ICML 2015 Papers

We investigate the statistical convergence rate of a Bayesian low-rank tensor estimator, and derive the minimax optimal rate for learning a low-rank tensor. Our problem setting is the regression problem where the regression coefficient forms a tensor structure. This problem setting occurs in many practical applications, such as collaborative filtering, multi-task learning, and spatio-temporal data analysis. The convergence rate of the Bayes tensor estimator is analyzed in terms of both in-sample and out-of-sample predictive accuracies. It is shown that a fast learning rate is achieved without any strong convexity of the observation. Moreover, we show that the method has adaptivity to the unknown rank of the true tensor, that is, the near optimal rate depending on the true rank is achieved even if it is not known a priori. Finally, we show the minimax optimal learning rate for the tensor estimation problem, and thus show that the derived bound of the Bayes estimator is tight and actually near minimax optimal.

Parameter inference in mechanistic models based on non-affine differential equations is computationally onerous, and various faster alternatives based on gradient matching have been proposed. A particularly promising approach is based on nonparametric Bayesian modelling with Gaussian processes, which exploits the fact that a Gaussian process is closed under differentiation. However, two alternative paradigms have been proposed. The first paradigm, proposed at NIPS 2008 and AISTATS 2013, is based on a product of experts approach and a marginalization over the derivatives of the state variables. The second paradigm, proposed at ICML 2014, is based on a probabilistic generative model and a marginalization over the state variables. The claim has been made that this leads to better inference results. In the present article, we offer a new interpretation of the second paradigm, which highlights the underlying assumptions, approximations and limitations. In particular, we show that the second paradigm suffers from an intrinsic identifiability problem, which the first paradigm is not affected by.

Big Data received from sources such as social media, in-stream monitoring systems, networks, and markets is often mined for discovering patterns, detecting anomalies, and making decisions or predictions. In distributed learning and real-time processing of Big Data, ensemble-based systems in which a fusion center (FC) is used to combine the local decisions of several classifiers, have shown to be superior to single expert systems. However, optimal design of the FC requires knowledge of the accuracy of the individual classifiers which, in many cases, is not available. Moreover, in many applications supervised training of the FC is not feasible since the true labels of the data set are not available. In this paper, we propose an unsupervised joint estimation-detection scheme to estimate the accuracies of the local classifiers as functions of data context and to fuse the local decisions of the classifiers. Numerical results show the dramatic improvement of the proposed method as compared with the state of the art approaches.

Significant progress has been made recently on theoretical analysis of estimators for the stochastic block model (SBM). In this paper, we consider the multi-graph SBM, which serves as a foundation for many application settings including dynamic and multi-layer networks. We explore the asymptotic properties of two estimators for the multi-graph SBM, namely spectral clustering and the maximum-likelihood estimate (MLE), as the number of layers of the multi-graph increases. We derive sufficient conditions for consistency of both estimators and propose a variational approximation to the MLE that is computationally feasible for large networks. We verify the sufficient conditions via simulation and demonstrate that they are practical. In addition, we apply the model to two real data sets: a dynamic social network and a multi-layer social network with several types of relations.

This paper presents new consistent algorithms for multiclass learning with complex performance measures, defined by arbitrary functions of the confusion matrix. This setting includes as a special case all loss-based performance measures, which are simply linear functions of the confusion matrix, but also includes more complex performance measures such as the multiclass G-mean and micro F_1 measures. We give a general framework for designing consistent algorithms for such performance measures by viewing the learning problem as an optimization problem over the set of feasible confusion matrices, and give two specific instantiations based on the Frank-Wolfe method for concave performance measures and on the bisection method for ratio-of-linear performance measures. The resulting algorithms are provably consistent and outperform a multiclass version of the state-of-the-art SVMperf method in experiments; for large multiclass problems, the algorithms are also orders of magnitude faster than SVMperf.

As deep nets are increasingly used in applications suited for mobile devices, a fundamental dilemma becomes apparent: the trend in deep learning is to grow models to absorb ever-increasing data set sizes; however mobile devices are designed with very little memory and cannot store such large models. We present a novel network architecture, HashedNets, that exploits inherent redundancy in neural networks to achieve drastic reductions in model sizes. HashedNets uses a low-cost hash function to randomly group connection weights into hash buckets, and all connections within the same hash bucket share a single parameter value. These parameters are tuned to adjust to the HashedNets weight sharing architecture with standard backprop during training. Our hashing procedure introduces no additional memory overhead, and we demonstrate on several benchmark data sets that HashedNets shrink the storage requirements of neural networks substantially while mostly preserving generalization performance.

