arrangement of subspaces, and the distribution of columns over the subspaces.

The algorithm involves several steps: (1) local neighborhoods; (2) local subspaces; (3) subspace refinement; (4) full matrix completion. This method can be applied to Internet distance matrix completion and topology identification.

Algorithms for Low-Rank Matrix Completion

edit

Various matrix completion algorithms have been proposed.[8] These include convex relaxation-based algorithm,[3] gradient-based algorithm,[11] and alternating minimization-based algorithm.[12]

Convex relaxation

edit

The rank minimization problem is NP-hard. One approach, proposed by Candès and Recht, is to form a convex relaxation of the problem and minimize the nuclear norm   (which gives the sum of the singular values of  ) instead of   (which counts the number of non zero singular values of  ).[3] This is analogous to minimizing the L1-norm rather than the L0-norm for vectors. The convex relaxation can be solved using semidefinite programming (SDP) by noticing that the optimization problem is equivalent to

 

The complexity of using SDP to solve the convex relaxation is  . State of the art solvers like SDPT3 can only handle matrices of size up to 100 by 100 [13] An alternative first order method that approximately solves the convex relaxation is the Singular Value Thresholding Algorithm introduced by Cai, Candès and Shen.[13]

Candès and Recht show, using the study of random variables on Banach spaces, that if the number of observed entries is on the order of   (assume without loss of generality  ), the rank minimization problem has a unique solution which also happens to be the solution of its convex relaxation with probability   for some constant  . If the rank of   is small ( ), the size of the set of observations reduces to the order of  . These results are near optimal, since the minimum number of entries that must be observed for the matrix completion problem to not be underdetermined is on the order of  .

This result has been improved by Candès and Tao.[6] They achieve bounds that differ from the optimal bounds only by polylogarithmic factors by strengthening the assumptions. Instead of the incoherence property, they assume the strong incoherence property with parameter  . This property states that:

  1.   for   and   for  
  2. The entries of   are bounded in magnitude by  

Intuitively, strong incoherence of a matrix   asserts that the orthogonal projections of standard basis vectors to   has magnitudes that have high likelihood if the singular vectors were distributed randomly.[7]

Candès and Tao find that when   is   and the number of observed entries is on the order of  , the rank minimization problem has a unique solution which also happens to be the solution of its convex relaxation with probability   for some constant  . For arbitrary  , the number of observed entries sufficient for this assertion hold true is on the order of  

Another convex relaxation approach [14] is to minimize the Frobenius squared norm under a rank constraint. This is equivalent to solving

 

By introducing an orthogonal projection matrix   (meaning  ) to model the rank of   via   and taking this problem's convex relaxation, we obtain the following semidefinite program

 

If Y is a projection matrix (i.e., has binary eigenvalues) in this relaxation, then the relaxation is tight. Otherwise, it gives a valid lower bound on the overall objective. Moreover, it can be converted into a feasible solution with a (slightly) larger objective by rounding the eigenvalues of Y greedily.[14] Remarkably, this convex relaxation can be solved by alternating minimization on X and Y without solving any SDPs, and thus it scales beyond the typical numerical limits of state-of-the-art SDP solvers like SDPT3 or Mosek.

This approach is a special case of a more general reformulation technique, which can be applied to obtain a valid lower bound on any low-rank problem with a trace-matrix-convex objective.[15]


Gradient descent

edit

Keshavan, Montanari and Oh[11] consider a variant of matrix completion where the rank of the   by   matrix  , which is to be recovered, is known to be  . They assume Bernoulli sampling of entries, constant aspect ratio  , bounded magnitude of entries of   (let the upper bound be  ), and constant condition number   (where   and   are the largest and smallest singular values of   respectively). Further, they assume the two incoherence conditions are satisfied with   and   where   and   are constants. Let   be a matrix that matches   on the set   of observed entries and is 0 elsewhere. They then propose the following algorithm:

  1. Trim   by removing all observations from columns with degree larger than   by setting the entries in the columns to 0. Similarly remove all observations from rows with degree larger than  .
  2. Project   onto its first   principal components. Call the resulting matrix  .
  3. Solve   where   is some regularization function by gradient descent with line search. Initialize   at   where  . Set   as some function forcing   to remain incoherent throughout gradient descent if   and   are incoherent.
  4. Return the matrix  .

