avoid most_freq_bin to be 0 in categorical features (#2824)

* avoid most_freq_bin to be 0 in categorical features

* Apply suggestions from code review

* add tests

* update test

* Apply suggestions from code review

* Apply suggestions from code review

src/io/bin.cpp

@@ -510,20 +510,24 @@ namespace LightGBM {
 
     if (!is_trivial_) {
       default_bin_ = ValueToBin(0);
+      most_freq_bin_ =
+          static_cast<uint32_t>(ArrayArgs<int>::ArgMax(cnt_in_bin));
       if (bin_type_ == BinType::CategoricalBin) {
-        CHECK(default_bin_ > 0);
+        if (most_freq_bin_ == 0) {
+          CHECK(num_bin_ > 1);
+          // FIXME: how to enable `most_freq_bin_ = 0` for categorical features
+          most_freq_bin_ = 1;
+        }
       }
-    }
-    if (!is_trivial_) {
-      most_freq_bin_ = static_cast<uint32_t>(ArrayArgs<int>::ArgMax(cnt_in_bin));
-      // calculate sparse rate
-      sparse_rate_ = static_cast<double>(cnt_in_bin[default_bin_]) / total_sample_cnt;
-      const double max_sparse_rate = static_cast<double>(cnt_in_bin[most_freq_bin_]) / total_sample_cnt;
-      if (most_freq_bin_ != default_bin_ && max_sparse_rate > 0.7f) {
-        sparse_rate_ = max_sparse_rate;
-      } else {
+      const double max_sparse_rate =
+          static_cast<double>(cnt_in_bin[most_freq_bin_]) / total_sample_cnt;
+      // When most_freq_bin_ != default_bin_, there are some additional data loading costs.
+      // so use most_freq_bin_  = default_bin_ when there is not so sparse
+      if (most_freq_bin_ != default_bin_ && max_sparse_rate < kSparseThreshold) {
         most_freq_bin_ = default_bin_;
       }
+      sparse_rate_ =
+          static_cast<double>(cnt_in_bin[most_freq_bin_]) / total_sample_cnt;
     } else {
       sparse_rate_ = 1.0f;
     }

tests/python_package_test/test_engine.py

@@ -335,6 +335,43 @@ class TestEngine(unittest.TestCase):
         self.assertGreater(ret, 0.999)
         self.assertAlmostEqual(evals_result['valid_0']['auc'][-1], ret, places=5)
 
+    def test_categorical_non_zero_inputs(self):
+        x = [1, 1, 1, 1, 1, 1, 2, 2]
+        y = [1, 1, 1, 1, 1, 1, 0, 0]
+
+        X_train = np.array(x).reshape(len(x), 1)
+        y_train = np.array(y)
+        lgb_train = lgb.Dataset(X_train, y_train)
+        lgb_eval = lgb.Dataset(X_train, y_train)
+
+        params = {
+            'objective': 'regression',
+            'metric': 'auc',
+            'verbose': -1,
+            'boost_from_average': False,
+            'min_data': 1,
+            'num_leaves': 2,
+            'learning_rate': 1,
+            'min_data_in_bin': 1,
+            'min_data_per_group': 1,
+            'cat_smooth': 1,
+            'cat_l2': 0,
+            'max_cat_to_onehot': 1,
+            'zero_as_missing': False,
+            'categorical_column': 0
+        }
+        evals_result = {}
+        gbm = lgb.train(params, lgb_train,
+                        num_boost_round=1,
+                        valid_sets=lgb_eval,
+                        verbose_eval=False,
+                        evals_result=evals_result)
+        pred = gbm.predict(X_train)
+        np.testing.assert_allclose(pred, y)
+        ret = roc_auc_score(y_train, pred)
+        self.assertGreater(ret, 0.999)
+        self.assertAlmostEqual(evals_result['valid_0']['auc'][-1], ret, places=5)
+
     def test_multiclass(self):
         X, y = load_digits(10, True)
         X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)