Recomendação associativa de tags na ausência de informação prévia

Abstract

Among the existing tag recommendation strategies, those that have consistently produced state-of-the-art performance often exploit term co-occurrence patterns with tags previously assigned to the target object, possibly jointly with other textual features. However, these techniques have been evaluated only for objects with previously assigned tags. Instead, we here focus on the recommendation of tags for objects with empty tags, a variation of a well known problem called cold start. We start by evaluating state-of-the-art co-occurrence based methods in face of cold start, using datasets collected from for popular Web 2.0 applications. Our results indicate that the effectiveness of these methods suffer greatly when they cannot rely on previously assigned tags in the target object. Moreover, we show that employing various automatic filtering strategies to generate a initial set of tags that enables the use of co-ocurrence patterns produces limited gains. We then propose a new approach that exploits both positive and negative relevance feedback (RF) from the users to iteratively select input tags to these methods. Our experimental results, comparing several variations of the proposed RF strategy against the original co-occurrence based methods and other state-of-the-art baselines in various scenarios, indicate that extending the methods to include relevance feedback from users leads to significant gains in precision over the best considered baseline in cold start scenarios, and also on objects that contain some initial tags (i.e., no cold start). We also show that our best RF-driven strategy is robust to the lack of user reliability and cooperation, remaining at least as good if not superior to the best baseline even if the feedback provided by user contains noise or if users select only a portion of the relevant tags that are shown to them in the feedback process.