ICON 2018ICON 2018 - Tutorial

Product Review Recommendation: Ranking, Selection and Personalization


Reviews ranked by their helpfulness facilitate purchase decisions by potential buyers. For a balanced coverage though, such selection of top‐k reviews should span the entire spectrum of aspects and sentiments of relevant products. Personalization makes review recommendation more effective as user preference varies not only across authors but also voters of reviews and raters of products. Review text helps capture user idiosyncrasies; a variety of techniques exist to represent this diverse data across users/products thus leading to further advances in factorization and AI methods for showing useful reviews to consumers ‐ an important need in online shopping. Proposed tutorial thus covers trending topics of interest to ICON community: ranking, elicitng/interpreting user preferences and deep learning with a unique perspective synergizing NLP(review text) with data mining (utility prediction).


90 mintues

Detailed bulleted outline of the tutorial

  1. Introduction (10 minutes)
    1. Role of Reviews:  Reviews form a low-cost and efficient feedback channel
      1. enable buyers to record their sentiments about the product and its aspects
      2. thus helping other consumers make purchase decisions and
      3. Provide feedback to merchants.
    2. Addressing Information Overload: Too many reviews for a product cannot be read. Show top K reviews
    3. Review recommendation: Select a small representative set of high-quality reviews which covers various aspects of the product.
    4.  Reviews may be ranked based on their quality scores. But the K reviews with the highest scores may contain redundancy in content or opinion.
    5. Personalized review recommendations are more helpful as different individuals may be interested in different aspects of the product.
    6. Identifying product aspects and associated sentiments: Topic models and word embedding can capture product features, user preferences and reviewer characteristics.
    7. Outline: In this tutorial we discuss methods of representing the reviews and the user interests in terms of the fine grained aspects of the product and the associated sentiments. We present methods to select reviews for recommendation to a user. We also discuss evaluation methodologies.
  2. Review scoring: Helpfulness prediction (10 minutes)
    1. Utility score of reviews is typically computed through a function learnt from social (e.g., number of past reviews, PageRank) and structural (e.g., length, readability) features and content of reviews.
      1. Social: [3] explores two roles users play in review helpfulness score, e.g., authors with high reputations are likely to write helpful reviews, and product raters find reviews from their connected authors more helpful.
      2. Structural features of reviews: length, readability
      3. Ranking of reviews by content using convolutional neural network.
    2. Personalized scoring: The quality of a review is not independent from its readers - latent factors of whom affect the evaluation.
  3. Review Selection: 10 minutes
  4. A user has to be recommended a subset of reviews. The selected set should have the following characteristics

    1. Comprehensive: Coverage over al important product aspects
    2. Representative: Representative distribution of product aspects and associated opinions as in the entire set of reviews
    3. Personalized: Emphasis on aspects of interest to the user.

    [4] proposes strategies for pruning review search space and objective functions/greedy algorithms for diverse, quality and representative review sets. [5] leverages short, concise and focused snippets for review selection by matching as few review sentences to as many micro-reviews.

  5. Context Modeling: 25+20 minutes
    1. Aspect Based Sentiment Analysis (Topic Models):  Review Selection depends on identifying the aspects and their sentiments from individual reviews and aggregating them to form product and user representation. There are different methods for the aspect and semantic extraction
      1. Rule Based
      2. Topic Model based: We will discuss some of the topic model based methods which are general and portable across domains.

      Matrix and tensor factor models help capture latent features of reviews, authors, raters and products. A topic model may be used to jointly discover the underlying aspects and sentiments guided by review helpfulness voting information.

      The demographic information of review authors is incorporated into topic modeling process in order to discover associations between market segments, topical aspects and sentiments [8].
      Latent factors like review expertise, writing style and judgement about fine-grained product aspects are explored for utility prediction through a hybrid HMM-LDA model [9]. [10] helps identify most valuable aspects of user’s potential experience with an item for item recommendation with aspects over which they have control. User preferences, item properties and their interaction are captured in a generative model for personalized review ranking [11].

    2. Distributed Representation: A shared layer helps couple latent factors learned for user behavior and item properties from reviews by two parallel networks [11]. Gated recurrent neural networks help translate user and item latent representations into concise abstractive tips with good linguistic quality simulating user experience and feelings [12]. By translating various sources (e.g., review, rating) into a unified representation space, heterogeneous information can be integrated for informed recommendation [13]. Highly-useful reviews are obtained through a novel, attention mechanism and provide explanations for users to make better and faster decisions [14]. [15] combines embedding method with an easy-to-interpret attention network, for explainable recommendations, with a tree-based model that learns decision rules form side information (e.g., user demographics).
  6. Review Recommendation Systems (15 minutes)
    1. User independent
    2. Personalized