Cepstral features have been widely used in audio applications. Domain knowledge has played an important role in designing different types of cepstral features proposed in the literature. In this paper, we present a novel approach for learning optimized cepstral features directly from audio data to better discriminate between different categories of signals in classification tasks. We employ multi-layer feedforward neural networks to model the cepstral feature extraction process. The network weights are initialized to replicate a reference cepstral feature like the mel frequency cepstral coefficient. We then propose a embedded approach that integrates feature learning with the training of a support vector machine (SVM) classifier. A single optimization problem is formulated where the feature and classifier variables are optimized simultaneously so as to refine the initial features and minimize the classification risk. Experimental results have demonstrated the effectiveness of the proposed feature learning approach, outperforming competing methods by a large margin on benchmark data.

Decision theory focuses on the problem of making decisions under uncertainty. This uncertainty arises from the unknown aspects of the state of the world the decision maker is in or the unknown utility function of performing actions. The uncertainty can be modeled as a probability distribution capturing our belief about the world the decision maker is in. Upon making new observations, the decision maker becomes more confident about this model. In addition, if there is a prior belief on this uncertainty that may have obtained from similar experiments, the Bayesian methods may be employed. The loss incurred by the decision maker can also be utilized for the optimal action selection. Most machine learning algorithms developed though focus on one of these aspects for learning and prediction; either learning the probabilistic model or minimizing the loss. In probabilistic models, approximate inference, the process of obtaining the desired model from the observations when its is not tractable, does not consider the task loss. On the other end of the spectrum, the common practice in learning is to minimize the task loss without considering the uncertainty of prediction model. Therefore, we investigate the intersection of decision theory and machine learning considering both uncertainty in prediction model and the task loss.

Matrix factorization (MF) is a popular collaborative filtering approach for recommender systems due to its simplicity and effectiveness. Existing MF methods either assume that all latent features are uncorrelated or assume that all are correlated. To address the important issue of what structure should be imposed on the features, we investigate the covariance matrix of the latent features learned from real data. Based on the findings, we propose an MF model with a sparse covariance prior which favors a sparse yet non-diagonal covariance matrix. Not only can this reflect the semantics more faithfully, but imposing sparsity can also have a side effect of preventing overfitting. Starting from a probabilistic generative model with a sparse covariance prior, we formulate the model inference problem as a maximum a posteriori (MAP) estimation problem. The optimization procedure makes use of stochastic gradient descent and majorization-minimization. For empirical validation, we conduct experiments using the MovieLens and Netflix datasets to compare the proposed method with two strong baselines which use different priors. Experimental results show that our sparse covariance prior can lead to performance improvement.

In this paper we offer a potential solution to the cold-start problem in group recommender systems. To do so, we use information about previous group recommendation events and copy ratings from a user who played a similar role in some previous group event. We show that copying in this way, i.e. conditioned on groups, is superior to copying nothing and also superior to copying ratings from the most similar user known to the system.

We study Bayesian reinforcement learning (RL) as a solution of the exploration-exploitation dilemma. As full Bayesian planning is intractable except for special cases, previous work has proposed several approximation methods. However, these were often computationally expensive or limited to Dirichlet priors. In this paper, we propose a new algorithm that is fast and of polynomial time for near Bayesian optimal policy with any prior distributions that are not greatly misspecified. Perhaps even more interestingly, the proposed algorithm can naturally avoid being misled by incorrect beliefs, while effectively utilizing useful parts of prior information. It can work well even when an utterly misspecified prior is assigned. In that case, the algorithm will follow PAC-MDP behavior instead, if an existing PAC-MDP algorithm does so. The proposed algorithm naturally outperformed other algorithms compared with it on a standard benchmark problem.

Multi-winner social choice considers the problem of selecting a slate of K options to realize some social objective. It has found application in the construction of political legislatures and committees, product recommendation, and related problems, and has recently attracted attention from a computational perspective. We address the multi-winner problem when facing incomplete voter preferences , using the notion of minimax regret to determine a  robust slate of options in the presence of preference uncertainty. We analyze the complexity of this problem and develop new exact and greedy robust optimization algorithms for its solution. Using these techniques, we also develop preference elicitation heuristics which, in practice, allow us to find near-optimal slates with considerable savings in the preference information required vis-a-vis complete votes.

Our initial line of research has shown that, to achieve the best performance on a constraint satisfaction problem, it may be beneficial to translate it to a satisfiability problem. For this translation, it is important to choose both the encoding and satisfiability solver in combination. By doing so, the contrasting performance among solvers on different representations of the same problem can be exploited. In taking these considerations into account, the performance of a solver portfolio augmented with multiple problem transformations can be improved significantly compared to restricting the portfolio to a single problem representation.

Location plays an essential role in our lives, bridging our online and offline worlds. This paper explores the interplay of people's location, interactions, and social ties within a large real-world dataset. We present and evaluate Flap, a system that solves two intimately related tasks: link and location prediction in online social networks. For link prediction, Flap infers social ties by considering patterns in friendship formation, the content of people's messages, and user location. We show that while each component is a weak predictor of friendship alone, combining them results in a strong model---accurately identifying the majority of friendships. For location prediction, Flap implements a scalable probabilistic model of human mobility, where we treat users with known GPS positions as noisy sensors of the location of their friends. We explore supervised and unsupervised learning scenarios, and focus on the efficiency of both learning and inference. We evaluate Flap on a large sample of highly active users from two distinct geographical areas and show that it (1) reconstructs the entire friendship graph with high accuracy even when no edges are given; and (2) infers people's fine-grained location, even when they keep their data private and we can only access the location of their friends. Our models significantly outperform current approaches to either task.

To tackle the potentially hard task of defining the reward function in a Markov Decision Process, we propose a new approach, based on Value Iteration, which interweaves the elicitation and optimization phases. We assume that rewards whose numeric values are unknown can only be ordered, and that a tutor is present to help comparing sequences of re- wards. We first show how the set of possible reward functions for a given preference relation can be rep- resented as a polytope. Then our algorithm, called Interactive Value Iteration, searches for an optimal policy while refining its knowledge about the pos- sible reward functions, by querying a tutor when necessary. We prove that the number of queries needed before finding an optimal policy is upper- bounded by a polynomial in the size of the problem, and we present experimental results which demon- strate that our approach is efficient in practice.

It is a classic result in database theory that conjunctive query (CQ) answering, which is NP-complete in general, is feasible in polynomial time when restricted to acyclic queries. Subsequent results identified more general structural properties of CQs (like bounded treewidth) which ensure tractable query evaluation. In this paper, we lift these tractability results to knowledge bases formulated in the lightweight description logics DL-Lite and ELH. The proof exploits known properties of query matches in these logics and involves a query-dependent modification of the data. To obtain a more practical approach, we propose a concrete polynomial-time algorithm for answering acyclic CQs based on rewriting queries into datalog programs. A preliminary evaluation suggests the interest of our approach for handling large acyclic CQs.