DUPUY Christophe

Most of the current recommendation systems are based on ratings (i.e., numbers between 0 and 5) to recommend a content (movie, restaurant.) to a user. The latter often has the possibility to comment on this content in the form of text in addition to rating it. It is difficult to extract information from a raw text while a simple note contains little information about the content and the user. In this thesis, we attempt to suggest a personalized readable text to the user to help him/her quickly form an opinion about a content. More specifically, we first build a thematic model predicting a personalized movie description from textual comments. Our model separates qualitative (i.e., opinionated) themes from descriptive themes by combining textual comments and number scores in an attached probabilistic model. We evaluate our model on an IMDB database and illustrate its performance through theme comparison. We then study parameter inference in large-scale latent variable models, including most theme models. We propose a unified treatment of online inference for latent variable models from non-canonical exponential families and make explicit the links between several previously proposed frequentist and Bayesian methods. We also propose a new inference method for frequentist parameter estimation that adapts MCMC methods to online inference of latent variable models by properly using local Gibbs sampling. For the latent Dirichlet allocation topic model, we provide an extensive set of experiments and comparisons with existing work in which our new approach performs better than previously proposed methods. Finally, we propose a new class of determinantal point processes (DPPs) that can be manipulated for parameter inference and learning in potentially sub-linear time in the number of objects. This class, based on a specific low-rank factorization of the marginal kernel, is particularly suited to a subclass of continuous PPDs and PPDs defined over an exponential number of objects. We apply this class to the modeling of text documents as samples of a PPD on sentences and propose a conditional maximum likelihood formulation for modeling topic proportions, which is made possible without any approximation with our class of PPDs. We present an application to document summarization with a PPD on 2 to the power of 500 objects, where the summaries are composed of readable sentences.

We study parameter inference in large-scale latent variable models. We first propose an unified treatment of online inference for latent variable models from a non-canonical exponential family, and draw explicit links between several previously proposed frequentist or Bayesian methods. We then propose a novel inference method for the frequentist estimation of parameters, that adapts MCMC methods to online inference of latent variable models with the proper use of local Gibbs sampling. Then, for latent Dirich-let allocation,we provide an extensive set of experiments and comparisons with existing work, where our new approach outperforms all previously proposed methods. In particular, using Gibbs sampling for latent variable inference is superior to variational inference in terms of test log-likelihoods. Moreover, Bayesian inference through variational methods perform poorly, sometimes leading to worse fits with latent variables of higher dimensionality.

We propose a new class of determinantal point processes (DPPs) which can be manipulated for inference and parameter learning in potentially sublinear time in the number of items. This class, based on a specific low-rank factorization of the marginal kernel, is particularly suited to a subclass of continuous DPPs and DPPs defined on exponentially many items. We apply this new class to modelling text documents as sampling a DPP of sentences, and propose a conditional maximum likelihood formulation to model topic proportions, which is made possible with no approximation for our class of DPPs. We present an application to document summarization with a DPP on $2^{500}$ items.

Streaming services such as Netflix, M-Go, and Hulu use advanced recommender systems to help their customers identify relevant content quickly and easily. These recommenders display the list of recommended movies organized in sublists labeled with the genre or some more specific labels. Unfortunately , existing methods to extract these labeled sublists require human annotators to manually label movies, which is time-consuming and biased by the views of annotators. In this paper, we design a method that relies on crowd sourced reviews to automatically identify groups of similar movies and label these groups. Our method takes the content of movie reviews available online as input for an algorithm based on Latent Dirichlet Allocation (LDA) that identifies groups of similar movies. We separate the set of similar movies that share the same combination of genre in sublists and personalize the movies to show in each sublist using matrix factorization. The results of a side-by-side comparison of our method against Technicolor's M-Go VoD service are encouraging.