We aim to detect complex events in long Internet videos that may last for hours. A major challenge in this setting is that only a few shots in a long video are relevant to the event of interest while others are irrelevant or even misleading. Instead of indifferently pooling the shots, we first define a novel notion of semantic saliency that assesses the relevance of each shot with the event of interest. We then prioritize the shots according to their saliency scores since shots that are semantically more salient are expected to contribute more to the final event detector. Next, we propose a new isotonic regularizer that is able to exploit the semantic ordering information. The resulting nearly-isotonic SVM classifier exhibits higher discriminative power. Computationally, we develop an efficient implementation using the proximal gradient algorithm, and we prove new, closed-form proximal steps. We conduct extensive experiments on three real-world video datasets and confirm the effectiveness of the proposed approach.

We consider the problem of recovering a complete (i.e., square and invertible) dictionary mb A_0, from mb Y = mb A_0 mb X_0 with mb Y ∈\mathbb R^n \times p. This recovery setting is central to the theoretical understanding of dictionary learning. We give the first efficient algorithm that provably recovers mb A_0 when mb X_0 has O(n) nonzeros per column, under suitable probability model for mb X_0. Prior results provide recovery guarantees when mb X_0 has only O(\sqrtn) nonzeros per column. Our algorithm is based on nonconvex optimization with a spherical constraint, and hence is naturally phrased in the language of manifold optimization. Our proofs give a geometric characterization of the high-dimensional objective landscape, which shows that with high probability there are no spurious local minima. Experiments with synthetic data corroborate our theory. Full version of this paper is available online: \urlhttp://arxiv.org/abs/1504.06785.

Local graph diffusions have proven to be valuable tools for solving various graph clustering problems. As such, there has been much interest recently in efficient local algorithms for computing them. We present an efficient local algorithm for approximating a graph diffusion that generalizes both the celebrated personalized PageRank and its recent competitor/companion - the heat kernel. Our algorithm is based on writing the diffusion vector as the solution of an initial value problem, and then using a waveform relaxation approach to approximate the solution. Our experimental results suggest that it produces rankings that are distinct and competitive with the ones produced by high quality implementations of personalized PageRank and localized heat kernel, and that our algorithm is a useful addition to the toolset of localized graph diffusions.

We consider classification and regression tasks where we have missing data and assume that the (clean) data resides in a low rank subspace. Finding a hidden subspace is known to be computationally hard. Nevertheless, using a non-proper formulation we give an efficient agnostic algorithm that classifies as good as the best linear classifier coupled with the best low-dimensional subspace in which the data resides. A direct implication is that our algorithm can linearly (and non-linearly through kernels) classify provably as well as the best classifier that has access to the full data.

We consider stochastic sequential learning problems where the learner can observe the average reward of several actions. Such a setting is interesting in many applications involving monitoring and surveillance, where the set of the actions to observe represent some (geographical) area. The importance of this setting is that in these applications, it is actually cheaper to observe average reward of a group of actions rather than the reward of a single action. We show that when the reward is smooth over a given graph representing the neighboring actions, we can maximize the cumulative reward of learning while minimizing the sensing cost. In this paper we propose CheapUCB, an algorithm that matches the regret guarantees of the known algorithms for this setting and at the same time guarantees a linear cost again over them. As a by-product of our analysis, we establish a Ω(\sqrt(dT)) lower bound on the cumulative regret of spectral bandits for a class of graphs with effective dimension d.

We present a new, fully generative model of optical telescope image sets, along with a variational procedure for inference. Each pixel intensity is treated as a Poisson random variable, with a rate parameter dependent on latent properties of stars and galaxies. Key latent properties are themselves random, with scientific prior distributions constructed from large ancillary data sets. We check our approach on synthetic images. We also run it on images from a major sky survey, where it exceeds the performance of the current state-of-the-art method for locating celestial bodies and measuring their colors.

A widely applied approach to causal inference from a time series X, often referred to as “(linear) Granger causal analysis”, is to simply regress present on past and interpret the regression matrix \hatB causally. However, if there is an unmeasured time series Z that influences X, then this approach can lead to wrong causal conclusions, i.e., distinct from those one would draw if one had additional information such as Z. In this paper we take a different approach: We assume that X together with some hidden Z forms a first order vector autoregressive (VAR) process with transition matrix A, and argue why it is more valid to interpret A causally instead of \hatB. Then we examine under which conditions the most important parts of A are identifiable or almost identifiable from only X. Essentially, sufficient conditions are (1) non-Gaussian, independent noise or (2) no influence from X to Z. We present two estimation algorithms that are tailored towards conditions (1) and (2), respectively, and evaluate them on synthetic and real-world data. We discuss how to check the model using X.

