Effects of sparse datasets on time interval-aware self-attention sequential recommendation models

Abstract

Recommendation models serve as crucial filters in managing information, yet they face a few crucial challenges, such as capturing user-item interaction behaviors in sparse datasets. Data sparsity refers to an issue where there is a lack of interactions or missing values in the recommendation dataset. A sparse dataset with a massive number of missing values and interactions leads to more dynamic user behaviors, which suffers a poor recommendation quality. The self-attention mechanism from Transformer can alleviate the effects of data sparsity in datasets by assigning weights to items of interaction behaviors. This allows the model to capture the user dependencies in complex user behavior, which is beneficial for sparse datasets with patterns that are not immediately apparent. This approach has shown its capability to handle large and sparse datasets, as seen in time interval-aware self-attention sequential recommendation model (TiSASRec). It utilized the self-attention mechanism, considering the timestamp and absolute positions of items to estimate the higher attention weights to show the importance of recent items. Thus, this study aims to investigate the effects of sparse datasets by comparing the performance of TiSASRec model with self-attention based sequential recommendation model (SASRec), which excludes time interval-awareness.