Source-linked AI summary
Graph Neural Team Recommendation: An Integrated Approach
Md Jamil Ahmed, Mahdis Saeedi, Hossein Fani
TL;DR
Existing neural team-recommendation methods overlook experts’ multi-hop relational structure and use disjoint skill pretraining. This paper reformulates recommendation as end-to-end graph link prediction, achieving state-of-the-art performance across two large-scale datasets.
Problem
Existing neural methods overlook multi-hop relational and structural information among experts and collaborations, while skill representations are pretrained separately from team recommendation.
Method
The approach represents successful teams as graph star subgraphs and learns end-to-end team–expert link predictions with graph neural networks.
Results
The end-to-end approach consistently and substantially outperforms transfer-based methods across graph neural networks, datasets, and metrics, establishing state-of-the-art performance.
Takeaways & Limitations
Directly modeling intra-team and cross-team dependencies removes disjoint training stages and uses graph structure for team recommendation.
Takeaways & Limitations
The prior skill representation disregards experts and their multi-hop collaborations and remains disjoint from the main recommendation task.
Abstract
from arXiv · showhide
Team recommendation aims to select an optimal subset of experts who can form an almost surely successful collaborative team for a given set of required skills. State-of-the-art methods are neural multi-label classifiers that transfer dense vector representations of skills into a sparse occurrence vector representing the optimal subset of experts. Such methods, however, overlook experts' relational and structural information encoded in the expert collaboration graph and, thus, fall short of capturing complex inter-dependencies among experts and their associated skills within teams. Moreover, the skills' dense vectors are pretrained disjointly and independently of the underlying neural classifier, hence, preventing end-to-end optimization. In this paper, we propose to reformulate the team recommendation problem into end-to-end link predictions in the expert collaboration graph to consume multi-hop intra-team and cross-team collaborations among experts while eschewing the unnecessary complexities of the disjoint two-phase training procedure. Our experiments on two large-scale datasets from various domains with distinct distributions of skills in teams demonstrate the superiority of the end-to-end approach and establish a new state of the art. Our code is available at https://github.com/fani-lab/OpeNTF.
1 Introduction
Team recommendation seeks successful expert teams for required skills, but existing neural classifiers overlook collaboration-graph structure and train skill representations separately. The paper proposes an end-to-end graph approach that captures multi-hop relations and reports state-of-the-art performance across datasets and graph networks.
- Team recommendation selects experts whose combined skills can solve difficult tasks across healthcare, peer review, and education.
- The proposed method models collaborations within and across teams in a graph and reformulates recommendation as end-to-end links between required skills and optimal experts.
- Existing neural methods transfer dense skill representations into sparse expert-selection vectors but overlook experts’ multi-hop relational and structural information.
- The approach captures multi-hop information for skills, experts, and teams directly in the expert collaboration graph.
- Experiments cover 3 graph structures, 6 graph neural networks, and 2 large-scale datasets with varied team-skill distributions.
2 Problem Definition
A team is a subset of experts assigned to cover required skills, with a binary success status. The recommender learns to map required skills to an expert subset associated with successful teams.
- A team combines an expert subset with a skill subset and is labeled successful or unsuccessful.
- The recommendation objective identifies experts that cover the required skills while avoiding subsets associated with unsuccessful teams.
- The learned mapping f_θ(s) = e outputs an expert subset for every successful training team.
3 Preliminaries
Prior neural recommenders encode skills in a heterogeneous skill-team-expert graph and feed aggregated skill vectors to a multi-label classifier. This representation ignores expert collaboration structure and is trained separately from recommendation.
- State-of-the-art methods map a dense skill vector to a multihot vector representing the optimal expert subset with a neural multi-label classifier.
- Their heterogeneous graph contains skill, successful-team, and expert nodes, with each successful team represented as a star connecting its skills and experts.
- Skill vectors are summed from constituent skill-node embeddings before being passed to the primary recommender.
- The pretrained skill encoder ignores experts’ multi-hop collaborations, remains disjoint from recommendation learning, and treats training teams independently.
4 Proposed Approach
The proposed approach replaces the separate skill-pretraining and multi-label-classification pipeline with graph-based link prediction trained directly on successful team-expert links. Message passing captures heterogeneous graph structure, and top-k predicted links form the recommended team.
