Hybrid collaborative filtering for high-involvement products: A solution to opinion sparsity and dynamics

High capital value goods that are purchased only after long and careful consideration, such as a car, truck, appliance (called high-involvement products) are increasingly being purchased online. For these products accurate recommendations are very important. Collaborative filtering (CF) is a commonly used approach for recommending products based on known preferences of similar users. Two challenges that limit the performance of CF in high-involvement products are the ratings' sparsity and dynamics. We use online reviews as an auxiliary information source and design a hybrid CF method to address the sparsity problem. Additionally, we consider the dynamics that may exist in online ratings and reviews. Specifically, we first investigate empirically the evolution of the ratings and reviews over time and sequence. The results show that there are both temporal and sequential dynamics in online ratings, and only temporal dynamics in online reviews. Next, considering these dynamics, we develop techniques to predict missing ratings based on online reviews and product similarities. The sparsity of the User-Item rating matrix is alleviated by filling these predicted ratings. The experiment, based on real-world datasets, demonstrates the superior performance of our recommendation approach for high-involvement products.