Using a clickstream sample of 2 billion URLs from many thousand volunteer Web users, we wish to analyze typical usage of keyword searches across the Web. In order to do this, we need to be able to determine whether a given URL represents a keyword search and, if so, which field contains the query. Although it is easy to recognize 'q' as the query field in 'http://www.google.com/search?hl=en&q=music', we must do this automatically for the long tail of diverse websites. This problem is the focus of this paper. Since the names, types and number of fields differ across sites, this does not conform to traditional text classification or to multi-class problem formulations. The problem also exhibits highly non-uniform importance across websites, since traffic follows a Zipf distribution.

Linear classifiers have been shown to be effective for many discrimination tasks. Irrespective of the learning algorithm itself, the final classifier has a weight to multiply by each feature. This suggests that ideally each input feature should be linearly correlated with the target variable (or anti-correlated), whereas raw features may be highly non-linear. In this paper, we attempt to re-shape each input feature so that it is appropriate to use with a linear weight and to scale the different features in proportion to their predictive value. We demonstrate that this pre-processing is beneficial for linear SVM classifiers on a large benchmark of text classification tasks as well as UCI datasets.

Locality Sensitive Hash functions are invaluable tools for approximate near neighbor problems in high dimensional spaces. In this work, we are focused on LSH schemes where the similarity metric is the cosine measure. The contribution of this work is a new class of locality sensitive hash functions for the cosine similarity measure based on the theory of concomitants, which arises in order statistics. Consider n i.i.d sample pairs, {(X1; Y1); (X2; Y2); : : : ;(Xn; Yn)} obtained from a bivariate distribution f(X, Y). Concomitant theory captures the relation between the order statistics of X and Y in the form of a rank distribution given by Prob(Rank(Yi)=j-Rank(Xi)=k). We exploit properties of the rank distribution towards developing a locality sensitive hash family that has excellent collision rate properties for the cosine measure.

We combine the speed and scalability of information retrieval with the generally superior classification accuracy offered by machine learning, yielding a two-phase text classifier that can scale to very large document corpora. We investigate the effect of different methods of formulating the query from the training set, as well as varying the query size. In empirical tests on the Reuters RCV1 corpus of 806,000 documents, we find runtime was easily reduced by a factor of 27x, with a somewhat surprising gain in F-measure compared with traditional text classification.

As the first decade of the 21st century comes to a close, growth in multimedia delivery infrastructure and public demand for applications built on this backbone are converging like never before. The push towards reaching truly interactive multimedia technologies becomes stronger as our media consumption paradigms continue to change. In this paper, we profile a technology leading the way in this revolution: active learning. Active learning is a strategy that helps alleviate challenges inherent in multimedia information retrieval through user interaction. We show how active learning is ideally suited for the multimedia information retrieval problem by giving an overview of the paradigm and component technologies used with special attention given to the application scenarios in which these technologies are useful. Finally, we give insight into the future of this growing field and how it fits into the larger context of multimedia information retrieval.

Interactively learning from a small sample of unlabeled examples is an enormously challenging task, one that often arises in vision applications. Relevance feedback and more recently active learning are two standard techniques that have received much attention towards solving this interactive learning problem. How to best utilize the user's effort for labeling, however, remains unanswered. It has been shown in the past that labeling a diverse set of points is helpful, however, the notion of diversity has either been dependent on the learner used, or computationally expensive. In this paper, we intend to address these issues in the bipartite ranking setting. First, we introduce a scheme for picking the query set which will be labeled by an oracle so that it will aid us in learning the ranker in as few active learning rounds as possible. Secondly, we propose a fundamentally motivated, information theoretic view of diversity and its use in a fast, non-degenerate active learning-based relevance feedback setting. Finally, we report comparative testing and results in a real-time image retrieval setting.

Several approaches to collaborative filtering have been studied but seldom have studies been reported for large (several millionusers and items) and dynamic (the underlying item set is continually changing) settings. In this paper we describe our approach to collaborative filtering for generating personalized recommendations for users of Google News. We generate recommendations using three approaches: collaborative filtering using MinHash clustering, Probabilistic Latent Semantic Indexing (PLSI), and covisitation counts. We combine recommendations from different algorithms using a linear model. Our approach is content agnostic and consequently domain independent, making it easily adaptable for other applications and languages with minimal effort. This paper will describe our algorithms and system setup in detail, and report results of running the recommendations engine on Google News.

In this paper we present a novel learning based method for restoring and recognizing images of digits that have been blurred using an unknown kernel. The novelty of our work is an iterative loop that alternates between recognition and restoration stages. In the restoration stage we model the image as an undirected graphical model over the image patches with the compatibility functions represented as non-parametric kernel densities. Compatibility functions are initially learned using uniform random samples from the training data. We solve the inference problem by an extended version of the non-parametric belief propagation algorithm in which we introduce the notion of partial messages. We close the loop by using the confidence scores of the recognition to non-uniformly sample from the training set in order to retrain the compatibility functions. We show experimental results on synthetic and license plate images.

In this paper, we present a novel learning based framework for performing super-resolution using multiple images. We model the image as an undirected graphical model over image patches in which the compatibility functions are represented as non-parametric kernel densities which are learnt from training data. The observed images are translation rectified and stitched together onto a high resolution grid and the inference problem reduces to estimating unknown pixels in the grid. We solve the inference problem by using an extended version of the non-parametric belief propagation algorithm. We show experimental results on synthetic digit images and real face images from the ORL face dataset.

In this paper, we investigate the use of audio and visual rather than only audio features for the task of speech separation in acoustically noisy environments. The success of existing independent component analysis (ICA) systems for the separation of a large variety of signals, including speech, is often limited by the ability of this technique to handle noise. In this paper, we introduce a Bayesian model for the mixing process that describes both the bimodality and the time dependency of speech sources. Our experimental results show that the online demixing process presented here outperforms both the ICA and the audio-only Bayesian model at all levels of noise.

In this paper, we develop a novel method for view-based recognition of human action/activity from videos. By observing just a few frames, we can identify the activity that takes place in a video sequence. The basic idea of our method is that activities can be positively identified from a sparsely sampled sequence of a few body poses acquired from videos. In our approach, an activity is represented by a set of pose and velocity vectors for the major body parts (hands, legs, and torso) and stored in a set of multidimensional hash tables. We develop a theoretical foundation that shows that robust recognition of a sequence of body pose vectors can be achieved by a method of indexing and sequencing and it requires only a few pose vectors (i.e., sampled body poses in video frames). We find that the probability of false alarm drops exponentially with the increased number of sampled body poses. So, matching only a few body poses guarantees high probability for correct recognition. Our approach is parallel, i.e., all possible model activities are examined at one indexing operation. In addition, our method is robust to partial occlusion since each body part is indexed separately. We use a sequence-based voting approach to recognize the activity invariant to the activity speed.

A novel method for view-based recognition of human activity is presented. The basic idea of our method is that activities can be positively identified from a sparsely sampled sequence of few body poses acquired from videos. In our approach, an activity is represented by a set of pose and velocity vectors for the major body parts (hands, legs and torso) and stored in a set of multidimensional hash tables. We show that robust recognition of a sequence of body pose vectors can be achieved by a method of indexing and sequencing and it requires only few vectors (i.e. sampled body poses in video frames). We find that the probability of false alarm drops exponentially with the increased number of sampled body poses. We also achieve speed invariant recognition by eliminating the time factor and replacing it with sequence information. Experiments performed with videos having 8 different activities show robust recognition even for different viewing directions.