- The method estimates recommendations directly through team-expert link prediction, eliminating the separate classifier parameters θ.
- Graph neural network node representations are computed by aggregating neighboring vectors through learnable message and update functions.
- The sole parameter set is learned by minimizing link-prediction loss over observed team-expert links from successful teams.
- The sigmoid link predictor uses heterogeneous relational and structural information to supervise team-expert predictions.
- At inference, experts with the top-k predicted probabilities are selected as the recommended team of size k.
- The end-to-end formulation captures intra-team and cross-team dependencies while avoiding separate pretraining and fine-tuning stages.
5 Experiments
The experiments compare end-to-end and transfer-based team recommendation across two heterogeneous datasets, graph structures, neural architectures, and evaluation settings. The benchmark design masks team–expert links and controls training conditions to support comparison.
- Experimental questions: The study evaluates whether end-to-end recommendation is superior across datasets, graph structures, and graph neural networks.The research questions cover comparative performance, cross-domain consistency, graph-structure effects, and model rankings.
- Datasets: The dblp dataset represents papers as successful teams with authors as experts and keywords as skills, while imdb represents movies with cast and crew as experts and genres as skills.Teams with fewer than three expert members were removed while preserving statistical distributions.
- Datasets: Both datasets have long-tailed distributions in which a few experts contribute to many teams while most experts participate sparsely.
- Baselines: The comparison includes end-to-end models and transfer-based baselines using metapath2vec or deep graph infomax skill embeddings with neural classifiers.The transfer-based baselines include variational Bayesian and non-variational classifiers.
- Models and graph settings: The evaluation includes strong heterogeneous-graph neural networks, including GraphSAGE and GIN, alongside attention-based architectures.The study also examines neighborhood sampling and multiple graph structures.
- Evaluation protocol: Testing masks expert members from 15% of randomly shuffled teams, while three-fold cross-validation masks team–expert links in the remaining 85%.The end-to-end models use the transfer-based baselines’ learning rate, epochs, and batch size for a controlled comparison.
6 Results
Across datasets, graph neural networks, and metrics, the end-to-end approach consistently and substantially outperforms transfer-based methods. Results also show that graph structure and attention-based architectures strongly influence performance.
- End-to-end versus transfer-based approaches: End-to-end models consistently and substantially outperform transfer-based methods across datasets, graph neural networks, and metrics.The reported advantage is attributed to combining multi-hop relational and structural information with supervised information about optimal experts.
- Cross-domain consistency: End-to-end performance improvements are consistent across both datasets, whereas transfer-based performance varies across classifier types, graph networks, and domains.
- Graph structures: End-to-end performance remains closely aligned with and without location links, with marginal differences except for HAN.HAN is no longer the best when location links are included because its meta-path-based attention mechanism overlooks those links.
- Graph structures: Neither approach performs comparably on the skill–expert bipartite graph, where removing team nodes creates a highly dense structure without contextual separation.The result highlights the importance of graph structure for effective message passing.
- Model performance: HAN outperforms all methods across datasets and metrics in the end-to-end setting, with GAT and GATv2 the closest runners-up in dblp.Attention-based models generally dominate, while GIN typically trails because of hyperparameter sensitivity.
7 Concluding Remarks and Future Work
The paper concludes that its end-to-end graph neural network achieves state-of-the-art team recommendation performance across two large-scale datasets, while identifying cold-start and team refinement as future directions.
- The end-to-end graph neural network achieves state-of-the-art performance on two large-scale datasets with distinct team-skill distributions.
- Future work will address cold-start recommendations for new experts and emerging skills with limited historical data.
- Future work will also study replacing experts in existing teams to maintain or improve performance.
8 GenAI Usage Disclosure
The authors disclose using GenAI tools during manuscript preparation and benchmark development to improve writing, support scripting, understand graph neural networks, and organize experimental results.
- GenAI tools were used to rephrase text and improve sentence structure, grammar, fluency, clarity, and coherence without changing technical content.
- GenAI tools helped develop scripts for parallel benchmark execution and support understanding of graph neural network characteristics.
- GenAI tools assisted with collecting bulk experimental results and establishing a reproducible pipeline.