Core Concepts
A simple and effective framework called DCA-SBRS that can be easily instantiated with existing representative accuracy-oriented session-based recommender systems to improve their diversity performance while preserving recommendation accuracy.
Abstract
The paper proposes a Diversified Category-aware Attentive framework called DCA-SBRS that can be used as a plugin to improve the diversity performance of existing accuracy-oriented session-based recommender systems (SBRSs) without significantly deteriorating their recommendation accuracy.
The key components of DCA-SBRS are:
Model-agnostic Diversity-oriented Loss (MDL) function: This loss function works with the accuracy-oriented loss (e.g., cross-entropy loss) to enhance the diversity of the recommended list by leveraging the items' category information and the estimated item scores from the given SBRS.
Non-invasive Category-aware Attention (NCA) mechanism: This mechanism utilizes the category information as directional guidance to replace the normal attention mechanism widely used in existing accuracy-oriented SBRSs. This helps maintain the recommendation accuracy while improving diversity.
Extensive experiments on three real-world datasets show that DCA-SBRS can help existing SOTA SBRSs achieve extraordinary performance in terms of diversity (e.g., 74.1% increase on ILD@10) and comprehensive performance (e.g., 52.3% lift on F-score@10), without significantly deteriorating recommendation accuracy compared to SOTA diversified SBRSs.
Stats
The average session length in the Diginetica, Retailrocket, and Tmall datasets is 4.85, 3.53, and 6.08, respectively.
The diversity score (DS) of the training set in the Diginetica, Retailrocket, and Tmall datasets is 0.3741, 0.4646, and 0.6575, respectively.
The repeat ratio (RR) of the training set in the Diginetica, Retailrocket, and Tmall datasets is 0.1301, 0.2488, and 0, respectively.