Topic Title: Improve recommender system and demand forecasting through the use of customer review data

Technical Area: Natural language processing



In e-commence system, large amounts of review data have been generated by customers considering various aspects towards products and services they received, i.e., quality, price, logistics and so on, reflecting customers’ positive or negative feelings. Even though review data is informative about user’s preference among products and the aspects they would value, however, this information has not been fully exploited. In this project, we propose to use review data in both recommendation and demand forecasting tasks, extracting additional information by deep neural network from customer feedbacks in order to improve the prediction accuracy.


Recommender system is particularly important for e-commence companies given the fact that customers are flooded by large amounts of products and services. It is challenging for companies to provide personalized and likely interested products to customers to improve user experience and profitability. Current recommender systems usually exploit features such as users’ profile data, behavior sequence (product glance view, orders, and rating scores, etc.), and environmental information (weather, holiday, location and events etc.), assuming that customers with common behavior in the past are likely to have similar needs in the future. Using customer’s review data can introduce high dimensional subjective semi-quantification of products which reflects specific details customers would emphasize as well as their sentiments towards them. This information helps to infer customers’ attitudes toward products with similar properties and can possibly alleviate the well-known cold start problem especially for new products. Another application of customer review data is demand forecasting, given that customer’s opinion of products might be affected by others’ comments. In extreme case, when most comments of a product are either positive or negative, the future demand quantity will possibly be affected.


Researchers have been trying to introduce customer review text as part of algorithm input to improve performance while predicting rating scores, analyzing comment topics, and modeling customers’ and products’ information simultaneously, etc. However, some aspects of customer review text have not been considered, e.g., evolution and variation of customer’s review data at different time and on different categories of products, as well as the cross correlation between individual’s review on products and product’s reviews received from all customers. Thus, we would invite researchers who are passionate and expertized in natural language processing and recommender system to further explore this topic and implement potential outcomes in our realistic tasks.



This project will focus on exploiting customers’ review data to improve the performance of current implemented algorithms in both recommender system and demand forecasting, and we target to investigate the correlation between customers’ behavior / demands and products’ review data using cutting-edge natural language processing techniques, and furthermore build a deep neural network framework incorporating both traditional features and customers reviews for our tasks.


Related Research Topics

One trend for recommendation and demand forecasting is to use neural network based models with deep learning techniques. De-noising Auto-encoders and Restricted Boltzmann Machines have been considered as collaborative based techniques to utilize the rating matrix for recommender system. Convolutional Neural Network and deep belief network have also been used to learn latent factors for recommendation to include both user profile and behavior. Other researchers explored the user and product latent factors in a joint manner and map them to a common space. Recurrent Neural Network has been used to process customers’ review data, which models the likelihood of reviews using the products’ latent factors and combines with the matrix factorization via a trade-off term to deal with the curse of data sparsity.


Recent studies have also shown that the customers’ review text can help improving the performance of recommender system, especially for those with few rating data. One of the reported studies applies topic modeling techniques on the review text to discover each individual’s latent opinion on each aspect as well as the relative emphasis on different aspects, achieving significant improvement comparing with models using only ratings or reviews. Further attempts have been made to combine customers’ review and products’ review simultaneously using a biased matrix factorization for rating prediction, and allows dynamic adjustment of the importance of latent factors. Moreover, sentiments of products have also been obtained from the review data, and has been used in recommender system as a particular quantification in per-aspect basis.