Steps 1 and 2 of the algorithm yield a matrix   very close to the true matrix   (as measured by the root mean square error (RMSE)) with high probability. In particular, with probability  ,   for some constant  .   denotes the Frobenius norm. Note that the full suite of assumptions is not needed for this result to hold. The incoherence condition, for example, only comes into play in exact reconstruction. Finally, although trimming may seem counter intuitive as it involves throwing out information, it ensures projecting   onto its first   principal components gives more information about the underlying matrix   than about the observed entries.

In Step 3, the space of candidate matrices   can be reduced by noticing that the inner minimization problem has the same solution for   as for   where   and   are orthonormal   by   matrices. Then gradient descent can be performed over the cross product of two Grassman manifolds. If   and the observed entry set is in the order of  , the matrix returned by Step 3 is exactly  . Then the algorithm is order optimal, since we know that for the matrix completion problem to not be underdetermined the number of entries must be in the order of  .

Alternating least squares minimization

edit

Alternating minimization represents a widely applicable and empirically successful approach for finding low-rank matrices that best fit the given data. For example, for the problem of low-rank matrix completion, this method is believed to be one of the most accurate and efficient, and formed a major component of the winning entry in the Netflix problem. In the alternating minimization approach, the low-rank target matrix is written in a bilinear form:

 ;

the algorithm then alternates between finding the best   and the best  . While the overall problem is non-convex, each sub-problem is typically convex and can be solved efficiently. Jain, Netrapalli and Sanghavi [12] have given one of the first guarantees for performance of alternating minimization for both matrix completion and matrix sensing.

The alternating minimization algorithm can be viewed as an approximate way to solve the following non-convex problem:

 

The AltMinComplete Algorithm proposed by Jain, Netrapalli and Sanghavi is listed here:[12]

  1. Input: observed set  , values  
  2. Partition   into   subsets   with each element of   belonging to one of the   with equal probability (sampling with replacement)
  3.   i.e., top-  left singular vectors of  
  4. Clipping: Set all elements of   that have magnitude greater than   to zero and orthonormalize the columns of  
  5. for   do
  6.  
  7.  
  8. end for
  9. Return  

They showed that by observing   random entries of an incoherent matrix  , AltMinComplete algorithm can recover   in   steps. In terms of sample complexity ( ), theoretically, Alternating Minimization may require a bigger   than Convex Relaxation. However empirically it seems not the case which implies that the sample complexity bounds can be further tightened. In terms of time complexity, they showed that AltMinComplete needs time

 .

It is worth noting that, although convex relaxation based methods have rigorous analysis, alternating minimization based algorithms are more successful in practice.[citation needed]

Applications

edit

Several applications of matrix completion are summarized by Candès and Plan[9] as follows:

Collaborative filtering

edit

Collaborative filtering is the task of making automatic predictions about the interests of a user by collecting taste information from many users. Companies like Apple, Amazon, Barnes and Noble, and Netflix are trying to predict their user preferences from partial knowledge. In these kind of matrix completion problem, the unknown full matrix is often considered low rank because only a few factors typically contribute to an individual's tastes or preference.

System identification

edit

In control, one would like to fit a discrete-time linear time-invariant state-space model

 

to a sequence of inputs   and outputs  . The vector   is the state of the system at time   and   is the order of the system model. From the input/output pair, one would like to recover the matrices   and the initial state  . This problem can also be viewed as a low-rank matrix completion problem.

Internet of things (IoT) localization

edit

The localization (or global positioning) problem emerges naturally in IoT sensor networks. The problem is to recover the sensor map in Euclidean space from a local or partial set of pairwise distances. Thus it is a matrix completion problem with rank two if the sensors are located in a 2-D plane and three if they are in a 3-D space.[16]

Social Networks Recovery

edit

Most of the real-world social networks have low-rank distance matrices. When we are not able to measure the complete network, which can be due to reasons such as private nodes, limited storage or compute resources, we only have a fraction of distance entries known. Criminal networks are a good example of such networks. Low-rank Matrix Completion can be used to recover these unobserved distances.[17]

