Research Workshop on AI for Preference Learning: Sentiment, Comparison, and Recommendation

Detecting the sentiment expressed by a document is a key task for many applications, e.g., modeling user preferences, monitoring consumer behaviors, assessing product quality. Traditionally, the sentiment analysis task primarily relies on textual content. Fueled by the rise of mobile phones that are often the only cameras on hand, documents on the Web (e.g., reviews, blog posts, tweets) are increasingly multimodal in nature, with photos in addition to textual content. A question arises whether the visual component could be useful for sentiment analysis as well. In this talk, we will discuss about the proposed Visual Aspect Attention Network or VistaNet, leveraging both textual and visual components. In many cases, with respect to sentiment detection, images play a supporting role to text, highlighting the salient aspects of an entity, rather than expressing sentiments independently of the text. Therefore, instead of using visual information as features, VistaNet relies on visual information as alignment for pointing out the important sentences of a document using attention mechanism.

A number of real-world applications require comparison of entities based on their textual representations. In this talk we discuss a topic model supervised by pairwise comparisons of documents. Such a model seeks to yield topics that help to differentiate entities along some dimension of interest, for example, product ratings. While previous supervised topic models consider document labels in an independent and pointwise manner, our proposed Comparative Latent Dirichlet Allocation (CompareLDA) learns predictive topic distributions that comply with the pairwise comparison observations.

Top-K recommendation seeks to deliver a personalized recommendation list of K items to a user. The dual objectives are (1) accuracy in identifying the items a user is likely to prefer, and (2) efficiency in constructing the recommendation list in real time. An established methodology in the literature based on matrix factorization (MF) which usually represents users and items as vectors in low-dimensional space, is an effective approach to recommender systems. Since most MF-based algorithms rely on inner product as predictor, the retrieval of MF recommendations can be formulated as maximum inner product search (MIPS) problem. Though there have been a number of works to quickly train MF models that can effectively handle millions of users and items, naive solutions for retrieving top preferred items involves ranking every single item for a user, which may inhibit truly real-time retrieval performance. In this presentation, we first explain the recommendation retrieval problem and the prohibitive complexity of linear scanning solution for real-time applications. We then provide a comprehensive summary of existing approaches in the literature that attempt to solve MIPS efficiently via approximate solutions such as indexing, sampling, and sequential scanning with upper-bounding.