Update causal_selection/data/results/evaluation_results.json with final benchmark (14 networks, 122 configs)

causal_selection/data/results/evaluation_results.json

@@ -1,59 +1,59 @@
 {
   "rf": {
-    "top_k_hit_rate": 0.6724137931034483,
-    "top_k_overlap_rate": 0.5344827586206894,
-    "mean_regret": 0.011830613186842965,
+    "top_k_hit_rate": 0.6639344262295082,
+    "top_k_overlap_rate": 0.5136612021857923,
+    "mean_regret": 0.010241234335280928,
     "median_regret": 0.0,
-    "max_regret": 0.2,
-    "ndcg_at_k": 0.9472865763738757,
-    "mean_pred_mse": 0.03897522735908129,
-    "mean_pred_mae": 0.1254012317597142
+    "max_regret": 0.1818181818181818,
+    "ndcg_at_k": 0.9508166448348011,
+    "mean_pred_mse": 0.03973197010929856,
+    "mean_pred_mae": 0.12461942879927632
   },
   "gbm": {
-    "top_k_hit_rate": 0.646551724137931,
-    "top_k_overlap_rate": 0.5114942528735633,
-    "mean_regret": 0.012502991165511557,
+    "top_k_hit_rate": 0.639344262295082,
+    "top_k_overlap_rate": 0.5136612021857924,
+    "mean_regret": 0.010800135428750801,
     "median_regret": 0.0,
     "max_regret": 0.1785714285714286,
-    "ndcg_at_k": 0.9330250596711344,
-    "mean_pred_mse": 0.047461864051339515,
-    "mean_pred_mae": 0.13108719283488715
+    "ndcg_at_k": 0.946114853798465,
+    "mean_pred_mse": 0.04519970192061315,
+    "mean_pred_mae": 0.1314004270489398
   },
   "feature_importance": {
-    "n_samples": 0.0025887084513785425,
-    "n_variables": 0.1477728757042395,
-    "n_over_p": 0.005618151414487232,
-    "avg_cardinality": 0.015302494551661026,
-    "max_cardinality": 0.022855101206383666,
-    "min_cardinality": 0.0001835133888817687,
-    "mean_entropy": 0.008604456302329989,
-    "std_entropy": 0.054754415922791864,
-    "max_entropy": 0.0946918257906736,
-    "mean_pairwise_MI": 0.02012542859186537,
-    "std_pairwise_MI": 0.04566962942247527,
-    "max_pairwise_MI": 0.2461266647849871,
-    "mean_normalized_MI": 0.012182980490283693,
-    "frac_high_MI_pairs": 0.00428653946971584,
-    "density_proxy": 0.015132300109205282,
-    "mean_chi2_stat": 0.00984875021400495,
-    "std_chi2_stat": 0.006687727708188509,
-    "max_chi2_stat": 0.03969793453146422,
-    "mean_cramers_v": 0.03650593773647407,
-    "max_cramers_v": 0.06675158054783932,
-    "frac_weak_deps": 0.011082215072850458,
-    "frac_strong_deps": 0.016512347580829506,
-    "mean_pval_marginal": 0.008176296972382794,
-    "frac_dep_marginal": 0.010906732971877857,
-    "mean_pval_conditional": 0.007731010247748832,
-    "frac_dep_conditional": 0.012652489124647821,
-    "v_structure_proxy": 0.007418659524534108,
-    "faithfulness_proxy": 0.008413010357448734,
-    "mean_mode_frequency": 0.013104346683199411,
-    "std_mode_frequency": 0.017345380497380575,
-    "mean_balance": 0.0047963758695245395,
-    "uniformity_score": 0.003931779101117974,
-    "frac_binary_vars": 0.017991987844377444,
-    "frac_high_card_vars": 0.00455035181274912
+    "n_samples": 0.0022988785978280417,
+    "n_variables": 0.3002160358395854,
+    "n_over_p": 0.006872254393885206,
+    "avg_cardinality": 0.0028305686233166464,
+    "max_cardinality": 0.008435204814464846,
+    "min_cardinality": 0.0001770838668010093,
+    "mean_entropy": 0.008222522299603973,
+    "std_entropy": 0.017998648497423015,
+    "max_entropy": 0.07417096856568049,
+    "mean_pairwise_MI": 0.008764692796618762,
+    "std_pairwise_MI": 0.011004684102768513,
+    "max_pairwise_MI": 0.23797074062266088,
+    "mean_normalized_MI": 0.018082515669353278,
+    "frac_high_MI_pairs": 0.004738823513844024,
+    "density_proxy": 0.0144191895645574,
+    "mean_chi2_stat": 0.010360513138847784,
+    "std_chi2_stat": 0.006645464114766113,
+    "max_chi2_stat": 0.02298569402792007,
+    "mean_cramers_v": 0.03408222768567789,
+    "max_cramers_v": 0.08407177394405024,
+    "frac_weak_deps": 0.010141702950701992,
+    "frac_strong_deps": 0.009112572595666956,
+    "mean_pval_marginal": 0.0077166705164384035,
+    "frac_dep_marginal": 0.011281737824431787,
+    "mean_pval_conditional": 0.009480231578275988,
+    "frac_dep_conditional": 0.01712523758129132,
+    "v_structure_proxy": 0.006504599870304651,
+    "faithfulness_proxy": 0.006613909659522786,
+    "mean_mode_frequency": 0.011838645199114772,
+    "std_mode_frequency": 0.016615483265606967,
+    "mean_balance": 0.00559360373300629,
+    "uniformity_score": 0.002850097843327797,
+    "frac_binary_vars": 0.004731515530616876,
+    "frac_high_card_vars": 0.006045507172039787
   },
   "selected_model": "rf"
 }