Optimizing Model Learning Performance on a Challenging Heck Reaction Yield Data Set

The development of machine learning (ML) in organic synthesis is limited due to the lack of available data sets. The ML-compatible literature data set for Heck reaction yields, named HeckLit, has been established. With 10,002 cases, the data set spans multiple reaction subclasses and covers a larger chemical space compared to high-throughput experimentation data sets, including nearly 3.6 × 10¹² accessible cases. HeckLit accelerates the advancement of ML-driven synthesis research; however, it suffers from the same dilemma as other literature-based data sets, sparse distribution and high-yield preference, leading to limited model learning ability on the test set, with R² = 0.318. Thus, feature distribution smoothing (FDS) and subset splitting training strategy (SSTS) are utilized to tackle this issue. Despite no improvement when using FDS, SSTS boosted the R² to 0.380. This optimization approach relies on the subset division. Therefore, we suggest a criterion for splitting. The SSTS opens up a new avenue for tackling the challenge of learning from large-scale data sets.