workbench 0.8.177__py3-none-any.whl → 0.8.227__py3-none-any.whl

workbench/algorithms/dataframe/proximity.py

@@ -1,384 +1,338 @@
 import pandas as pd
 import numpy as np
-from sklearn.preprocessing import StandardScaler
-from sklearn.neighbors import NearestNeighbors
-from typing import List, Dict
+from abc import ABC, abstractmethod
+from typing import List, Dict, Optional, Union
 import logging
-import pickle
-import os
-import json
-from pathlib import Path
-from enum import Enum
 
 # Set up logging
 log = logging.getLogger("workbench")
 
 
-# ^Enumerated^ Proximity Types (distance or similarity)
-class ProximityType(Enum):
-    DISTANCE = "distance"
-    SIMILARITY = "similarity"
+class Proximity(ABC):
+    """Abstract base class for proximity/neighbor computations."""
 
-
-class Proximity:
     def __init__(
         self,
         df: pd.DataFrame,
         id_column: str,
         features: List[str],
-        target: str = None,
-        track_columns: List[str] = None,
-        n_neighbors: int = 10,
+        target: Optional[str] = None,
+        include_all_columns: bool = False,
     ):
         """
         Initialize the Proximity class.
 
         Args:
-            df (pd.DataFrame): DataFrame containing data for neighbor computations.
-            id_column (str): Name of the column used as the identifier.
-            features (List[str]): List of feature column names to be used for neighbor computations.
-            target (str, optional): Name of the target column. Defaults to None.
-            track_columns (List[str], optional): Additional columns to track in results. Defaults to None.
-            n_neighbors (int): Number of neighbors to compute. Defaults to 10.
+            df: DataFrame containing data for neighbor computations.
+            id_column: Name of the column used as the identifier.
+            features: List of feature column names to be used for neighbor computations.
+            target: Name of the target column. Defaults to None.
+            include_all_columns: Include all DataFrame columns in neighbor results. Defaults to False.
         """
-        self.df = df.dropna(subset=features).copy()
         self.id_column = id_column
-        self.n_neighbors = min(n_neighbors, len(self.df) - 1)
-        self.target = target
         self.features = features
-        self.scaler = None
-        self.X = None
-        self.nn = None
-        self.proximity_type = None
-        self.track_columns = track_columns or []
-
-        # Right now we only support numeric features, so remove any columns that are not numeric
-        non_numeric_features = self.df[self.features].select_dtypes(exclude=["number"]).columns.tolist()
-        if non_numeric_features:
-            log.warning(f"Non-numeric features {non_numeric_features} aren't currently supported...")
-            self.features = [f for f in self.features if f not in non_numeric_features]
-
-        # Build the proximity model
-        self.build_proximity_model()
-
-    def build_proximity_model(self) -> None:
-        """Standardize features and fit Nearest Neighbors model.
-        Note: This method can be overridden in subclasses for custom behavior."""
-        self.proximity_type = ProximityType.DISTANCE
-        self.scaler = StandardScaler()
-        self.X = self.scaler.fit_transform(self.df[self.features])
-        self.nn = NearestNeighbors(n_neighbors=self.n_neighbors + 1).fit(self.X)
-
-    def all_neighbors(self) -> pd.DataFrame:
+        self.target = target
+        self.include_all_columns = include_all_columns
+
+        # Store the DataFrame (subclasses may filter/modify in _prepare_data)
+        self.df = df.copy()
+
+        # Prepare data (subclasses can override)
+        self._prepare_data()
+
+        # Compute target range if target is provided
+        self.target_range = None
+        if self.target and self.target in self.df.columns:
+            self.target_range = self.df[self.target].max() - self.df[self.target].min()
+
+        # Build the proximity model (subclass-specific)
+        self._build_model()
+
+        # Precompute landscape metrics
+        self._precompute_metrics()
+
+        # Define core columns for output (subclasses can override)
+        self._set_core_columns()
+
+        # Project the data to 2D (subclass-specific)
+        self._project_2d()
+
+    def _prepare_data(self) -> None:
+        """Prepare the data before building the model. Subclasses can override."""
+        pass
+
+    def _set_core_columns(self) -> None:
+        """Set the core columns for output. Subclasses can override."""
+        self.core_columns = [self.id_column, "nn_distance", "nn_id"]
+        if self.target:
+            self.core_columns.extend([self.target, "nn_target", "nn_target_diff"])
+
+    @abstractmethod
+    def _build_model(self) -> None:
+        """Build the proximity model. Must set self.nn (NearestNeighbors instance)."""
+        pass
+
+    @abstractmethod
+    def _transform_features(self, df: pd.DataFrame) -> np.ndarray:
+        """Transform features for querying. Returns feature matrix for nearest neighbor lookup."""
+        pass
+
+    @abstractmethod
+    def _project_2d(self) -> None:
+        """Project the data to 2D for visualization. Updates self.df with 'x' and 'y' columns."""
+        pass
+
+    def isolated(self, top_percent: float = 1.0) -> pd.DataFrame:
         """
-        Compute nearest neighbors for all rows in the dataset.
+        Find isolated data points based on distance to nearest neighbor.
+
+        Args:
+            top_percent: Percentage of most isolated data points to return (e.g., 1.0 returns top 1%)
 
