A framework for leveraging properties of user reviews in recommendation

Wang, Xi (2022) A framework for leveraging properties of user reviews in recommendation. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2022XiWangPhD.pdf] PDF
Download (2MB)


With the growing volume of information online, it is increasingly harder for users to identify useful information to support their choices when interacting with different items. Review-based recommendation systems, which leverage reviews posted by users on items to estimate the users’ preferences, have been shown to be a credible solution for addressing the problem of identifying their preferred items. However, the actual usefulness of these reviews impact the effectiveness of the resulting recommender systems, especially with the enormous volume of available reviews online. In particular, as argued by the widely cited users’ adoption of information framework, users exhibit distinct preferences for reviews depending on the properties of these reviews (e.g. length, sentiment) when making decisions. Therefore, we argue that not all reviews are equally useful for different users. We aim to effectively modelling the personalised usefulness of reviews through the use of reviews’ properties when developing review-based recommendation techniques. Note that, few studies in the literature investigated the effectiveness of leveraging the properties of reviews to develop effective review-based recommendation approaches.

This thesis aims to address this research gap by proposing a review-based recommendation framework. Such a framework models the personalised usefulness of reviews according to various reviews’ properties, including the reviews’ age, length, sentiment, ratings, helpfulness as judged by the users and helpfulness as predicted by a review helpfulness classifier. In particular, the thesis addresses two main challenges: (i) the availability of the attributes of reviews and (ii) the users’ preferences estimation.

The first challenge refers to the difficulty of extracting particular review properties from their corresponding attributes. For example, extraction of the age property relies on the availability of the timestamps of the corresponding reviews. We address the availability of the reviews’ attributes to extract their sentiment and helpfulness properties with classification techniques. The sentiment property of reviews is estimated through effective state-of-the-art sentiment classifiers. We first evaluate the estimated reviews’ sentiment in comparison to the users’ ratings in typical recommendation approaches. Then, we introduce a sentiment attention mechanism to encode the estimated reviews’ sentiment. Our experiments show that the sentiment property can effectively replace the users’ ratings when estimating the user preferences. Moreover, by leveraging the estimated sentiment property of reviews, our proposed review-based rating prediction model shows improved performance compared to state-of-the-art rating prediction models. Next, the extraction of the reviews’ helpfulness property leverages the reviews’ helpful votes (i.e. a type of feedback given by other reviewers providing information on whether the corresponding review is helpful to them). However, the number of helpful votes for each review are not commonly available. In particular, we propose a novel weakly supervised review helpfulness classification correction approach (i.e. the Negative Confidence-aware Weakly Supervised (NCWS) approach), which leverages the confidence in a given review being unhelpful with respect to its age. We experimentally show that NCWS-based classifiers significantly outperform existing review helpfulness classifiers on two public review datasets. Moreover, the estimated helpfulness of reviews by NCWS-based classifiers can significantly improve the performance of a review-based rating prediction model.

Next, to address our second challenge pertaining to the users’ preferences estimation, we aim to estimate their preferences when using reviews exhibiting different properties to accurately predict their preferred items. In particular, we propose two novel ranking-based recommendation approaches (named RPRM and BanditProp), which models the users’ preferences using different review properties with different techniques. The RPRM model applies the attention mechanism to model the usefulness of reviews according to different review properties. Unlike RPRM, the BanditProp model leverages existing bandit algorithms and introduces a novel contextual bandit algorithm to tackle the users’ preference estimation of using specific reviews’ properties to identify useful reviews. Our experiments show that RPRM can outperform stateof-the-art review-based recommendation models, and BanditProp can significantly outperform RPRM on two publicly available review datasets. These results validate the effectiveness of leveraging the review properties when examining the usefulness of reviews to improve the performances of review-based recommendation techniques.

Overall, we contribute an effective review-based recommendation framework that make accurate recommendations by leveraging the reviews’ associated properties. This framework includes models for extracting properties from reviews, and various techniques that are required to integrate the learned properties, which, in turn and according to our conducted experiments, provide good approximations of a given users’ item preferences. These contributions make progress in the development of review-based recommendation techniques and could inspire future directions of research in recommendation systems.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Colleges/Schools: College of Science and Engineering > School of Computing Science
Supervisor's Name: Ounis, Professor Iadh and Macdonald, Dr. Craig
Date of Award: 2022
Depositing User: Theses Team
Unique ID: glathesis:2022-82979
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 17 Jun 2022 08:26
Last Modified: 17 Jun 2022 08:26
Thesis DOI: 10.5525/gla.thesis.82979
URI: https://theses.gla.ac.uk/id/eprint/82979
Related URLs:

Actions (login required)

View Item View Item


Downloads per month over past year