Time-Split Cross-Validation as a Method for Estimating the Goodness of Prospective Prediction

By Robert P. Sheridan. 2013

305 citations recorded by the source corpus.

Subfields: Computational Theory and Mathematics, Environmental Chemistry, Materials Chemistry, Statistics and Probability, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Theory and Mathematics

1. Gisbert Schneider
2. Tommi S. Jaakkola
3. Regina Barzilay
4. Francesca Grisoni
5. Jure Leskovec
6. Wengong Jin
7. Marinka Zitnik
8. José Jiménez-Luna

In Environmental Chemistry

1. Connor W. Coley
2. Gisbert Schneider
3. Deng Cai
4. Jimeng Sun
5. Regina Barzilay
6. Mike Preuss
7. Junying Li
8. Wengong Jin

In Materials Chemistry

1. Gisbert Schneider
2. Tommi S. Jaakkola
3. Regina Barzilay
4. Francesca Grisoni
5. Jure Leskovec
6. Wengong Jin
7. Connor W. Coley
8. José Jiménez-Luna

In Statistics and Probability

1. Deng Cai
2. Xiaofei He
3. Connor W. Coley
4. Gisbert Schneider
5. Percy Liang
6. Yue Zhao
7. Jure Leskovec
8. Jimeng Sun

In Evaluation & Benchmarks

1. Deng Cai
2. Jure Leskovec
3. Connor W. Coley
4. Kevin Yang
5. Gisbert Schneider
6. Percy Liang
7. Yue Zhao
8. Cao Xiao

Citations in the corpus

• Planning chemical syntheses with deep neural networks and symbolic AI (2017, background): “[37] Table 1 shows the metrics for the expansion policy.”
• Strategies for Pre-training Graph Neural Networks (2019, neutral): “Prior work has shown that this scaffold split provides a more realistic estimate of model performance in prospective evaluation compared to random split (Chen et al., 2012; Sheridan, 2013).”
• Analyzing Learned Molecular Representations for Property Prediction (2019, positive): “Overall, our results confirm the findings of Sheridan13 that scaffold and chronological splits are more difficult than random splits, and hence scaffold splits should be preferred over random splits during evaluation.”
• Learning Graph-Level Representation for Drug Discovery (2017, methodology): “For chemical data, random splitting datasets into train/valid/test collections, common in machine learning, is often not correct(Sheridan 2013).”
• Molecular de-novo design through deep reinforcement learning (2017, neutral): “In order to decrease the nearest neighbour similarity between the training and testing structures [34–36], the actives were grouped in clusters based on their molecular similarity.”
• Pre-training Graph Neural Networks (2019, background): “Prior work has shown that this scaffold split provides a more realistic estimate of model performance in prospective evaluation compared to random split [6, 45].”
• Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development (2021, background): “Similar to theADMEprediction, as the drug structures of interest evolve over time (Sheridan 2013), toxicity prediction requires a model to generalize to a set of novel drugs with small structural similarity to the existing drug set.”