         Returns:
-            pd.DataFrame: A DataFrame of neighbors and their distances.
+            DataFrame of observations above the percentile threshold, sorted by distance (descending)
         """
-        distances, indices = self.nn.kneighbors(self.X)
-        results = []
+        percentile = 100 - top_percent
+        threshold = np.percentile(self.df["nn_distance"], percentile)
+        isolated = self.df[self.df["nn_distance"] >= threshold].copy()
+        isolated = isolated.sort_values("nn_distance", ascending=False).reset_index(drop=True)
+        return isolated if self.include_all_columns else isolated[self.core_columns]
 
-        for i, (dists, nbrs) in enumerate(zip(distances, indices)):
-            query_id = self.df.iloc[i][self.id_column]
+    def proximity_stats(self) -> pd.DataFrame:
+        """
+        Return distribution statistics for nearest neighbor distances.
 
-            # Process neighbors
-            for neighbor_idx, dist in zip(nbrs, dists):
-                # Skip self (neighbor index == current row index)
-                if neighbor_idx == i:
-                    continue
-                results.append(self._build_neighbor_result(query_id=query_id, neighbor_idx=neighbor_idx, distance=dist))
+        Returns:
+            DataFrame with proximity distribution statistics (count, mean, std, percentiles)
+        """
+        return (
+            self.df["nn_distance"].describe(percentiles=[0.01, 0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95, 0.99]).to_frame()
+        )
+
+    def target_gradients(
+        self,
+        top_percent: float = 1.0,
+        min_delta: Optional[float] = None,
+        k_neighbors: int = 4,
+        only_coincident: bool = False,
+    ) -> pd.DataFrame:
+        """
+        Find compounds with steep target gradients (data quality issues and activity cliffs).
 
