Inductive Transfer Learning for Graph-Based Recommenders

Conference: NeurIPS 2025 arXiv: 2510.22799 Code: None Area: Audio & Speech Keywords: Graph Neural Networks, Transfer Learning, Recommender Systems, Inductive Inference, Zero-Shot Recommendation

TL;DR

This paper proposes NBF-Rec, a graph-based recommendation model built upon the Neural Bellman-Ford Network, which supports inductive transfer learning across datasets with completely disjoint users and items, enabling zero-shot cross-domain recommendation and lightweight fine-tuning adaptation.

Background & Motivation

Background: Graph neural network-based recommender systems (e.g., LightGCN) perform well within a single domain, but are primarily trained in a transductive manner and cannot generalize to new users, new items, or new datasets.

Limitations of Prior Work: - Existing cross-domain recommendation methods assume overlapping users or items between the source and target domains, limiting their applicability. - Methods based on adversarial training, contrastive disentanglement, and meta-learning still rely on aligned entity spaces or domain-specific supervision. - Large-scale pre-trained models (P5, GPTRec) require substantial pre-training and inference resources, and depend on textual or visual side information.

Key Challenge: Transfer learning has become standard practice in NLP and CV, yet it remains largely unexplored in graph-based recommendation—particularly in scenarios with completely disjoint users and items.

Goal: Enable inductive transfer learning across user-item graphs with fully disjoint entities, supporting both zero-shot recommendation and fine-tuning adaptation.

Key Insight: The path-aggregation message-passing mechanism of NBFNet is leveraged to dynamically compute node representations (rather than pre-learned embeddings), thereby achieving inductive generalization; edge feature encoding is integrated to enhance the capture of interaction-level information.

Core Idea: Rather than learning node-specific parameters, the model learns the message-passing process itself, enabling generalization to entirely unseen user-item graphs.

Method

Overall Architecture

NBF-Rec builds upon NBFNet (Neural Bellman-Ford Network) and frames recommendation as a link prediction task on a bipartite graph. Given a query user \(u\), the model dynamically computes representations for all nodes via multi-layer message passing and produces a score for each candidate item.

Key Designs

1. Query-Conditioned Initialization

   • Function: Initialize representations of all nodes conditioned on the query user.
   • Mechanism: \(h_v^{(0)} = \mathbf{1}(u = v)\), i.e., only the node corresponding to the query user is initialized to 1, while all others are set to 0. All information propagates outward from the query user.
   • Design Motivation: Ensures that representations are dynamically computed for each query without relying on pre-computed node embeddings.

2. Edge Feature Embedding

   • Function: Encode raw edge features (ratings, timestamps, categories, play counts, etc.) into embeddings suitable for message passing.
   • Mechanism: A two-level MLP structure: \(g(r) = \text{MLP}_{\text{emb}}(\text{MLP}_{\text{proj}}(r))\)
     • \(\text{MLP}_{\text{proj}}\): a dataset-specific projection MLP that handles heterogeneous edge features across different datasets
     • \(\text{MLP}_{\text{emb}}\): a shared backbone embedding MLP
   • Novelty: While the original NBFNet relies solely on graph structural information, NBF-Rec incorporates edge features to enable learning of richer interaction patterns.

3. Message Passing Mechanism

   • Function: At each layer \(t\), aggregate neighbor messages to update node representations.
   • Mechanism: \(M_v^{(t)} = \{\text{MESSAGE}(h_x^{(t-1)}, \mathbf{w}_q(x,r,v)) \mid (x,r,v) \in \mathcal{E}(v)\}\)
     • Edge weights: \(\mathbf{w}_q(x,r,v) = \text{MLP}_t(g(r))\), with a separate MLP per layer
     • Message function: non-parametric DistMult operation
     • Node update: aggregation (summation) + linear transformation + layer normalization + activation
     • Residual connection incorporating initial embeddings: \(\text{AGGREGATE}(M_v^{(t)} \cup \{h_v^{(0)}\})\)

4. Score Generation

   • Function: After \(T\) layers of message passing, compute a recommendation score for each node.
   • Mechanism: \(\text{score}(u,q,v) = \text{MLP}_{\text{score}}(\text{concat}(h_v^{(T)}, h_v^{(0)}))\)
   • The final-layer embedding and the initial embedding are concatenated and passed through an MLP to produce a scalar score.

5. Key to Inductive Generalization

   • The model learns no node-specific parameters; all parameters reside in the message-passing MLPs and aggregation operations.
   • Heterogeneous edge feature formats across datasets are handled via dataset-specific projection MLPs.
   • Representations are computed dynamically at inference time, requiring no pre-computed embeddings.

Loss & Training

• Cross-Entropy Loss: $\(\mathcal{L} = -\log p(u,q,v) - \sum_{i=1}^{n} \frac{1}{n} \log(1-p(u'_i, q, v'_i))\)$ where \((u,q,v)\) is a positive sample and \(\{(u'_i, q, v'_i)\}\) are strictly negative samples drawn via uniform random sampling (i.e., not present in the training set).
• Batch-Edge Removal During Training: Edges in the current batch are removed from the message-passing graph, forcing the model to rely on non-trivial paths rather than direct connections when learning relational patterns.
• Three Settings: end-to-end training / zero-shot transfer / fine-tuning.

Computational Complexity

The total forward-pass complexity is \(\mathcal{O}(T|E| + |V|)\), linear in the number of nodes and edges. Although inference overhead is higher than that of LightGCN (which pre-computes embeddings), the model supports inductive generalization.

