multi_task_forest1.py & experiment.ipynb

multi_task_forest1.py

@@ -0,0 +1,109 @@
+# -*- coding: utf-8 -*-
+"""MultiTaskClf.ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/1d7ipMRMx9b10KeaPRe-yLQq5tAo6Wnvd
+"""
+
+import numpy as np
+from sklearn.metrics import accuracy_score
+from treeple import ObliqueRandomForestClassifier
+
+class MultiTaskForestClassifier:
+    def __init__(self, clf_type="SPORF", **kwargs):
+        if clf_type == "SPORF":
+            self.model_cls = ObliqueRandomForestClassifier
+            self.default_params = {
+                "n_estimators": 200,
+                "feature_combinations": 2.0,
+                "max_depth": 20,
+                "min_samples_split": 5,
+                "min_samples_leaf": 1,
+                "max_features": 0.5,
+                "bootstrap": True
+            }
+        elif clf_type == "MORF":
+            self.model_cls = MORFClassifier  # Liora
+            self.default_params = { ... }
+        elif clf_type == "HonestForest":
+            self.model_cls = HonestForestClassifier  # Riya
+            self.default_params = { ... }
+        else:
+            raise ValueError(f"Unsupported tree: {clf_type}")
+
+        self.params = {**self.default_params, **kwargs}
+        self.model = None
+        self.task_data = {}
+
+    def add_task(self, task_id, X, y):
+        self.task_data[task_id] = (X, y)
+
+    def fit(self, task_ids):
+        X_all, y_all, task_labels = [], [], []
+        for task_id in task_ids:
+            X, y = self.task_data[task_id]
+            X_all.append(X)
+            y_all.append(y)
+            task_labels.append(np.full(len(y), task_id))
+
+        X_all = np.vstack(X_all)
+        y_all = np.concatenate(y_all)
+        task_labels = np.concatenate(task_labels)
+        X_all = np.column_stack((X_all, task_labels))
+
+        self.model = self.model_cls(**self.params, random_state=42)
+        self.model.fit(X_all, y_all)
+
+    def predict(self, X, task_id):
+        X_task = np.column_stack((X, np.full(len(X), task_id)))
+        return self.model.predict(X_task)
+
+    def score(self, X, y, task_id):
+        return accuracy_score(y, self.predict(X, task_id))
+
+    def evaluate_transfer_general(self, forward_train_ids, forward_test_id, backward_train_ids, backward_test_ids, do_reverse=False):
+        """
+        - Forward: train on `forward_train_ids`, test on `forward_test_id`
+        - Backward: train on `backward_train_ids`, test individually on each in `backward_test_ids`
+        - Reverse (not mandatory. default as False but can change to true): train on `forward_test_id`, test on `forward_train_ids`
+        """
+        results = {}
+
+        # Forward
+        self.fit(forward_train_ids)
+        X_test, y_test = self.task_data[forward_test_id]
+        forward_acc = self.score(X_test, y_test, task_id=forward_test_id)
+        results["forward_transfer"] = {
+            "train_on": forward_train_ids,
+            "test_on": forward_test_id,
+            "accuracy": forward_acc
+        }
+
+        # Backward
+        self.fit(backward_train_ids)
+        backward_accuracies = {}
+        for tid in backward_test_ids:
+            X, y = self.task_data[tid]
+            backward_accuracies[f"task{tid}"] = {
+                "train_on": backward_train_ids,
+                "test_on": tid,
+                "accuracy": self.score(X, y, task_id=tid)
+            }
+        results["backward_transfer"] = backward_accuracies
+
+        # Reverse
+        if do_reverse:
+            self.fit([forward_test_id])
+            reverse_accuracies = {}
+            for tid in forward_train_ids:
+                X, y = self.task_data[tid]
+                reverse_accuracies[f"task{tid}"] = {
+                    "train_on": [forward_test_id],
+                    "test_on": tid,
+                    "accuracy": self.score(X, y, task_id=tid)
+                }
+            results["reverse_transfer"] = reverse_accuracies
+
+        return results

