Denis Parra Santander

A list of my research interests and what I have done on each of them:

I started working on this topic during my internship at Telefonica I+D in the summer of 2010 with Dr. Xavier Amatriain. Focusing in the music domain, we were wondering if there was a reliable way to map implicit behaviour (albums or songs playcounts) with explicit preference (in our research, in the form of ratings) in such a way that we could use the inferred explicit information to provide recommendations of music to users. We performed a user study with last.fm users and using ANOVA and simple regression we were able to find some factors that could explain the variance of the dependent variable (the rating that users gave to albums in the user study). Our work opened many questions that we are currently doing research on.

On this area I have been working in the Latent Communities Project, which main goal is to develop a set of tools for analysts in order to use Social Network Analysis tools to analyze social networks and their explicit and implicit communities. We are a team working on this project, but I particularly integrated the Open Source tools Gephi (for network visualization and manipulation) with Mallet (in the particular, for Topic Modelling) in such a way that Latent Dirichlet Allocation (LDA) can be used to identify communities in networks, as described in the paper "An lda-based community structure discovery approach for large-scale social networks" by Zhang et al.

In neurophysiology interactions between neurons are modeled by way of activation which propagates from one neuron to another via connections called synapses to transmit information using chemical signals. The first spreading activation models were used in cognitive psychology to model these processes of memory retrieval. This framework was later exploited in Artificial Intelligence as a processing framework for semantic networks and ontologies, and applied to Information Retrieval as the result of direct transfer of information retrieval ideas from cognitive sciences to AI, and it has also been used for trust propagation on the Web.
In our research, results of numerical simulations show that spreading activation algorithms allow discriminating the degree of connectivity of users between certain graph structures connecting users via resources and tags.

I worked in using the data from collaborative tagging systems such as delicious or citeulike to develop new approaches for recommender systems. My first step on this area was the Workshop paper Evaluation of Collaborative Filtering Algorithms for Recommending Articles on CiteULike that I will present on the Hypertext 2009 conference in Torino, Italy. On this paper, I compare three Collaborative Filtering approaches to recommend scientific articles to CiteULike users. The methods are Classic Collaborative Filtering (CCF), Neighbor-weighted Collaborative Filtering (NwCF) and Okapi BM25 based similarity (BM25).