Embedding User Behavioral Aspect in TF-IDF Like Representation

Abstract

Term Frequency – Inverse Document Frequency (TF-IDF) computes weight for each word in a document which increases proportionally to the number of times the word appears in a specific document but is counterbalanced by the number of times it occurs in the collection of documents. TF-IDF is the state-of-the-art for computing relevancy scores between documents. However, it is based on statistical learning alone and doesn’t directly capture the conceptual contents of the text or the behavioral aspects of the writer. Hence, in this work we show how relatively low dimensional user behavioral vectors extracted from the same text, from which TF-IDF vectors are extracted, can be used to enrich the performance of TF-IDF. We extract User-Concerns embedded in user reviews and append them to TF-IDF vectors to train a deep rating prediction model. Our experiments show that adding such conceptual knowledge to TF-IDF vectors can significantly enhance the performance of TF-IDF vectors by only adding very little complexity.