-        return pd.DataFrame(results)
+        Uses a two-phase approach:
+        1. Quick filter using nearest neighbor gradient
+        2. Verify using k-neighbor median to handle cases where the nearest neighbor is the outlier
+
+        Args:
+            top_percent: Percentage of compounds with steepest gradients to return (e.g., 1.0 = top 1%)
+            min_delta: Minimum absolute target difference to consider. If None, defaults to target_range/100
+            k_neighbors: Number of neighbors to use for median calculation (default: 4)
+            only_coincident: If True, only consider compounds that are coincident (default: False)
+
+        Returns:
+            DataFrame of compounds with steepest gradients, sorted by gradient (descending)
+        """
+        if self.target is None:
+            raise ValueError("Target column must be specified")
+
+        epsilon = 1e-6
+
+        # Phase 1: Quick filter using precomputed nearest neighbor
+        candidates = self.df.copy()
+        candidates["gradient"] = candidates["nn_target_diff"] / (candidates["nn_distance"] + epsilon)
+
+        # Apply min_delta
+        if min_delta is None:
+            min_delta = self.target_range / 100.0 if self.target_range > 0 else 0.0
+        candidates = candidates[candidates["nn_target_diff"] >= min_delta]
+
+        # Filter based on mode
+        if only_coincident:
+            # Only keep coincident points (nn_distance ~= 0)
+            candidates = candidates[candidates["nn_distance"] < epsilon].copy()
+        else:
+            # Get top X% by initial gradient
+            percentile = 100 - top_percent
+            threshold = np.percentile(candidates["gradient"], percentile)
+            candidates = candidates[candidates["gradient"] >= threshold].copy()
+
+        # Phase 2: Verify with K-neighbor median to filter out cases where nearest neighbor is the outlier
+        results = []
+        for _, row in candidates.iterrows():
+            cmpd_id = row[self.id_column]
+            cmpd_target = row[self.target]
+
+            # Get K nearest neighbors (excluding self)
+            nbrs = self.neighbors(cmpd_id, n_neighbors=k_neighbors, include_self=False)
+
+            # Calculate median target of k neighbors, excluding the nearest neighbor (index 0)
+            neighbor_median = nbrs.iloc[1:k_neighbors][self.target].median()
+            median_diff = abs(cmpd_target - neighbor_median)
+
+            # Only keep if compound differs from neighborhood median
+            # This filters out cases where the nearest neighbor is the outlier
+            if median_diff >= min_delta:
+                results.append(
+                    {
+                        self.id_column: cmpd_id,
+                        self.target: cmpd_target,
+                        "nn_target": row["nn_target"],
+                        "nn_target_diff": row["nn_target_diff"],
+                        "nn_distance": row["nn_distance"],
+                        "gradient": row["gradient"],  # Keep Phase 1 gradient
+                        "neighbor_median": neighbor_median,
+                        "neighbor_median_diff": median_diff,
+                    }
+                )
+
+        # Handle empty results
+        if not results:
+            return pd.DataFrame(
+                columns=[
+                    self.id_column,
+                    self.target,
+                    "nn_target",
+                    "nn_target_diff",
+                    "nn_distance",
+                    "gradient",
+                    "neighbor_median",
+                    "neighbor_median_diff",
+                ]
+            )
+
+        results_df = pd.DataFrame(results)
+        results_df = results_df.sort_values("gradient", ascending=False).reset_index(drop=True)
+        return results_df
 
     def neighbors(
         self,
-        query_df: pd.DataFrame,
-        radius: float = None,
+        id_or_ids: Union[str, int, List[Union[str, int]]],
+        n_neighbors: Optional[int] = 5,
+        radius: Optional[float] = None,
         include_self: bool = True,
     ) -> pd.DataFrame:
         """
-        Return neighbors for rows in a query DataFrame.
+        Return neighbors for ID(s) from the existing dataset.
 
         Args:
-            query_df: DataFrame containing query points
+            id_or_ids: Single ID or list of IDs to look up
+            n_neighbors: Number of neighbors to return (default: 5, ignored if radius is set)
             radius: If provided, find all neighbors within this radius
-            include_self: Whether to include self in results (if present)
+            include_self: Whether to include self in results (default: True)
 
         Returns:
             DataFrame containing neighbors and distances
-
-        Note: The query DataFrame must include the feature columns. The id_column is optional.
         """
-        # Check if all required features are present
-        missing = set(self.features) - set(query_df.columns)
-        if missing:
-            raise ValueError(f"Query DataFrame is missing required feature columns: {missing}")
+        # Normalize to list
+        ids = [id_or_ids] if not isinstance(id_or_ids, list) else id_or_ids
 
-        # Check if id_column is present
-        id_column_present = self.id_column in query_df.columns
+        # Validate IDs exist
+        missing_ids = set(ids) - set(self.df[self.id_column])
+        if missing_ids:
+            raise ValueError(f"IDs not found in dataset: {missing_ids}")
 
