The Limits of Graph Samplers for Training Inductive Recommender Systems: Extended results

TL;DR

The paper tackles the high training cost of graph-based inductive recommender systems by systematically evaluating pre-training on subsampled graphs. It introduces multiple graph-sampling techniques and applies three inductive models (PinSAGE, GInRec, INMO) to three real-world datasets, measuring ranking performance and training time. Key findings show that training on about 50% of the data can maintain performance while substantially reducing training time (up to ~86%), but aggressive reductions degrade accuracy, with temporal sampling and user-focused strategies often performing best; INMO demonstrates exceptional robustness to downsampling, offering large time savings. The work highlights limitations of current sampling methods for heterogeneous graphs and points to future directions for designing temporal-aware, structure-aware samplers and more robust inductive methods.

Abstract

Inductive Recommender Systems are capable of recommending for new users and with new items thus avoiding the need to retrain after new data reaches the system. However, these methods are still trained on all the data available, requiring multiple days to train a single model, without counting hyperparameter tuning. In this work we focus on graph-based recommender systems, i.e., systems that model the data as a heterogeneous network. In other applications, graph sampling allows to study a subgraph and generalize the findings to the original graph. Thus, we investigate the applicability of sampling techniques for this task. We test on three real world datasets, with three state-of-the-art inductive methods, and using six different sampling methods. We find that its possible to maintain performance using only 50% of the training data with up to 86% percent decrease in training time; however, using less training data leads to far worse performance. Further, we find that when it comes to data for recommendations, graph sampling should also account for the temporal dimension. Therefore, we find that if higher data reduction is needed, new graph based sampling techniques should be studied and new inductive methods should be designed.

Figures (7)

• Figure 1: Sampling example with illustration of sampling graph and the correlation between CG samples and KG samples.
• Figure 2: #Ratings on items.
• Figure 3: Kolmogorov-Smirnov D-statistic of degree distribution per node type.
• Figure 4: Illustration of user ratio and rating skew in the YD.
• Figure 5: INMO's relative performance.