Semi-supervised support vector machines (S3VMs) are a kind of popular approaches which try to improve learning performance by exploiting unlabeled data. Though S3VMs have been found helpful in many situations, they may degenerate performance and the resultant generalization ability may be even worse than using the labeled data only. In this paper, we try to reduce the chance of performance degeneration of S3VMs. Our basic idea is that, rather than exploiting all unlabeled data, the unlabeled instances should be selected such that only the ones which are very likely to be helpful are exploited, while some highly risky unlabeled instances are avoided. We propose the S3VM- us method by using hierarchical clustering to select the unlabeled instances. Experiments on a broad range of data sets over eighty-eight different settings show that the chance of performance degeneration of S3VM- us is much smaller than that of existing S3VMs.

The new user problem (aka user cold start) is very common in online recommender systems. Active collaborative filtering (active CF) tries to solve this problem by intelligently soliciting user feedback in order to build an initial user profile with minimal costs. Existing methods only query the user for feedback on items, while users can have preferences over items as well as certain item attributes. In this paper, we extend active CF via user feedback on both items and attributes. For example, when making movie recommendations, the system can ask users for not only their favorite movies, but also attributes such as genres, actors, etc. We design a unified active CF framework for incorporating both item and attribute feedback based on the random walk model. We test the active CF algorithm on real-world movie recommendation data sets to demonstrate that appropriately querying for both item and feature feedback can significantly reduce the overall user effort measured in terms of number of queries. We show that we can achieve much better recommendation quality as compared to traditional active CF methods that support only item feedback.

Inspired by “GoogleTM Sets” and Bayesian sets, we consider the problem of retrieving complex objects and relations among them, i.e., ground atoms from a logical concept, given a query consisting of a few atoms from that concept. We formulate this as a within-network relational learning problem using few labels only and describe an algorithm that ranks atoms using a score based on random walks with restart (RWR): the probability that a random surfer hits an atom starting from the query atoms. Specifically, we compute an initial ranking using personalized PageRank. Then, we find paths of atoms that are connected via their arguments, variablize the ground atoms in each path, in order to create features for the query. These features are used to re-personalize the original RWR and to finally compute the set completion, based on Label Propagation. Moreover, we exploit that RWR techniques can naturally be lifted and show that lifted inference for label propagation is possible. We evaluate our algorithm on a realworld relational dataset by finding completions of sets of objects describing the Roman city of Pompeii. We compare to Bayesian sets and show that our approach gives very reasonable set completions.

Since learning in Boltzmann machines is typically quite slow, there is a need to restrict connections within hidden layers. However, theresulting states of hidden units exhibit statistical dependencies. Based on this observation, we propose using l1/l2 regularization upon the activation probabilities of hidden units in restricted Boltzmann machines to capture the local dependencies among hidden units. This regularization not only encourages hidden units of many groups to be inactive given observed data but also makes hidden units within a group compete with each other for modeling observed data. Thus, the l1/l2 regularization on RBMs yields sparsity at both the group and the hidden unit levels. We call RBMs trained with the regularizer sparse group RBMs (SGRBMs). The proposed SGRBMs are appliedto model patches of natural images, handwritten digits and OCR English letters. Then to emphasize that SGRBMs can learn more discriminative features we applied SGRBMs to pretrain deep networks for classification tasks. Furthermore, we illustrate the regularizer can also be applied to deep Boltzmann machines, which lead to sparse group deep Boltzmann machines. When adapted to the MNIST data set, a two-layer sparse group Boltzmann machine achieves an error rate of 0.84%, which is, to our knowledge, the best published result on the permutation-invariant version of the MNIST task.

Possessing a sufficient level of situation awareness is essential for effective decision making in dynamic environments. In video games, this includes being aware to some extent of the intentions of the opponents. Such high-level awareness hinges upon inferences over the lower-level situation awareness provided by the game state. Traditional plan recognizers are completely passive processes that leave all the initiative to the observed agent. In a situation where the opponent's intentions are unclear, the observer is forced to wait until further observations of the opponent's actions are made to disambiguate the pending goal hypotheses. With the plan recognizer we propose, in contrast, the observer would take the initiative and provoke the opponent, with the expectation that his reaction will give cues as to what his true intentions actually are.

