Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Inductive Transfer Learning for Graph-Based Recommenders

Florian Grötschla, Elia Trachsel, Luca A. Lanzendörfer, Roger Wattenhofer

TL;DR

The paper tackles the limitation of transductive graph-based recommender systems by enabling inductive transfer across disjoint user-item graphs. It introduces NBF-Rec, which performs query-time dynamic message passing with edge features, deriving node representations without target-domain training. The approach demonstrates competitive zero-shot performance across seven real-world datasets and yields further gains through lightweight fine-tuning, while enabling transfer analyses across domains. This work broadens the applicability of graph-based recommendations by reducing the need for retraining when deploying across new domains or platforms.

Abstract

Graph-based recommender systems are commonly trained in transductive settings, which limits their applicability to new users, items, or datasets. We propose NBF-Rec, a graph-based recommendation model that supports inductive transfer learning across datasets with disjoint user and item sets. Unlike conventional embedding-based methods that require retraining for each domain, NBF-Rec computes node embeddings dynamically at inference time. We evaluate the method on seven real-world datasets spanning movies, music, e-commerce, and location check-ins. NBF-Rec achieves competitive performance in zero-shot settings, where no target domain data is used for training, and demonstrates further improvements through lightweight fine-tuning. These results show that inductive transfer is feasible in graph-based recommendation and that interaction-level message passing supports generalization across datasets without requiring aligned users or items.

Inductive Transfer Learning for Graph-Based Recommenders

TL;DR

The paper tackles the limitation of transductive graph-based recommender systems by enabling inductive transfer across disjoint user-item graphs. It introduces NBF-Rec, which performs query-time dynamic message passing with edge features, deriving node representations without target-domain training. The approach demonstrates competitive zero-shot performance across seven real-world datasets and yields further gains through lightweight fine-tuning, while enabling transfer analyses across domains. This work broadens the applicability of graph-based recommendations by reducing the need for retraining when deploying across new domains or platforms.

Abstract

Graph-based recommender systems are commonly trained in transductive settings, which limits their applicability to new users, items, or datasets. We propose NBF-Rec, a graph-based recommendation model that supports inductive transfer learning across datasets with disjoint user and item sets. Unlike conventional embedding-based methods that require retraining for each domain, NBF-Rec computes node embeddings dynamically at inference time. We evaluate the method on seven real-world datasets spanning movies, music, e-commerce, and location check-ins. NBF-Rec achieves competitive performance in zero-shot settings, where no target domain data is used for training, and demonstrates further improvements through lightweight fine-tuning. These results show that inductive transfer is feasible in graph-based recommendation and that interaction-level message passing supports generalization across datasets without requiring aligned users or items.
Paper Structure (19 sections, 6 equations, 4 figures, 1 table)

This paper contains 19 sections, 6 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Formulation
  5. NBF-Rec Model Architecture
  6. Initialization.
  7. Message Computation.
  8. Node Update.
  9. Score Generation.
  10. Training
  11. Computational Complexity
  12. Experimental Evaluation
  13. Datasets
  14. Experimental Setup
  15. Results and Analysis
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Schematic overview of NBF-Rec’s inductive transfer learning pipeline. The model is trained on user-item interaction graphs from one or more source domains and applied to unseen target domains. Transferable patterns are learned from structural and interaction features without requiring shared users or items.
  • Figure 2: Performance of NBF-Rec in zero-shot, fine-tuned, and end-to-end settings across seven datasets. Metrics are Hits@10 or NDCG@20, depending on the dataset (see Table \ref{['tab:datasets']}). Bars indicate 95% confidence intervals over 3 runs. Zero-shot transfer achieves competitive results.
  • Figure 3: Cross-dataset transfer heatmap: each column indicates the training dataset and each row the test dataset. Cell colors represent relative performance (normalized by row). NBF-Rec achieves strong transfer when trained on Amazon Beauty, outperforming several dataset-specific models.
  • Figure 4: Performance comparison between NBF-Rec and NBFNet across all datasets under end-to-end, zero-shot, and fine-tuned settings. Incorporating interaction-level features and inductive training enables consistent improvements in recommendation quality.