-        # None of the features can be NaNs, so report rows with NaNs and then drop them
-        rows_with_nan = query_df[self.features].isna().any(axis=1)
+        # Filter to requested IDs and preserve order
+        query_df = self.df[self.df[self.id_column].isin(ids)]
+        query_df = query_df.set_index(self.id_column).loc[ids].reset_index()
 
-        # Print the ID column for rows with NaNs
-        if rows_with_nan.any():
-            log.warning(f"Found {rows_with_nan.sum()} rows with NaNs in feature columns:")
-            log.warning(query_df.loc[rows_with_nan, self.id_column])
+        # Transform query features (subclass-specific)
+        X_query = self._transform_features(query_df)
 
-        # Drop rows with NaNs in feature columns and reassign to query_df
-        query_df = query_df.dropna(subset=self.features)
-
-        # Transform the query features using the model's scaler
-        X_query = self.scaler.transform(query_df[self.features])
-
-        # Get neighbors using either radius or k-nearest neighbors
+        # Get neighbors
         if radius is not None:
             distances, indices = self.nn.radius_neighbors(X_query, radius=radius)
         else:
-            distances, indices = self.nn.kneighbors(X_query)
+            distances, indices = self.nn.kneighbors(X_query, n_neighbors=n_neighbors)
 
         # Build results
-        all_results = []
+        results = []
         for i, (dists, nbrs) in enumerate(zip(distances, indices)):
-            # Use the ID from the query DataFrame if available, otherwise use the row index
-            query_id = query_df.iloc[i][self.id_column] if id_column_present else f"query_{i}"
+            query_id = query_df.iloc[i][self.id_column]
 
             for neighbor_idx, dist in zip(nbrs, dists):
-                # Skip if the neighbor is the query itself and include_self is False
                 neighbor_id = self.df.iloc[neighbor_idx][self.id_column]
+
+                # Skip self if requested
                 if not include_self and neighbor_id == query_id:
                     continue
 
-                all_results.append(
-                    self._build_neighbor_result(query_id=query_id, neighbor_idx=neighbor_idx, distance=dist)
-                )
+                results.append(self._build_neighbor_result(query_id=query_id, neighbor_idx=neighbor_idx, distance=dist))
 
-        return pd.DataFrame(all_results)
+        df_results = pd.DataFrame(results)
+        df_results["is_self"] = df_results["neighbor_id"] == df_results[self.id_column]
+        df_results = df_results.sort_values([self.id_column, "is_self", "distance"], ascending=[True, False, True])
+        return df_results.drop("is_self", axis=1).reset_index(drop=True)
 
-    def _build_neighbor_result(self, query_id, neighbor_idx: int, distance: float) -> Dict:
+    def _precompute_metrics(self) -> None:
         """
-        Internal: Build a result dictionary for a single neighbor.
+        Precompute landscape metrics for all compounds.
 
-        Args:
-            query_id: ID of the query point
-            neighbor_idx: Index of the neighbor in the original DataFrame
-            distance: Distance between query and neighbor
+        Adds columns to self.df:
+        - nn_distance: Distance to nearest neighbor
+        - nn_id: ID of nearest neighbor
 
-        Returns:
-            Dictionary containing neighbor information
+        If target is specified, also adds:
+        - nn_target: Target value of nearest neighbor
+        - nn_target_diff: Absolute difference from nearest neighbor target
         """
-        neighbor_id = self.df.iloc[neighbor_idx][self.id_column]
+        log.info("Precomputing proximity metrics...")
 