In this paper, we model the pair-wise similarities of a setof documents as a weighted network with a single cutoffparameter. Such a network can be thought of an ensemble of unweighted graphs, each consisting of edges withweights greater than the cutoff value. We look at this network ensemble as a complex system with a temperature parameter, and refer to it as a Latent Network. Ourexperiments on a number of datasets from two different domains show that certain properties of latent networks like clustering coefficient, average shortest path,and connected components exhibit patterns that are significantly divergent from randomized networks. We explain that these patterns reflect the network phase transition as well as the existence of a community structure in document collections. Using numerical analysis,we show that we can use the aforementioned networkproperties to predicts the clustering Normalized MutualInformation (NMI) with high correlation (rho > 0.9). Finally we show that our clustering method significantlyoutperforms other baseline methods (NMI > 0.5)

Automatic semantic annotation of structured data enables unsupervised integration of data from heterogeneous sources but is difficult to perform accurately due to the presence of many numeric fields and proper-noun fields that do not allow reference-based approaches and the absence of natural language text that prevents the use of language-based approaches. In addition, several of these semantic types have multiple heterogeneous representations, while sharing syntactic structure with other types. In this work, we propose a new approach to use conditional random fields (CRFs) to perform semantic annotation of structured data that takes advantage of the structure and labels of the tokens for higher accuracy of field labeling, while still allowing the use of exact inference techniques. We compare our approach with a linear-CRF based model that only labels fields and also with a regular-expression based approach.

In constraint solving, a critical bottleneck is the formulation of aneffective constraint model of an input problem. The Conjure system describedin this paper, a substantial step forward over prototype versions of Conjurepreviously reported, makes a valuable contribution to the automation ofconstraint modelling by automatically producing constraint models from theirspecifications in the abstract constraint specification language Essence. Aset of rules is used to refine an abstract specification into a concreteconstraint model. We demonstrate that this set of rules is readily extensibleto increase the space of possible constraint models Conjure can produce. Ourempirical results confirm that Conjure can reproduce successfully the kernelsof the constraint models of 32 benchmark problems found in the literature.

We present a novel learning-based method for generating optimal motion plans for high-dimensional motion planning problems. In order to cope with the curse of dimensional- ity, our method proceeds in a fashion similar to block co- ordinate descent in finite-dimensional optimization: at each iteration, the motion is optimized over a lower dimensional subspace while leaving the path fixed along the other dimen- sions. Naive implementations of such an idea can produce vastly suboptimal results. In this work, we show how a prof- itable set of directions in which to perform this dimensional descent procedure can be learned efficiently. We provide suf- ficient conditions for global optimality of dimensional de- scent in this learned basis, based upon the low-dimensional structure of the planning cost function. We also show how this dimensional descent procedure can easily be used for problems that do not exhibit such structure with monotonic convergence. We illustrate the application of our method to high dimensional shape planning and arm trajectory planning problems.

For the problem of allocating resources among multiple strategic agents, the well-known Vickrey-Clarke-Groves (VCG) mechanism is efficient, strategy-proof, and it never incurs a deficit. However, in general, under the VCG mechanism, payments flow out of the system of agents, which reduces the agents' utilities. VCG redistribution mechanisms aim to return as much of the VCG payments as possible back to the agents, without affecting the desirable properties of the VCG mechanism. Most previous research on VCG redistribution mechanisms has focused on settings with homogeneous items and/or settings with unit-demand agents. In this paper, we study VCG redistribution mechanisms in the more general setting of combinatorial auctions. We show that when the gross substitutes condition holds, we are able to design mechanisms that guarantee to redistribute a large fraction of the VCG payments.