See also

edit

References

edit
  1. ^ Johnson, Charles R. (1990). "Matrix completion problems: A survey". Matrix Theory and Applications. Proceedings of Symposia in Applied Mathematics. Vol. 40. pp. 171–198. doi:10.1090/psapm/040/1059486. ISBN 9780821801543.
  2. ^ Laurent, Monique (2008). "Matrix Completion Problems". Encyclopedia of Optimization. Vol. 3. pp. 221–229. doi:10.1007/978-0-387-74759-0_355. ISBN 978-0-387-74758-3.
  3. ^ a b c d e Candès, E. J.; Recht, B. (2009). "Exact Matrix Completion via Convex Optimization". Foundations of Computational Mathematics. 9 (6): 717–772. arXiv:0805.4471. doi:10.1007/s10208-009-9045-5.
  4. ^ Recht, B. (2009). "A Simpler Approach to Matrix Completion" (PDF). Journal of Machine Learning Research. 12: 3413–3430. arXiv:0910.0651. Bibcode:2009arXiv0910.0651R.
  5. ^ Xu, Zhiqiang (2018). "The minimal measurement number for low-rank matrix recovery". Applied and Computational Harmonic Analysis. 44 (2): 497–508. arXiv:1505.07204. doi:10.1016/j.acha.2017.01.005. S2CID 11990443.
  6. ^ a b Candès, E. J.; Tao, T. (2010). "The Power of Convex Relaxation: Near-Optimal Matrix Completion". IEEE Transactions on Information Theory. 56 (5): 2053–2080. arXiv:0903.1476. doi:10.1109/TIT.2010.2044061. S2CID 1255437.
  7. ^ a b Tao, T. (10 March 2009). "The power of convex relaxation: near-optimal matrix completion". What's new.
  8. ^ a b Nguyen, L.T.; Kim, J.; Shim, B. (10 July 2019). "Low-Rank Matrix Completion: A Contemporary Survey". IEEE Access. 7 (1): 94215–94237. arXiv:1907.11705. Bibcode:2019arXiv190711705N. doi:10.1109/ACCESS.2019.2928130. S2CID 198930899.
  9. ^ a b c Candès, E. J.; Plan, Y. (2010). "Matrix Completion with Noise". Proceedings of the IEEE. 98 (6): 925–936. arXiv:0903.3131. doi:10.1109/JPROC.2009.2035722. S2CID 109721.
  10. ^ a b Eriksson, B.; Balzano, L.; Nowak, R. (2011). "High-Rank Matrix Completion and Subspace Clustering with Missing Data". arXiv:1112.5629 [cs.IT].
  11. ^ a b Keshavan, R. H.; Montanari, A.; Oh, S. (2010). "Matrix Completion from a Few Entries". IEEE Transactions on Information Theory. 56 (6): 2980–2998. arXiv:0901.3150. doi:10.1109/TIT.2010.2046205. S2CID 53504.
  12. ^ a b c Jain, P.; Netrapalli, P.; Sanghavi, S. (2013). "Low-rank Matrix Completion using Alternating Minimization". Proceedings of the 45th annual ACM symposium on Symposium on theory of computing. ACM. pp. 665–674. arXiv:1212.0467. doi:10.1145/2488608.2488693. ISBN 978-1-4503-2029-0. S2CID 447011.
  13. ^ a b Cai, J.-F.; Candès, E. J.; Shen, Z. (2010). "A Singular Value Thresholding Algorithm for Matrix Completion". SIAM Journal on Optimization. 20 (4): 1956–1982. arXiv:0810.3286. doi:10.1137/080738970. S2CID 1254778.
  14. ^ a b Bertsimas, Dimitris; Cory-Wright, Ryan; Pauphilet, Jean (2021). "Mixed-Projection Conic Optimization: A New Paradigm for Modeling Rank Constraints". Operations Research. 70 (6): 3321–3344. arXiv:2009.10395. doi:10.1287/opre.2021.2182. S2CID 221836263.
  15. ^ Bertsimas, Dimitris; Cory-Wright, Ryan; Pauphilet, Jean (2021). "A New Perspective on Low-Rank Optimization". Optimization Online. arXiv:2105.05947.
  16. ^ Nguyen, L.T.; Kim, J.; Kim, S.; Shim, B. (2019). "Localization of IoT Networks Via Low-Rank Matrix Completion". IEEE Transactions on Communications. 67 (8): 5833–5847. doi:10.1109/TCOMM.2019.2915226. S2CID 164605437.
  17. ^ Mahindre, G.; Jayasumana, A.P.; Gajamannage, K.; Paffenroth, R. (2019). "On Sampling and Recovery of Topology of Directed Social Networks – A Low-Rank Matrix Completion Based Approach". 2019 IEEE 44th Conference on Local Computer Networks (LCN). IEEE. pp. 324–331. doi:10.1109/LCN44214.2019.8990707. ISBN 978-1-7281-1028-8. S2CID 211206354.