Research

In the world of recommender systems, so-called content-based methods are an important approach that rely on the availability of detailed product or item descriptions to drive the recommendation process. For example, recommendations can be generated for a target user by selecting unseen products that are similar to the products that the target user has liked or purchased in the past. To do this, content-based methods must be able to compute the similarity between pairs of products (unseen products and liked products, for example) and typically this This work is supported by Science Foundation Ireland under grant 07/CE/I1147. The Insight Centre for Data Analytics is supported by Science Foundation Ireland under Grant Number SFI/12/RC/2289. R. Dong CLARITY: Centre for Sensor Web Technologies School of Computer Science and Informatics University College Dublin, Dublin, Ireland E-mail: ruihai.dong@ucd.ie M. P. O'Mahony Insight Centre for Data Analytics School of Computer Science and Informatics University College Dublin, Dublin, Ireland E-mail: michael.omahony@ucd.ie M. Schaal CLARITY: Centre for Sensor Web Technologies School of Computer Science and Informatics University College Dublin, Dublin, Ireland E-mail: markus.schaal@ucd.ie K. McCarthy Insight Centre for Data Analytics School of Computer Science and Informatics University College Dublin, Dublin, Ireland E-mail: kevin.mccarthy@ucd.ie B. Smyth Insight Centre for Data Analytics School of Computer Science and Informatics University College Dublin, Dublin, Ireland E-mail: barry.smyth@ucd.ie 2 Ruihai Dong et al. is achieved by comparing product features or other descriptive elements. The ap- proach works well when product descriptions are readily available and when they are detailed enough to aord an eective similarity comparison. But this is not always the case. Detailed product descriptions may not be available since they can be expensive to create and maintain. In this article we consider another source of product descriptions in the form of the user-generated reviews that frequently accompany products on the web. We ask whether it is possible to mine these re- views, unstructured and noisy as they are, to produce useful product descriptions that can be used in a recommendation system. In particular we describe a novel approach to product recommendation that harnesses not only the features that can be mined from user-generated reviews but also the expressions of sentiment that are associated with these features. We present a recommendation ranking strategy that combines similarity and sentiment to suggest products that are sim- ilar but superior to a query product according to the opinion of reviewers, and we demonstrate the practical benets of this approach across a variety of Amazon product domains.