Key Experimental Results

Datasets

Seven real-world recommendation datasets:

Dataset	#Users	#Items	#Interactions	Domain
ML-1M	5,950	2,811	364,654	Movies
LastFM	1,867	1,867	39,717	Music
Amazon B.	52,204	57,289	293,912	E-commerce
Gowalla	29,858	70,839	712,504	Location check-in
Epinions	21,008	13,887	266,791	Product reviews
BookX	12,720	18,318	276,334	Books
Yelp18	31,668	38,048	1,097,007	Local businesses

Main Results

Zero-Shot / Fine-Tuning / End-to-End Comparison

Pre-training sources: Amazon Beauty + Epinions

Setting	ML-1M	LastFM	Amazon B.	Gowalla	Epinions	BookX	Yelp18
Zero-shot	Competitive	Below par	—	Below par	—	Near baseline	Near baseline
Fine-tuning	Improved	Substantially improved	Improved	Substantially improved	Improved	Improved	Improved
End-to-end	Baseline	Baseline	Baseline	Baseline	Baseline	Baseline	Baseline

Key observations: - On ML-1M, BookX, and Yelp18, zero-shot performance falls within 5% of the end-to-end baseline. - Fine-tuning consistently improves performance across all datasets, with the most notable gains on LastFM and Gowalla.

Ablation Study

Cross-Dataset Transfer Heatmap

Key Finding	Description
Asymmetric transfer	Transfer from LastFM→ML-1M is effective, but not vice versa
Self-transfer not optimal	BookX and Amazon Fashion benefit more from transfer from other datasets than from themselves
Edge feature impact	Low-information edge features (BookX, LastFM) generally reduce transferability
Graph scale vs. features	Gowalla (large graph, sparse features) transfers better than Yelp (rich features but weak transferability)

NBF-Rec vs. NBFNet Comparison

• NBF-Rec consistently outperforms NBFNet (graph structure only) in zero-shot and fine-tuning settings.
• In end-to-end training, NBF-Rec performs on par with or slightly better than NBFNet.
• This confirms the contribution of edge feature embeddings to transfer learning.

Key Findings

• Inductive transfer learning is feasible for graph-based recommendation: Even when users and items are completely disjoint, NBF-Rec achieves meaningful zero-shot recommendation through the learned message-passing process.
• Pre-training on a different domain can outperform pre-training on the target domain: Cross-domain inductive biases are sometimes more effective than in-domain signals.
• Edge features are a double-edged sword: Rich edge features improve in-domain performance but do not necessarily enhance cross-domain transferability.
• Lightweight fine-tuning bridges the gap: A small amount of in-domain supervision is sufficient to bring the zero-shot model close to fully supervised performance.

Highlights & Insights

• This is the first demonstration of scalable inductive transfer across fully disjoint user-item graphs, representing an important milestone in graph-based recommendation research.
• Elegant design: Built as an extension of NBFNet, the core modification is the introduction of edge feature embeddings and a dataset-specific projection MLP, making the engineering implementation relatively lightweight.
• The discovery of asymmetric transferability is particularly intriguing and suggests that source domain selection is a research question worthy of deeper investigation.
• No reliance on textual or visual side information: The model operates purely on interaction graph structure and edge features, granting it broad applicability.

Limitations & Future Work

• Inference cost: Each query requires a full message-passing forward pass, resulting in higher inference cost than methods that pre-compute embeddings.
• Medium-scale datasets: The largest dataset, Yelp18, contains only approximately one million interactions; scalability to industrial-scale data remains unverified.
• Edge feature engineering: Different datasets require different feature preprocessing pipelines, and the dataset-specific projection MLPs add implementation complexity.
• No direct comparison with large-scale pre-trained recommendation models (P5, GPTRec) under identical conditions.
• The negative sampling strategy is relatively simple (uniform random); more advanced strategies could potentially yield further gains.
• Future directions include multi-graph joint pre-training, evaluation on larger-scale datasets, and more efficient inference strategies.

Related Work & Insights

• NBFNet: The Neural Bellman-Ford Network was originally developed for knowledge graph link prediction; this paper successfully transfers the framework to recommender systems.
• ULTRA: Zero-shot transfer for knowledge graph completion, representing the closest methodological relative.
• LightGCN: A representative transductive method, serving as the primary comparison baseline.
• Insight: The message-passing mechanism itself can serve as transferable "knowledge" across domains, as opposed to node embeddings. This observation may generalize to other graph learning tasks.

Rating

• Novelty: ⭐⭐⭐⭐ First demonstration of scalable inductive recommendation transfer across purely interaction-based graphs with disjoint entities; the problem formulation is novel
• Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive evaluation across 7 datasets, three settings, and cross-domain transfer heatmap analysis
• Writing Quality: ⭐⭐⭐⭐ Clear methodology, well-specified equations, and well-designed experiments
• Value: ⭐⭐⭐⭐ Opens a new direction for transfer learning in graph-based recommendation; the lightweight design has practical potential

Related Papers

• [NeurIPS 2025] Sound Logical Explanations for Mean Aggregation Graph Neural Networks
• [NeurIPS 2025] Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation
• [ICLR 2026] AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer
• [ACL 2026] Learning Invariant Modality Representation for Robust Multimodal Learning from a Causal Inference Perspective
• [ICCV 2025] Learning to See Inside Opaque Liquid Containers using Speckle Vibrometry