|
About The Talk
Recommendation explanations help to make sense of recommendations, increasing the likelihood of adoption. While they are strongly related to explainable recommendations, which seek to provide not only accurate recommendations but also accompanying explanations for those recommendations, the task of explanation can be decoupled from that of recommendation, casting the recommendation explanation as a research problem in its own right. We can categorize recommendation explanation into integrated and pipeline approaches. The former aims at a single interpretable model for both recommendation and explanation tasks. The latter produces explanations after having recommendations by another recommendation model.
We are interested in mining product textual data for recommendation explanation. Although it is an unstructured datatype, textual data may come from manufacturers, sellers, and consumers. It also appears in many places, e.g., title, summary, description, review, question and answers, etc., that could be a rich source of information for recommendation explanations. Recommendation explanations appear in different forms such as content-based explanation, rules, topics, or social, etc. In this dissertation, we focus on diverse natural language explanation and encompass both integrated and pipeline approaches. Depending on the characteristics of the models, information used, as well as the desired experience that the system wants users to have, the research can produce different forms of recommendation explanations. It could be generic, as they explain how the recommendation engine works, evaluative, accessing the quality of a product and of itself, or comparative, accessing a recommendation in comparison to another reference product. For pipeline approach, we proposed a post-hoc recommendation explanation approach, called SEER, that synthesizes an explanation by selecting sentences (opinion phrases could be substituted to match user preference) from other reviewers satisfying an aspect demand of interests from the target user. For integrated approach, we proposed a comparative explainable recommendation model, called ComparER, which anchors reference products on the previously adopted products in a user history. Experiments on Amazon Reviews dataset show the efficacies of the proposed methods.
|
