JultikaUniversity of Oulu repository

Abstract

Several businesses/smart applications rely on personalizing their services to adapt to the user’s preferences. Personalized services are developed using recommendation systems based on user’s feedback on products/services, needs, habits and social or demographic characteristics. Several businesses from e-commerce (suggesting users what to buy) to hospitality services (suggesting which hotel to book) focus on using recommendation systems to achieve a personalized experience for their users. Majority of recommendation systems make use of only product ratings shared by the users, this may pose challenges like sparsity of ratings. The wide availability of other attributes of products or users like textual product reviews provided by users or product descriptions in e-commerce and hospitality domains present a gold mine of additional personalising information with which to supplement their ratings based recommendation system.

Recommendation systems majorly involves two tasks: rating (predict ratings that user might assign to a product) and ranking (recommend products based on predicted rank scores) prediction tasks. In this thesis, we propose a novel hybrid recommendation system using the state-of-the-art DeepFM model which makes use of multiple textual features derived from product reviews particularly contextual sentence embedding vectors, average sentiment scores and linguistic cues such as presence/absence of negation in the product reviews in combination with ratings shared by users to enhance the prediction of the desired ratings or rank scores. We evaluated our system with commercial datasets from Amazon and Datafiniti for both tasks: predicting rating and recommendations based on predicted rank scores. We utilised different metrics for both types of tasks. From our evaluation we infer that using contextual sentence embedding vectors extracted using BERT, average sentiment scores and presence/absence of negation in the product reviews obtained from VADER, does impact the prediction of ratings and recommendations based on predicted scores of the recommendation system which only utilises product ratings as user preferences. Furthermore, we can conclude from our evaluation that (A) contextual embedding vectors and average sentiment scores together along with ratings in the proposed hybrid system improves prediction of desired ratings, (B) contextual embedding vectors, average sentiment scores and presence/absence of negation in the product reviews together along with ratings in the proposed hybrid system improves prediction of desired ratings as well, (C) contextual embedding vectors and average sentiment scores together along with ratings in the proposed hybrid system improves recommendations based on rank scores.