TreeCCA: Canonical Correlation Analysis via Gradient-Boosted Trees
TreeCCA trains gradient-boosted tree ensembles as CCA encoders, using an Eckart-Young loss to make them work inside standard GBT libraries.
The paper presents TreeCCA as a tabular, interpretable alternative to linear and neural CCA methods. Its custom loss supplies per-sample gradients for XGBoost or LightGBM, while tree gain importances show which features drive cross-view correlation. In the reported tests, TreeCCA matches or beats Deep CCA on synthetic benchmarks, recovers sparse support where PMD finds no signal, and reaches comparable UCI HAR performance at lower cost. ArXiv · AI/CL/LG's note
The paper presents TreeCCA as a tabular, interpretable alternative to linear and neural CCA methods. Its custom loss supplies per-sample gradients for XGBoost or LightGBM, while tree gain importances show which features drive cross-view correlation. In the reported tests, TreeCCA matches or beats Deep CCA on synthetic benchmarks, recovers sparse support where PMD finds no signal, and reaches comparable UCI HAR performance at lower cost. ArXiv · AI/CL/LG's note
score 4