As the size of the market in online social media and e-commerce has grown dramatically, the number of online commercial contents has increased at an exponential rate, and consequently, the recommender systems have been used to assist customers to explore their preferred products among lots of products in an efficient way. Furthermore, deep neural networks and an increase of training datasets have improved the accuracy of sequential recommendation approaches which take into account the sequential patterns of user logs, e.g., purchase histories of a user.
However, incorporating only the individual's recent logs may not be sufficient in properly reflecting global preferences and trends across all users and items. In response, we propose a novel self-attentive sequential recommender system with topic modeling-based category embedding as a novel approach to exploit global information in the process of sequential recommendation.
The self-attention module in our model effectively leverages the sequential patterns from the user's recent history. In addition, our novel category embedding approach, combined with the categorical probabilities calculated by topic modeling, effectively captures global information that the user generally prefers. We further propose a categorical preference gate to efficiently combine the sequential information and the global information for the given user's previous purchased logs or the user's general preference. Furthermore, as the diverse recommendation has become important for real-world applications, our model also incorporates a linearly-transformed noise vector obtained by random sampling so as to make users explore various items as well as to prevent loss in user information and the overfitting problem.
Experimental studies using public datasets, MovieLens and Gowalla, show that our model outperforms other state-of-the-art sequential recommendation models, and the results of the ablation test show that each module in our model plays important rule for improving the model performance. Additional qualitative experiments show that the proposed category embedding, combined with the proposed categorical preference gate, effectively provides global preference information. In addition, the linearly-transformed random noise vector added to the user embedding vector for personal recommendation effectively provides the diverse recommendation without loss in the user's general preference information, and the model further provides options to users for directly controlling the trade-off between their own preferences and exploring various items.