Exploring User and Item Representation, Justification Generation, and Data Augmentation for Conversational Recommender Systems

Conversational Recommender Systems (CRS) aim to provide personalized and contextualized recommendations through natural language conversations with users. The objective of my proposed dissertation is to capitalize on the recent developments in conversational interfaces to advance the field of Recommender Systems in several directions. I aim to address several problems in recommender systems: user and item representation, justification generation, and data sparsity. A critical challenge in CRS is learning effective representations of users and items that capture their preferences and characteristics. First, we focus on user representation, where we use a separate corpus of reviews to learn user representation. We attempt to map conversational users into the space of reviewers using semantic similarity between the conversation and the texts of reviews. Second, we improve item representation by incorporating textual features such as item descriptions into the user-item interaction graph, which captures a great deal of semantic and behavioral information unavailable from the purely topological structure of the interaction graph. Justifications for recommendations enhance the explainability and transparency of CRS; however, existing approaches, such as rule-based and template-based methods, have limitations. In this work, we propose an extractive method using a corpus of reviews to identify relevant information for generating concise and coherent justifications. We address the challenge of data scarcity for CRS by generating synthetic conversations using SOTA generative pre trained transformers (GPT). These synthetic conversations are used to augment the data used for training the CRS. In addition, we also evaluate if the GPTs exhibit emerging abilities of CRS (or a non-conversational RecSys) due to the large amount of data they are trained on, which potentially includes the reviews and opinions of users.