A search engine usually outputs a list of K web pages. The user examines this list, from the first web page to the last, and chooses the first attractive page. This model of user behavior is known as the cascade model. In this paper, we propose cascading bandits, a learning variant of the cascade model where the objective is to identify K most attractive items. We formulate our problem as a stochastic combinatorial partial monitoring problem. We propose two algorithms for solving it, CascadeUCB1 and CascadeKL-UCB. We also prove gap-dependent upper bounds on the regret of these algorithms and derive a lower bound on the regret in cascading bandits. The lower bound matches the upper bound of CascadeKL-UCB up to a logarithmic factor. We experiment with our algorithms on several problems. The algorithms perform surprisingly well even when our modeling assumptions are violated.

CUR matrix decomposition computes the low rank approximation of a given matrix by using the actual rows and columns of the matrix. It has been a very useful tool for handling large matrices. One limitation with the existing algorithms for CUR matrix decomposition is that they cannot deal with entries in a \it partially observed matrix, while incomplete matrices are found in many real world applications. In this work, we alleviate this limitation by developing a CUR decomposition algorithm for partially observed matrices. In particular, the proposed algorithm computes the low rank approximation of the target matrix based on (i) the randomly sampled rows and columns, and (ii) a subset of observed entries that are randomly sampled from the matrix. Our analysis shows the relative error bound, measured by spectral norm, for the proposed algorithm when the target matrix is of full rank. We also show that only O(n r\ln r) observed entries are needed by the proposed algorithm to perfectly recover a rank r matrix of size n\times n, which improves the sample complexity of the existing algorithms for matrix completion. Empirical studies on both synthetic and real-world datasets verify our theoretical claims and demonstrate the effectiveness of the proposed algorithm.

In marketing planning, advertisers seek to maximize the number of customers by allocating given budgets to each media channel effectively. The budget allocation problem with a bipartite influence model captures this scenario; however, the model is problematic because it assumes there is only one advertiser in the market. In reality, there are many advertisers which are in conflict of advertisement; thus we must extend the model for such a case. By extending the budget allocation problem with a bipartite influence model, we propose a game-theoretic model problem that considers many advertisers. By simulating our model, we can analyze the behavior of a media channel market, e.g., we can estimate which media channels are allocated by an advertiser, and which customers are influenced by an advertiser. Our model has many attractive features. First, our model is a potential game; therefore, it has a pure Nash equilibrium. Second, any Nash equilibrium of our game has 2-optimal social utility, i.e., the price of anarchy is 2. Finally, the proposed model can be simulated very efficiently; thus it can be used to analyze large markets.

Splicing refers to the elimination of non-coding regions in transcribed pre-messenger ribonucleic acid (RNA). Discovering splice sites is an important machine learning task that helps us not only to identify the basic units of genetic heredity but also to understand how different proteins are produced. Existing methods for splicing prediction have produced promising results, but often show limited robustness and accuracy. In this paper, we propose a deep belief network-based methodology for computational splice junction prediction. Our proposal includes a novel method for training restricted Boltzmann machines for class-imbalanced prediction. The proposed method addresses the limitations of conventional contrastive divergence and provides regularization for datasets that have categorical features. We tested our approach using public human genome datasets and obtained significantly improved accuracy and reduced runtime compared to state-of-the-art alternatives. The proposed approach was less sensitive to the length of input sequences and more robust for handling false splicing signals. Furthermore, we could discover non-canonical splicing patterns that were otherwise difficult to recognize using conventional methods. Given the efficiency and robustness of our methodology, we anticipate that it can be extended to the discovery of primary structural patterns of other subtle genomic elements.

By reducing optimization to a sequence of small subproblems, working set methods achieve fast convergence times for many challenging problems. Despite excellent performance, theoretical understanding of working sets is limited, and implementations often resort to heuristics to determine subproblem size, makeup, and stopping criteria. We propose Blitz, a fast working set algorithm accompanied by useful guarantees. Making no assumptions on data, our theory relates subproblem size to progress toward convergence. This result motivates methods for optimizing algorithmic parameters and discarding irrelevant variables as iterations progress. Applied to L1-regularized learning, Blitz convincingly outperforms existing solvers in sequential, limited-memory, and distributed settings. Blitz is not specific to L1-regularized learning, making the algorithm relevant to many applications involving sparsity or constraints.

This paper addresses the problem of predicting the missing edges of a bipartite graph where each side of the vertices has its own intrinsic structure. We propose a new optimization framework to map the two sides of the intrinsic structures onto the manifold structure of the edges via a graph product, and to reduce the original problem to vertex label propagation over the product graph. This framework enjoys flexible choices in the formulation of graph products, and supports a rich family of graph transduction schemes with scalable inference. Experiments on benchmark datasets for collaborative filtering, citation network analysis and prerequisite prediction of online courses show advantageous performance of the proposed approach over other state-of-the-art methods.