RAVI K YADAV

This page is a survey of existing deep learning based SBR systems.

A traditional system utilizes user profiles and historical actions to recommend user’s next interaction, whereas SBRs utilizes limited (short) and dynamic sessions. For latest mobile stream medias platforms, where traditional collaborative signals are limited, SBRs can predict next best action for user based on sessions with limited interaction information.

Most of the recent development in SBR focuses on leveraging Graph Neural Networks (GNNs) to model the user sessions and actions. It is found that exponential growth in the model complexity may not result in the significant performance gains on benchmark scores. Some of the areas to focus while building GNN-based models are, GNN propagation, inductive bias, popularity bias, position information, target information, global context, and memory optimization. Session graphs are usually very sparse. Paper here proposes an update to the read out module that takes on more responsibility in model reasoning process, while relaxing the requirement for GNN propagation.

Instance-view readout module generates the overall preference with attention mechanism related to single item (either every item or last clicked item). Without the reasoning process by GNN propagation, instance-view readout lack information about higher-level connections between items/nodes. the paper proposes to combine high-level-view with instance-view readout to achieve multi-level reasoning over item/node transitions. Since building such high-level concepts for real world use-cases is infeasible, the authors recommends incorporating inductive biases to prune the search space.

To get more updates on data science and machine learning –