검색 상세

Evidential Meta-Learning for Molecular Property Prediction

  • 함경표 (아주대학교 일반대학원)

초록/요약

The usefulness of supervised molecular property prediction is well-recognized in many applications. However, inadequacy and imbalance of labeled data make the learning problem difficult. Moreover, the reliability of the predictions is also a huddle in the distribution of supervised molecular property prediction in the cost and safety-critical application fields, such as drug discovery. We propose an EM3P2 model as a supervised molecular property prediction method that addresses the problem of data insufficiency and reliability. Our proposed method trains an evidential graph isomorphism network classifier using multi-task molecular property datasets on top of a model-agnostic meta-learning (MAML) scheme. Our model is a well-orchestrated combination of evidential neural networks for estimating model prediction uncertainty, graph isomorphism networks for embedding vector input molecular graphs, and data balance-aware model-agnostic meta-learning for generating a meta-model that adapts to new tasks with little labeled data.

more

목차

1 Introduction 1
2 Method 4
2.1 Evidential Meta Learning Framework 5
2.1.1 EM3P2 Meta Training 6
2.1.2 EM3P2 Meta Testing 9
2.2 Graph Embedding and Classification Model 10
2.2.1 Graph Isomorphism Network 10
2.2.2 Evidential Multi-layer Perceptron(EMLP) 10
2.2.3 Estimating Uncertainty 11
2.3 Loss Function 12
2.3.1 Evidential Loss Function 12
2.3.2 Belief Regularizer 13
2.3.3 Accuracy Versus Uncertainty Loss Calibration 13
2.3.4 Overall Loss Function 14
3 Experiment 15
3.1 Dataset 15
3.2 Data Summary 16
3.2.1 Tox21 16
3.2.2 Sider 17
3.3 Compared Methods 19
3.3.1 Reproducibility Setting 19
3.3.2 Evaluation 20
4 Results 23
4.1 Comparison with State-of-the-art 23
4.2 Case Study of accuracy according to vacuity threshold 29
4.3 Case Study of Query balancing 31
4.4 Case Study of Belief Quantification with Softmax 32
5 Conclusion 35

more
반출 Meta View 목록