混合类型目标变量的多目标预测求助：分类目标预测效果不佳

Hey there! Let’s work through why your $target1$ (the binary classification variable) is underperforming while $target2$ (the continuous one) is doing great. I’ve got targeted fixes to try out, based on common pitfalls with mixed multi-output tasks and random forests.

First off, let’s address a likely core issue: you’re using a MultiOutputClassifier for a mixed task—one classification, one regression. Sklearn’s MultiOutputClassifier trains a separate classifier for each target, which means it’s forcing your continuous $target2$ into a classification framework (even if you didn’t intend it). Alternatively, if you switched to MultiOutputRegressor to handle $target2$, it’s treating $target1$ as a regression problem, which explains the wonky classification results. Let’s fix that first.

1. Adjust Your Model Architecture for Mixed Targets #

You’ve got two distinct tasks here—don’t force them into a one-size-fits-all multi-output wrapper. Try these approaches:

• Split into separate models (simplest & most effective)
  Train a dedicated classifier for $target1$ and regressor for $target2$. This lets each model optimize for its specific task without interference. Example code:

  from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
  from sklearn.model_selection import train_test_split
  
  # Split data into features and both targets
  X_train, X_test, y1_train, y1_test, y2_train, y2_test = train_test_split(
      X, target1, target2, test_size=0.2, random_state=42
  )
  
  # Train classification model for target1
  clf = RandomForestClassifier(n_estimators=200, class_weight="balanced", random_state=42)
  clf.fit(X_train, y1_train)
  
  # Train regression model for target2
  reg = RandomForestRegressor(n_estimators=200, random_state=42)
  reg.fit(X_train, y2_train)
  

• Use a multi-task model (for shared feature learning)
  If you want the model to learn shared features across both tasks, try frameworks that support mixed objectives like XGBoost or neural networks. For example, with XGBoost:

  import xgboost as xgb
  import pandas as pd
  
  # Combine targets into a single dataframe
  y_train = pd.concat([y1_train, y2_train], axis=1)
  y_test = pd.concat([y1_test, y2_test], axis=1)
  
  # Define mixed objectives: binary classification for target1, regression for target2
  dtrain = xgb.DMatrix(X_train, label=y_train)
  dtest = xgb.DMatrix(X_test, label=y_test)
  
  params = {
      "objective": ["binary:logistic", "reg:squarederror"],
      "eval_metric": ["logloss", "rmse"],
      "n_estimators": 200,
      "random_state": 42
  }
  
  model = xgb.train(params, dtrain, num_boost_round=200, evals=[(dtest, "test")])
  

2. Diagnose Data Issues for $target1$ #

If splitting models doesn’t fix things, the problem is likely in your data for the classification task:

• Check for class imbalance
  Count the ratio of 0s and 1s in $target1$. If one class makes up 90%+ of the data, random forests will default to predicting the majority class (killing your minority class accuracy). Fixes:
  • Set class_weight="balanced" in RandomForestClassifier to weigh minority classes more heavily.
  • Use SMOTE to oversample the minority class, or undersample the majority class.
  • Evaluate with F1-score or AUC-ROC instead of accuracy—accuracy is useless for imbalanced data.
• Test feature relevance
  Calculate how much each feature correlates with $target1$ using mutual information (sklearn.feature_selection.mutual_info_classif) or chi-squared tests. If most features are strongly tied to $target2$ but not $target1$:
  • Use feature selection (like SelectKBest or RFE) to pick features that actually relate to $target1$.
  • Build task-specific features (e.g., aggregate stats grouped by $target1$ labels).
• Verify data quality
  Double-check for label errors (e.g., 1s marked as 0s) or missing/garbage values in features that matter for $target1$.

3. Tune the Random Forest Classifier’s Hyperparameters #

Random forests have hyperparameters that make a huge difference for classification. Focus on these for $target1$:

• Increase n_estimators (try 200-300) to reduce variance and improve generalization.
• Adjust tree depth: set max_depth=None (let trees grow fully) but add min_samples_split=5 or min_samples_leaf=2 to prevent overfitting.
• Use criterion="entropy" instead of "gini" if your classes have subtle differences.
• Use grid search to automate tuning:

  from sklearn.model_selection import GridSearchCV
  
  param_grid = {
      "n_estimators": [100, 200, 300],
      "max_depth": [None, 10, 20],
      "class_weight": ["balanced", None],
      "criterion": ["gini", "entropy"]
  }
  
  grid_search = GridSearchCV(
      RandomForestClassifier(random_state=42),
      param_grid,
      cv=5,
      scoring="f1"  # Use F1 for imbalanced data
  )
  grid_search.fit(X_train, y1_train)
  best_clf = grid_search.best_estimator_
  

4. Use the Right Evaluation Metrics #

Stop relying on accuracy! For binary classification, especially with imbalanced data, use these metrics to get a true picture of performance:

• F1-score: Balances precision and recall to measure how well you’re capturing the minority class.
• AUC-ROC: Measures the model’s ability to distinguish between 0s and 1s.
• Confusion matrix: Shows exactly where you’re making mistakes (e.g., missing 1s or misclassifying 0s).

内容的提问来源于stack exchange，提问作者Rajeev Motwani