User-item recommendation is a central challenge for search engines, social media platforms, and streaming services, due to the need to model both relational structures and temporal dynamics. Many existing solutions address these two aspects separately, limiting their ability to fully capture user behavior. In this work, we attempt to bridge that gap by evaluating Lightweight Memory Networks (LiMNet), a model designed to preserve causal relationships within sequences of temporal interactions. To assess its potential, we developed a benchmarking framework for user-item interaction prediction. We compared LiMNet against Jodie, a state-of-the-art baseline, across three real-world datasets: Wikipedia page edits, Reddit post submissions, and LastFM music streams. These datasets vary in scale and interaction patterns, providing a comprehensive testbed. Our experiments reveal two key findings. First, in its simplified form, LiMNet consistently underperforms compared to Jodie in predictive accuracy. Second, normalizing embeddings within the cross-RNN mechanism significantly improves LiMNet’s performance. These results highlight both the limitations and strengths of LiMNet in interaction prediction, offering insights into architectural choices that influence performance and suggesting avenues for future research in temporal recommendation systems.
Titouan Jean Mazier (Wed,) studied this question.