-        # Basic neighbor info
-        neighbor_info = {
-            self.id_column: query_id,
-            "neighbor_id": neighbor_id,
-            "distance": distance,
-        }
+        # Get nearest neighbors for all points (n=2 because index 0 is self)
+        X = self._transform_features(self.df)
+        distances, indices = self.nn.kneighbors(X, n_neighbors=2)
 
-        # Determine which additional columns to include
-        relevant_cols = [self.target, "prediction"] if self.target else []
-        relevant_cols += [c for c in self.df.columns if "_proba" in c or "residual" in c]
-        relevant_cols += ["outlier"]
+        # Extract nearest neighbor (index 1, since index 0 is self)
+        self.df["nn_distance"] = distances[:, 1]
+        self.df["nn_id"] = self.df.iloc[indices[:, 1]][self.id_column].values
 
-        # Add user-specified columns
-        relevant_cols += self.track_columns
+        # If target exists, compute target-based metrics
+        if self.target and self.target in self.df.columns:
+            # Get target values for nearest neighbor
+            nn_target_values = self.df.iloc[indices[:, 1]][self.target].values
+            self.df["nn_target"] = nn_target_values
+            self.df["nn_target_diff"] = np.abs(self.df[self.target].values - nn_target_values)
 
-        # Add values for each relevant column that exists in the dataframe
-        for col in filter(lambda c: c in self.df.columns, relevant_cols):
-            neighbor_info[col] = self.df.iloc[neighbor_idx][col]
+            # Precompute target range for min_delta default
+            self.target_range = self.df[self.target].max() - self.df[self.target].min()
 
-        return neighbor_info
+        log.info("Proximity metrics precomputed successfully")
 
-    def serialize(self, directory: str) -> None:
+    def _build_neighbor_result(self, query_id, neighbor_idx: int, distance: float) -> Dict:
         """
-        Serialize the Proximity model to a directory.
+        Build a result dictionary for a single neighbor.
 
         Args:
-            directory: Directory path to save the model components
-        """
-        # Create directory if it doesn't exist
-        os.makedirs(directory, exist_ok=True)
-
-        # Save metadata
-        metadata = {
-            "id_column": self.id_column,
-            "features": self.features,
-            "target": self.target,
-            "track_columns": self.track_columns,
-            "n_neighbors": self.n_neighbors,
-        }
-
-        with open(os.path.join(directory, "metadata.json"), "w") as f:
-            json.dump(metadata, f)
-
-        # Save the DataFrame
-        self.df.to_pickle(os.path.join(directory, "df.pkl"))
+            query_id: ID of the query point
+            neighbor_idx: Index of the neighbor in the original DataFrame
+            distance: Distance between query and neighbor
 
-        # Save the scaler and nearest neighbors model
-        with open(os.path.join(directory, "scaler.pkl"), "wb") as f:
-            pickle.dump(self.scaler, f)
+        Returns:
+            Dictionary containing neighbor information
+        """
+        neighbor_row = self.df.iloc[neighbor_idx]
+        neighbor_id = neighbor_row[self.id_column]
 
-        with open(os.path.join(directory, "nn_model.pkl"), "wb") as f:
-            pickle.dump(self.nn, f)
+        # Start with basic info
+        result = {
+            self.id_column: query_id,
+            "neighbor_id": neighbor_id,
+            "distance": 0.0 if distance < 1e-6 else distance,
+        }
 
-        log.info(f"Proximity model serialized to {directory}")
+        # Add target if present
+        if self.target and self.target in self.df.columns:
+            result[self.target] = neighbor_row[self.target]
 
-    @classmethod
-    def deserialize(cls, directory: str) -> "Proximity":
-        """
-        Deserialize a Proximity model from a directory.
+        # Add prediction/probability columns if they exist
+        for col in self.df.columns:
+            if col == "prediction" or "_proba" in col or "residual" in col or col == "in_model":
+                result[col] = neighbor_row[col]
 
-        Args:
-            directory: Directory path containing the serialized model components
+        # Include all columns if requested
+        if self.include_all_columns:
+            result.update(neighbor_row.to_dict())
+            # Restore query_id after update (neighbor_row may have overwritten id column)
+            result[self.id_column] = query_id
+            result["neighbor_id"] = neighbor_id
 
