Aspect-based Recommendation

  • Author / Creator
    Mirzaei, Maryam
  • The problem of aspect-based recommendation---recommending an "item" to a "recommendation recipient" based on "aspects", i.e., information about the characteristic features of the item that may be of interest to the recommendation recipient or what makes an item a good match for a recommendation recipient---is attracting substantial research attention recently.

    In this work, we study two problems in this general area.

    First, we consider the problem of assigning reviewers to papers submitted for publication to a conference. We cast this problem as the recommendation of a set of experts as appropriate reviewers for a paper. Papers in this case correspond to "recommendation recipient", and we consider the thematic areas or topics of a paper as their "aspects". Potential reviewers correspond to "items", and we consider the expertise areas of reviewers when considering the importance of the papers aspects . The paper aspects can be inferred from terms extracted from the paper description (title and abstract); the reviewer’s expertise can similarly be extracted from the descriptions of the papers they have authored. Our reviewer-recommendation algorithm assigns to each submitted paper a set of reviewers who can evaluate all ‌aspects of the paper, while at the same time, maximizing the relevant expertise of the reviewers and balancing their workload.

    Next, we consider the problem of personalized and explainable aspect-based recommendations of products and services based on online reviews. Our algorithm recommends items to users by capturing the dependencies between the sentiments that reviews express towards different item aspects, and using the importance of these aspects for each target user. In this scenario, the algorithm effectively predicts the user’s sentiments toward candidate item aspects, and uses these predicted sentiments as de-facto explanations for the items it selects to recommend.

    In all stages of our work we experimentally validate our methods on a variety of datasets from different domains and we experimentally demonstrate its superior performance relative to other state-of-the-art approaches.

  • Subjects / Keywords
  • Graduation date
    Fall 2021
  • Type of Item
  • Degree
    Doctor of Philosophy
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.