Abstract:
Recommender systems have a major contribution, that is, it allows users to interact with content efficiently. A recommender system advises users by filtering items based on users’ previous actions. Collaborative filtering (CF) is one of the recommender system algorithms which is built on explicit feedback (e.g., user ratings) and implicit feedback (e.g., number of clicks and purchases). It compares a target user with others who have similar preferences. Further, there are two well known types of CF: user-based CF and item-based CF. User-based CF assumes that people who have similar tastes tend to react to items similarly. For item-based CF, it tries to find look-alike items instead of look-alike users. Nowadays, many research attempt to apply the neural network into CF because there is a limitation in CF that CF can learn only linear representation, but the neural network can learn both linear and non-linear representation. Autoencoder reconstructs the input data in the output layer by encoding the input data into a low dimensional middle layer called the hidden layer to form latent representation, and then the output from the hidden layer is decoded by the output layer to reconstruct the data. From previous works, we have noticed that relations between a target user (item) and other users (other items) were not utilized and this relationship is a plus point for collaborative filtering technique. Therefore, we have proposed the autoencoder recommender system model that learns a representation of similarity between a target user (item) and other users (items). Finally, the experimental results have shown that the proposed model performs better than state of the art methods.
