JasonBatman/XGBoost_vs_LightGBM

Comparison of XGBoost and LightGBM (speed, accuracy and complexity)

Comparisons of XGB a8f670d(2017/11/02) and LGB 7a166fb(2017/11/01)

• LightGBM.ipynb: Modified version of marugari's work
• exp013
  • model : XGB(hist_depthwise, hist_lossguie, hist_GPU, GPU), LGB
  • objective : Binary classification
  • metric : Logloss
  • dataset : make_classification
  • n_train : 0.5M, 1M, 2M, 4M
  • n_valid : n_train/4
  • n_features : 32
  • n_clusters_per_class : 8
  • n_rounds : 100
  • max_depth : 5, 10, 15
  • num_leaves : 2 ** max_depth
• exp014
  • model : XGB(hist_depthwise, hist_lossguie, hist_GPU, GPU), LGB
  • objective : Binary classification
  • metric : Logloss
  • dataset : make_classification
  • n_train : 1,2,4,8,16,32,64 * 10K
  • n_valid : n_train/4
  • n_features : 256
  • n_clusters_per_class : 8
  • n_rounds : 100
  • max_depth : 5, 10
  • num_leaves : 2 ** max_depth

The following codes were run on older versions of XGBoost and LightGBM

• exp010
  • model : XGB(CPU, EQBIN_depthwise, EQBIN_lossguie, GPU), LGB
  • objective : Binary classification
  • metric : Logloss
  • dataset : make_classification
  • n_train : 0.5M, 1M, 2M
  • n_valid : n_train/4
  • n_features : 32
  • n_rounds : 100
  • n_clusters_per_class : 8
  • max_depth : 5, 10, 15
• exp011
  • model : XGB(EQBIN_depthwise, EQBIN_lossguie), LGB
  • objective : Binary classification
  • metric : Logloss
  • dataset : make_classification
  • n_train : 0.5M, 1M, 2M
  • n_valid : n_train/4
  • n_features : 32
  • n_clusters_per_class : 8
  • n_rounds : 200
  • max_depth : 5, 10, 15, 20
  • num_leaves : 32, 256, 1024, 4096, 16384
• exp012
  • model : XGB(EQBIN_depthwise, EQBIN_lossguie), LGB
  • objective : Binary classification
  • metric : Logloss
  • dataset : make_classification
  • n_train : 1, 2, 4, 8, 16, 32 * 10000
  • n_valid : n_train/4
  • n_features : 256
  • n_clusters_per_class : 8
  • n_rounds : 100
  • max_depth : 5, 10, 15, 20
  • num_leaves : 32, 256, 1024, 4096