Relevance, Novelty, Diversity and Personalization in Tag Recommendation

The design and evaluation of tag recommendation methods have historically focused on maximizing the relevance of the suggested tags for a given object, such as a movie or a song. Tag relevance can be defined in two perspectives. In an object-centered perspective, a tag is relevant if it correctly describes the content of the target object, while in a personalized perspective, a relevant tag not only describes well the content of the target object, but also matches the interests of the target user. However, even enriched by a personalized perspective, relevance by itself may not be enough to guarantee recommendation usefulness. Promoting novelty and diversity in tag recommendation not only increases the chances that the user will select some of the recommended tags, but also promotes complementary information (i.e., tags), which helps cover multiple aspects or topics related to the target object. Yet, no prior work has tackled novelty and diversity in the specific context of tag recommendation. In this thesis, we aim at proposing novel solutions that effectively address multiple aspects related to the tag recommendation problem, notably, relevance, novelty, diversity and personalization of the suggested tags. We evaluate our strategies using real data from five Web 2.0 applications, namely, Bibsonomy, LastFM, MovieLens, YahooVideo and YouTube. Our experimental results demonstrate the effectiveness of our new methods over state-of-the-art approaches, and attest the viability to effectively increase novelty and diversity with only a slight impact (if any) on relevance. We also found that our proposed syntactic attributes are responsible for significant improvements (up to 17% in precision) over the best relevance-driven method in a cold start scenario. In addition, we assessed the benefits of personalization to provide better descriptions of the target object, with average gains of 15% in relevance over the best object-centered approach.

Multimodal Semantics and Affective Computing from Multimedia Content

Advancements in technologies and increasing popularities of social media websites have enabled people to view, create, and share user-generated content (UGC) on the web. This results in a huge amount of UGC (e.g., photos, videos, and texts) on the web. Since such content depicts ideas, opinions, and interests of users, it requires analyzing the content efficiently to provide personalized services to users. Thus, it necessitates determining semantics and sentiments information from UGC. Such information help in decision making, learning, and recommendations. Since this chapter is based on the intuition that semantics and sentiment information are exhibited by different representations of data, the effectiveness of multimodal techniques is shown in semantics and affective computing. This chapter describes several significant multimedia analytics problems such as multimedia summarization, tag-relevance computation, multimedia recommendation, and facilitating e-learning and their solutions.