Bridging Datasets and Hyperparameters: GCN-Based Link Prediction for Recommendation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Hyperparameter recommendation through meta-learning (HPR-MtL) is effective but existing methods overlook within-dataset and within-hyperparameter interactions.
• Current HPR-MtL approaches often use KNN, linear regression, or collaborative filtering, focusing mainly on dataset-hyperparameter performance observations.

Purpose of the Study:

• To propose a novel hyperparameter recommendation framework that captures both homogeneous and heterogeneous interactions.
• To enhance the accuracy of hyperparameter recommendations by modeling similarity relationships within datasets and hyperparameter configurations.

Main Methods:

• Formulated hyperparameter recommendation as a link prediction problem on a bipartite graph, representing datasets and hyperparameters as nodes.
• Introduced a graph convolutional network (GCN) to simultaneously capture interactions within datasets, within hyperparameters, and between datasets and hyperparameters.
• Evaluated the framework on ResNet and Vision Transformer (ViT) models across 105 real-world classification datasets.

Main Results:

• The proposed GCN-based framework demonstrated superior performance compared to state-of-the-art hyperparameter recommendation baselines.
• Experiments showed significant improvements in recommendation accuracy across multiple evaluation metrics.
• The method effectively captures complex interactions crucial for accurate hyperparameter selection.

Conclusions:

• The novel GCN framework offers a more comprehensive approach to hyperparameter recommendation by integrating diverse interaction types.
• This method advances the field of meta-learning for automated machine learning (AutoML).
• The findings suggest a promising direction for future research in intelligent hyperparameter optimization.