2023-Temporal Neural Team Formation with Negative Sampling

[thangk]

Link: https://scholar.uwindsor.ca/etd/8949

Main problem

Previous team formation approaches suffer from two major limitations: (1) they neglect the temporal nature of human collaborations where experts' skills, interests, and collaborative ties change over time, and (2) neural models are prone to overfitting when training data exhibits the long-tail phenomenon (few experts have many successful collaborations while most participate sparingly). Traditional approaches treat team formation as a static problem using bag-of-teams training with i.i.d assumptions, failing to capture how experts evolve and adapt over time. Additionally, most approaches only consider successful teams while ignoring unsuccessful examples that could provide valuable learning signals.

Proposed method

The authors propose a temporal neural team formation approach with three key innovations: (1) Streaming training strategy that processes teams chronologically rather than randomly shuffled, allowing models to learn temporal evolution of experts' skills and collaborative patterns, (2) Three negative sampling heuristics (uniform, unigram, and smoothed unigram) to incorporate virtually unsuccessful teams using closed-world assumption, and (3) Temporal skill modeling to capture how expertise evolves over time. The streaming approach enables neural models to understand trajectory changes and predict future optimal team positions in latent space by learning from ordered collaboration sequences.

My Summary

This work represents the first comprehensive investigation of temporality in neural team formation, addressing a critical gap in existing literature. The streaming training strategy is particularly innovative as it allows models to capture the dynamic nature of expert collaborations and skill evolution over time. The empirical results across four diverse datasets (DBLP, IMDB, USPT, GitHub) demonstrate that incorporating temporal aspects and negative sampling significantly improves team formation performance.

The results show that Bayesian neural models with negative sampling consistently achieve superior performance, while streaming training with temporal skills outperforms traditional shuffled training across all datasets and evaluation metrics (MAP, NDCG, Precision, Recall, ROC-AUC).

The negative sampling heuristics effectively address the long-tail distribution problem in collaboration data, with smoothed unigram distribution showing particular promise for handling sparse training minibatches. This temporal approach opens new directions for modeling the dynamic nature of human expertise and collaboration patterns.

Datasets

DBLP.v12: 99,375 teams, 14,214 authors, 29,661 skills (1979-2018)
IMDB: Movie teams with cast/crew as experts, genres as skills (1914-2020)
USPT: Patent teams with inventors as experts, subcategories as skills (1976-2019)
GITH: Repository teams with contributors as experts, languages as skills (2008-2022)