multitaskclf.py

@@ -0,0 +1,78 @@
+# -*- coding: utf-8 -*-
+"""MultiTaskClf.ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/1d7ipMRMx9b10KeaPRe-yLQq5tAo6Wnvd
+"""
+
+import numpy as np
+from sklearn.metrics import accuracy_score
+from treeple import ObliqueRandomForestClassifier
+
+class MultiTaskForestClassifier:
+    def __init__(self, clf_type="SPORF", **kwargs):
+        if clf_type == "SPORF":
+            self.model_cls = ObliqueRandomForestClassifier
+            self.default_params = {
+                "n_estimators": 200,
+                "feature_combinations": 2.0,
+                "max_depth": 20,
+                "min_samples_split": 5,
+                "min_samples_leaf": 1,
+                "max_features": 0.5,
+                "bootstrap": True
+            }
+        elif clf_type == "MORF":
+            self.model_cls = MORFClassifier  # Liora
+            self.default_params = { ... }
+        elif clf_type == "HonestForest":
+            self.model_cls = HonestForestClassifier  # Riya
+            self.default_params = { ... }
+        else:
+            raise ValueError(f"Unsupported tree: {clf_type}")
+
+        self.params = {**self.default_params, **kwargs}
+        self.model = None
+        self.task_data = {}
+
+    def add_task(self, task_id, X_train, y_train, X_test, y_test):
+        """Add both training and testing data for each task."""
+        self.task_data[task_id] = {
+            "train": (X_train, y_train),
+            "test": (X_test, y_test)
+        }
+
+    def get_task_ids(self):
+      return list(self.task_data.keys())
+
+    def fit(self):
+        """Train on all tasks jointly (multi-task learning)."""
+        X_all, y_all, task_labels = [], [], []
+        for task_id, data in self.task_data.items():
+            X, y = data["train"]
+            X_all.append(X)
+            y_all.append(y)
+            task_labels.append(np.full(len(y), task_id))
+
+        X_all = np.vstack(X_all)
+        y_all = np.concatenate(y_all)
+        task_labels = np.concatenate(task_labels)
+
+        # Add task ID as a feature
+        X_all = np.column_stack((X_all, task_labels))
+
+        self.model = self.model_cls(**self.params, random_state=42)
+        self.model.fit(X_all, y_all)
+
+    def predict(self, X, task_id):
+        """Make predictions for a specific task."""
+        X_task = np.column_stack((X, np.full(len(X), task_id)))
+        return self.model.predict(X_task)
+
+
+    def score(self, task_id):
+        """Return accuracy on the held-out test set for a specific task."""
+        X_test, y_test = self.task_data[task_id]["test"]
+        return accuracy_score(y_test, self.predict(X_test, task_id))

multitaskclf_ratiosplit.py

@@ -0,0 +1,96 @@
+# -*- coding: utf-8 -*-
+"""MultiTaskClf_ratiosplit.ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/1d7ipMRMx9b10KeaPRe-yLQq5tAo6Wnvd
+"""
+
+import numpy as np
+from sklearn.metrics import accuracy_score
+from treeple import ObliqueRandomForestClassifier
+
+"""
+MultiTaskForestClassifier:
+A unified multi-task learning wrapper for SPORF, MORF, and HonestForest.
+Trains on all tasks jointly and evaluates per-task performance.
+"""
+
+class MultiTaskForestClassifier:
+    def __init__(self, clf_type="SPORF", task_ratios=None, random_state=42, **kwargs):
+        if clf_type == "SPORF":
+            self.model_cls = ObliqueRandomForestClassifier
+            self.default_params = {
+                "n_estimators": 200,
+                "feature_combinations": 2.0,
+                "max_depth": 20,
+                "min_samples_split": 5,
+                "min_samples_leaf": 1,
+                "max_features": 0.5,
+                "bootstrap": True
+            }
+        elif clf_type == "MORF":
+            self.model_cls = MORFClassifier
+            self.default_params = { ... }
+        elif clf_type == "HonestForest":
+            self.model_cls = HonestForestClassifier
+            self.default_params = { ... }
+        else:
+            raise ValueError(f"Unsupported tree: {clf_type}")
+
+        self.params = {**self.default_params, **kwargs}
+        self.model = None
+        self.task_data = {}
+
+        if task_ratios is None:
+            self.task_ratios = {0: 0.1, 1: 0.9}
+        else:
+            self.task_ratios = task_ratios
+
+        self.random_state = random_state
+
+
+    def add_task(self, task_id, X, y, test_size=0.2):
+        from sklearn.model_selection import train_test_split
+        if task_id in self.task_ratios:
+            ratio = self.task_ratios[task_id]
+            X, _, y, _ = train_test_split(X, y, train_size=ratio, stratify=y, random_state=self.random_state)
+
+        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=test_size, stratify=y, random_state=self.random_state)
+        self.task_data[task_id] = {
+            "train": (X_train, y_train),
+            "test": (X_test, y_test)
+    }
+
+    def get_task_ids(self):
+      return list(self.task_data.keys())
+
+    def fit(self):
+        """Train on all tasks jointly (multi-task learning)."""
+        X_all, y_all, task_labels = [], [], []
+        for task_id, data in self.task_data.items():
+            X, y = data["train"]
+            X_all.append(X)
+            y_all.append(y)
+            task_labels.append(np.full(len(y), task_id))
+
+        X_all = np.vstack(X_all)
+        y_all = np.concatenate(y_all)
+        task_labels = np.concatenate(task_labels)
+        X_all = np.column_stack((X_all, task_labels))
+
+        self.model = self.model_cls(**self.params, random_state=42)
+        self.model.fit(X_all, y_all)
+
+    def predict(self, X, task_id):
+        """Make predictions for a specific task."""
+        X_task = np.column_stack((X, np.full(len(X), task_id)))
+        return self.model.predict(X_task)
+
+
+    def score(self, task_id):
+        """Return accuracy on the held-out test set for a specific task."""
+        X_test, y_test = self.task_data[task_id]["test"]
+        return accuracy_score(y_test, self.predict(X_test, task_id))
+