Embedding unstructured side information in product recommendation

Various researchers have already engaged in using auxiliary side information within recommender applications to improve the quality and accuracy of recommendations. This side information has either been in the form of structured information such as product specifications and user demographic information or unstructured information such as product reviews. The abundance of unstructured information compared to structured information entices the use of such unstructured information in the recommendation process. Existing works that employ unstructured content have been confined to standard text modeling technique such as the use of frequency measures or topic modeling techniques. In this paper, we propose to model unstructured content about both products and users through the exploitation of word embedding techniques. More specifically, we propose to learn both user and product representations from any type of unstructured textual contents available in different external information sources using recurrent neural networks. We then apply our learnt product and user representations on two recommendation frameworks based on matrix factorization and link prediction to enhance the recommendation task. Experimental results on four datasets constructed from the Rotten Tomatoes website (movie review aggregator database) have shown the effectiveness of our proposed approach in different real-world situations compared to the state of the art.