Generative Causality-Driven Network for Graph Multi-Task Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Causal Inference

Background:

• Multi-task learning (MTL) leverages shared knowledge to address data sparsity.
• Graph multi-task learning (GMTL) uses graph neural networks (GNNs) but relies on heuristics, leading to spurious correlations.
• Existing GMTL methods struggle with accurately identifying beneficial task relationships.

Purpose of the Study:

• To propose a novel framework, Generative Causality-driven Network (GCNet), for learning causal task structures.
• To improve generalization ability and model robustness in multi-task learning.
• To overcome the limitations of heuristic-based task graph construction in GMTL.

Main Methods:

• GCNet employs a feature-level generator to create structure priors.
• An output-level generator, modeled as a causal energy-based model (EBM), refines structures in the output space.
• Theoretical derivation of intervention contrastive estimation for efficient causal EBM training.

Main Results:

• GCNet effectively learns causal structures between tasks.
• The proposed causal framework enhances generalization and robustness.
• Experimental results show GCNet outperforms competitive MTL baselines on synthetic and real-world datasets.

Conclusions:

• GCNet offers a principled approach to learning task relationships for improved MTL.
• The causal framework addresses limitations of heuristic-based GMTL.
• GCNet demonstrates superior performance and robustness in multi-task learning scenarios.