Aspect-based sentiment analysis for social recommender systems.

Social recommender systems harness knowledge from social content, experiences and interactions to provide recommendations to users. The retrieval and ranking of products, using similarity knowledge, is central to the recommendation architecture. To enhance recommendation performance, having an effective representation of products is essential. Social content such as product reviews contain experiential knowledge in the form of user opinions centred on product aspects. Making sense of these for recommender systems requires the capability to reason with text. However, Natural Language Processing (NLP) toolkits trained on formal text documents encounter challenges when analysing product reviews, due to their informal nature. This calls for novel methods and algorithms to capitalise on textual content in product reviews together with other knowledge resources. In this thesis, methods to utilise user purchase preference knowledge - inferred from the viewed and purchased product behaviour - are proposed to overcome the challenges encountered in analysing textual content. This thesis introduces three major methods to improve the performance of social recommender systems. First, an effective aspect extraction method that combines strengths of both dependency relations and frequent noun analysis is proposed. Thereafter, this thesis presents how extracted aspects can be used to structure opinionated content enabling sentiment knowledge to enrich product representations. Second, a novel method to integrate aspect-level sentiment analysis and implicit knowledge extracted from users' product purchase preferences analysis is presented. The role of sentiment distribution and threshold analysis on the proposed integration method is also explored. Third, this thesis explores the utility of feature selection techniques to rank and select relevant aspects for product representation. For this purpose, this thesis presents how established dimensionality reduction approaches from text classification can be employed to select a subset of aspects for recommendation purposes. Finally, a comprehensive evaluation of all the proposed methods in this thesis is presented using a computational measure of 'better' and Mean Average Precision (MAP) with seven real-world datasets