Exploiting relevance, novelty and diversity in tag recommendation

Abstract

Tag recommendation methods have mostly focused on maximizing relevance, but other aspects may be as important for recommendation usefulness. We here define novelty and diversity for tag recommendation, and propose two new recommendation strategies that consider these aspects jointly with relevance. We evaluate the proposed strategies using real datasets from popular Web 2.0 applications, achieving gains over the state-of-the-art of up to 21% in relevance, 45% in novelty and 2.5% in diversity.

Recomendação Colaborativa de Conteúdos Educacionais para Dispositivos Portáteis

“…Foram tomados como parâmetros os itens mais citados em m-learning pelos autores Huang et al (2012), Ozdamli and Cavus (2011), Lichtnow et al (2006), Belém et al (2012) e Lemos et al (2012), como Colaboração/Interatividade (Col/Int) entre aluno-aluno e aluno-professor, Adaptação de conteúdos (Adap), Avaliação/Feedback (Aval/Feed), Recomendação de conteúdos (RecCont) e Informações de Contexto (Contexto). Outros parâmetros foram adicionados para caracterização dos trabalhos, pois são considerados relevantes na arquitetura proposta.…”

Topic diversity in tag recommendation

“…When recommending tags for a target object, relevance refers to how well the recommended tags describe the contents of the target object. However, relevance by itself may not be enough to guarantee recommendation usefulness and effectiveness [4,25]. For example, a list of synonyms that well describe the object's content is arguably relevant, but also redundant and less useful than a more diversified list covering more aspects related to the object.…”

“…For example, a list of synonyms that well describe the object's content is arguably relevant, but also redundant and less useful than a more diversified list covering more aspects related to the object. As argued in [4], objects on the Web 2.0 may be multifaceted, being related to various aspects and topics.…”

“…While diversity is an important aspect in tag recommendation, our previous work [4] is, to our knowledge, the only one to address aspects other than relevance for tag recommendation. In [4], we defined the diversity of a list of recommended tags implicitly, as the average semantic distance [4,13,19] between each pair of tags in the list, such that a set of synonyms or semantically similar words has low diversity.…”

“…In [4], we defined the diversity of a list of recommended tags implicitly, as the average semantic distance [4,13,19] between each pair of tags in the list, such that a set of synonyms or semantically similar words has low diversity. However, this approach has the disadvantage of being more subject to data sparsity, because the estimates of the semantic similarities are based on tag co-occurrences [4].…”