Category-Aware Sequential Recommendation with Time Intervals of Purchases

The goal of a sequential recommendation system is to predict the next item a user is likely to purchase based on their buying history. Previous research has considered the time intervals between purchases by analyzing patterns in the items, but have neglected the important information at the category level. To overcome this shortcoming, this paper presents two category-aware sequential recommendation models which effectively integrate category information into the user's purchase sequence representation. The first model fuses item embedding with the corresponding category embedding, thus directly infusing category-specific details into the representation of purchasing history, thereby enriching the insight into user behavior. On the other hand, the dual model employs a specialized sub-network to identify patterns within item categories, and this category-level representation indirectly influences the item-level representation of user behavior through an attention mechanism. The results of experiments on Amazon datasets reveal that the inclusion of category data notably improves the hit ratio in sequential recommendation. The proposed models outperform the baseline model particularly in situations involving shorter user sequences. Further, merging purchase records from multiple product datasets across different categories during the training phases leads to even more substantial improvements in the hit ratios.