Recommender systems are software tools that help users to obtain a list of items (i.e., products, services, people, etc.) as answer to (implicit) information needs. In such systems, user and item attributes, as well as user past behavior, are used to estimate a user’s preferences and hence provide her with personalized suggestions of items of potential relevance. In this context, there are many domains –such as restaurants, hotels and e-commerce– where users usually rate available items, and provide textual reviews supporting their ratings. These reviews are a very useful source of information about the user preferences. In particular, the opinions (sentiments) the user has about specific aspects (features, components, etc.) of the items can be exploited to improve the quality of her profile. Recommender systems that use such aspect information are called aspect-based recommender systems. In this master thesis, we address the aspect-based recommendation task as a threestage problem. Firstly, performing an aspect extraction process where potential item aspects are identified in textual reviews. Secondly, estimating the opinion about each aspect a user has commented on. Finally, exploiting the obtained aspects and their opinion polarities (i.e., positive, neutral, and negative) to generate effective personalized recommendations. For the aspect extraction task, we have developed and empirically compared two state-of-the-art, popular methods, namely Double Propagation –which exploits semantic relations between the words in the review to establish which of such words may correspond to aspects–, and Topic Models –which find latent topics as a proxy for the aspects. Next, for the aspect opinion polarity estimation stage, we have followed the common strategy of using a lexicon –i.e., a list of well-known words that are positive or negative in general domains–, but differently to previous work, we have used Natural Language Processing techniques and resources to better estimate the opinion polarity in negated adjectives and negative sentences. Finally, for the aspect-recommendation stage, we have implemented and evaluated numerous recommenders of several types, such as content-based, collaborative filtering, and hybrid, with and without using aspect-based information. We have conducted an exhaustive experimentation on several domains with relatively large datasets, and computing a wide array of metrics. The obtained results show that considering the opinion about item aspects generates valuable recommendations that improve the performance of personalized recommendation methods, and have empirically proved that content-based recommendation approaches with an appropriate aspect-based user representation achieves the best performance results.

Máster Universitario en Investigación e Innovación en Tecnologías de la Información y las Comunicaciones