Eerika Savia

Combining eye movements and collaborative filtering for proactive information retrieval

We study a new task, proactive information retrieval by combining implicit relevance feedback and collaborative filtering. We have constructed a controlled experimental setting, a prototype application, in which the users try to find interesting scientific articles by browsing their titles. Implicit feedback is inferred from eye movement signals, with discriminative hidden Markov models estimated from existing data in which explicit relevance feedback is available. Collaborative filtering is carried out using the User Rating Profile model, a state-of-the-art probabilistic latent variable model, computed using Markov Chain Monte Carlo techniques. For new document titles the prediction accuracy with eye movements, collaborative filtering, and their combination was significantly better than by chance. The best prediction accuracy still leaves room for improvement but shows that proactive information retrieval and combination of many sources of relevance feedback is feasible.

Dependencies between stimuli and spatially independent fMRI sources: towards brain correlates of natural stimuli.

Natural stimuli are increasingly used in functional magnetic resonance imaging (fMRI) studies to imitate real-life situations. Consequently, challenges are created for novel analysis methods, including new machine-learning tools. With natural stimuli it is no longer feasible to assume single features of the experimental design alone to account for the brain activity. Instead, relevant combinations of rich enough stimulus features could explain the more complex activation patterns. We propose a novel two-step approach, where independent component analysis is first used to identify spatially independent brain processes, which we refer to as functional patterns. As the second step, temporal dependencies between stimuli and functional patterns are detected using canonical correlation analysis. Our proposed method looks for combinations of stimulus features and the corresponding combinations of functional patterns. This two-step approach was used to analyze measurements from an fMRI study during multi-modal stimulation. The detected complex activation patterns were explained as resulting from interactions of multiple brain processes. Our approach seems promising for analysis of data from studies with natural stimuli.

Latent grouping models for user preference prediction

We tackle the problem of new users or documents in collaborative filtering. Generalization over users by grouping them into user groups is beneficial when a rating is to be predicted for a relatively new document having only few observed ratings. Analogously, generalization over documents improves predictions in the case of new users. We show that if either users and documents or both are new, two-way generalization becomes necessary. We demonstrate the benefits of grouping of users, grouping of documents, and two-way grouping, with artificial data and in two case studies with real data. We have introduced a probabilistic latent grouping model for predicting the relevance of a document to a user. The model assumes a latent group structure for both users and items. We compare the model against a state-of-the-art method, the User Rating Profile model, where only the users have a latent group structure. We compute the posterior of both models by Gibbs sampling. The Two-Way Model predicts relevance more accurately when the target consists of both new documents and new users. The reason is that generalization over documents becomes beneficial for new documents and at the same time generalization over users is needed for new users.

EXPLORATORY MODELING OF YEAST STRESS RESPONSE AND ITS REGULATION WITH gCCA AND ASSOCIATIVE CLUSTERING

We model dependencies between m multivariate continuous-valued information sources by a combination of (i) a generalized canonical correlations analysis (gCCA) to reduce dimensionality while preserving dependencies in m - 1 of them, and (ii) summarizing dependencies with the remaining one by associative clustering. This new combination of methods avoids multiway associative clustering which would require a multiway contingency table and hence suffer from curse of dimensionality of the table. The method is applied to summarizing properties of yeast stress by searching for dependencies (commonalities) between expression of genes of bakers yeast Saccharomyces cerevisiae in various stressful treatments, and summarizing stress regulation by finally adding data about transcription factor binding sites.

On text-based estimation of document relevance

This work is part of a proactive information retrieval project that aims at estimating relevance from implicit user feedback. The noisy feedback signal needs to be complemented with all available information, and textual content is one of the natural sources. Here we take the first steps by investigating whether this source is at all useful in the challenging setting of estimating the relevance of a new document based on only few samples with known relevance. It turns out that even sophisticated unsupervised methods like multinomial PCA (or latent Dirichlet allocation) cannot help much. By contrast, feature extraction supervised by relevant auxiliary data may help.

Two-Way Grouping by One-Way Topic Models

We tackle the problem of new users or documents in collaborative filtering. Generalization over users by grouping them into user groups is beneficial when a rating is to be predicted for a relatively new document having only few observed ratings. The same applies for documents in the case of new users. We have shown earlier that if there are both new users and new documents, two-way generalization becomes necessary, and introduced a probabilistic Two-Way Model for the task. The task of finding a two-way grouping is a non-trivial combinatorial problem, which makes it computationally difficult. We suggest approximating the Two-Way Model with two URP models; one that groups users and one that groups documents. Their two predictions are combined using a product of experts model. This combination of two one-way models achieves even better prediction performance than the original Two-Way Model.