Chen et al. demonstrate a new algorithm that makes recommendation systems more efficient and effective, especially for systems that keep track of user behaviors such as those found on shopping websites. The proposed algorithm makes good use of the history records left by users. Unlike currently known systems, the proposed system selects only the relevant items for the recommendation rather than using the user’s entire shopping history, which makes the system more effective and efficient.
The newly proposed memory-augmented neural network (MANN) system makes use of two types of known successful models: the sequential recommendation model and memory-augmented neural networks. The sequential recommendation model “embeds the transition information between adjacent behaviors into the item latent factors for recommendation.” The model is effective in working with “local sequential patterns between every two adjacent records.” Memory-augmented neural networks can store historical hidden states. MANN takes advantage of both models, allowing the model to capture the essence of user behaviors on selected items. It thus can better predict user behavior based on past records.
The authors test their model in a large real-world dataset from Amazon, which includes about 7000 customers who purchased about 67000 items in four categories. Results show improved performance over other state-of-the-art models.
The model proposed by the authors is novel. It can be used in applications where the historical behaviors of users are stored.