SR-PredictAO: Session-based Recommendation with High-Capability Predictor Add-On

TL;DR

This work tackles session-based next-item prediction by identifying a bottleneck in predictor capacity within encoder–predictor models and proposing a general augmentation framework, SR-PredictAO. The approach adds a high-capability predictor, instantiated as Neural Decision Forest (NDF-SR), and a Random User Behavior Alleviator within the predictor path, along with a Merger to fuse predictions from both the base and augmented paths. Empirical results on YooChoose and Diginetica show consistent gains across three strong base models, achieving up to 2.9% improvements in HR@20 and 2.3% in MRR@20 with statistical significance. The framework is end-to-end trainable and model-agnostic within the encoder–predictor paradigm, offering a practical pathway to boost SR systems without heavy redesign of encoders.

Abstract

Session-based recommendation, aiming at making the prediction of the user's next item click based on the information in a single session only, even in the presence of some random user's behavior, is a complex problem. This complex problem requires a high-capability model of predicting the user's next action. Most (if not all) existing models follow the encoder-predictor paradigm where all studies focus on how to optimize the encoder module extensively in the paradigm, but they overlook how to optimize the predictor module. In this paper, we discover the critical issue of the low-capability predictor module among existing models. Motivated by this, we propose a novel framework called *Session-based Recommendation with Predictor Add-On* (SR-PredictAO). In this framework, we propose a high-capability predictor module which could alleviate the effect of random user's behavior for prediction. It is worth mentioning that this framework could be applied to any existing models, which could give opportunities for further optimizing the framework. Extensive experiments on two real-world benchmark datasets for three state-of-the-art models show that *SR-PredictAO* out-performs the current state-of-the-art model by up to 2.9% in HR@20 and 2.3% in MRR@20. More importantly, the improvement is consistent across almost all the existing models on all datasets, and is statistically significant, which could be regarded as a significant contribution in the field.

Figures (4)

• Figure 1: (a) The overview of the base model, (b) Framework SR-PredictAO; Given an input session $S$, the encoder module generates the latent variable $\bm{z}$. In (a), $\bm{z}$ is passed to the base model predictor module to obtain the predicted probability distribution over all items. In (b), $\bm{z}$ is passed to both the base model predictor module and the new predictor module (called NDF-SR) to obtain two predicted probability distributions over all items. Then, module Merger combines the two distributions to output the final distribution.
• Figure 2: The overview of the NDT, decision function gives the split score for root and internal nodes, and the leaves nodes' result is the probability of the session reaching the node
• Figure 3: The hyper-parameter study results of SR-PredAO(SGNN-HN)
• Figure 4: DoF of NDF-SR with different depth and pruning rates (r), dotted lines represent DoF of linear model

Theorems & Definitions (1)

• Lemma 4.1