-        Returns:
-            Proximity: A new Proximity instance
-        """
-        directory_path = Path(directory)
-        if not directory_path.exists() or not directory_path.is_dir():
-            raise ValueError(f"Directory {directory} does not exist or is not a directory")
-
-        # Load metadata
-        with open(os.path.join(directory, "metadata.json"), "r") as f:
-            metadata = json.load(f)
-
-        # Load DataFrame
-        df_path = os.path.join(directory, "df.pkl")
-        if not os.path.exists(df_path):
-            raise FileNotFoundError(f"DataFrame file not found at {df_path}")
-        df = pd.read_pickle(df_path)
-
-        # Create instance but skip _prepare_data
-        instance = cls.__new__(cls)
-        instance.df = df
-        instance.id_column = metadata["id_column"]
-        instance.features = metadata["features"]
-        instance.target = metadata["target"]
-        instance.track_columns = metadata["track_columns"]
-        instance.n_neighbors = metadata["n_neighbors"]
-
-        # Load scaler and nn model
-        with open(os.path.join(directory, "scaler.pkl"), "rb") as f:
-            instance.scaler = pickle.load(f)
-
-        with open(os.path.join(directory, "nn_model.pkl"), "rb") as f:
-            instance.nn = pickle.load(f)
-
-        # Load X from scaler transform
-        instance.X = instance.scaler.transform(instance.df[instance.features])
-
-        log.info(f"Proximity model deserialized from {directory}")
-        return instance
-
-
-# Testing the Proximity class
-if __name__ == "__main__":
-
-    pd.set_option("display.max_columns", None)
-    pd.set_option("display.width", 1000)
-
-    # Create a sample DataFrame
-    data = {
-        "ID": [1, 2, 3, 4, 5],
-        "Feature1": [0.1, 0.2, 0.3, 0.4, 0.5],
-        "Feature2": [0.5, 0.4, 0.3, 0.2, 0.1],
-        "Feature3": [2.5, 2.4, 2.3, 2.3, np.nan],
-    }
-    df = pd.DataFrame(data)
-
-    # Test the Proximity class
-    features = ["Feature1", "Feature2", "Feature3"]
-    prox = Proximity(df, id_column="ID", features=features, n_neighbors=3)
-    print(prox.all_neighbors())
-
-    # Test the neighbors method
-    print(prox.neighbors(query_df=df.iloc[[0]]))
-
-    # Test the neighbors method with radius
-    print(prox.neighbors(query_df=df.iloc[0:2], radius=2.0))
-
-    # Test with data that isn't in the 'train' dataframe
-    query_data = {
-        "ID": [6],
-        "Feature1": [0.31],
-        "Feature2": [0.31],
-        "Feature3": [2.31],
-    }
-    query_df = pd.DataFrame(query_data)
-    print(prox.neighbors(query_df=query_df))
-
-    # Test with Features list
-    prox = Proximity(df, id_column="ID", features=["Feature1"], n_neighbors=2)
-    print(prox.all_neighbors())
-
-    # Create a sample DataFrame
-    data = {
-        "foo_id": ["a", "b", "c", "d", "e"],  # Testing string IDs
-        "Feature1": [0.1, 0.2, 0.3, 0.4, 0.5],
-        "Feature2": [0.5, 0.4, 0.3, 0.2, 0.1],
-        "target": [1, 0, 1, 0, 5],
-    }
-    df = pd.DataFrame(data)
-
-    # Test with String Ids
-    prox = Proximity(
-        df,
-        id_column="foo_id",
-        features=["Feature1", "Feature2"],
-        target="target",
-        track_columns=["Feature1", "Feature2"],
-        n_neighbors=3,
-    )
-    print(prox.all_neighbors())
-
-    # Test the neighbors method
-    print(prox.neighbors(query_df=df.iloc[0:2]))
-
-    # Time neighbors with all IDs versus calling all_neighbors
-    import time
-
-    start_time = time.time()
-    prox_df = prox.neighbors(query_df=df, include_self=False)
-    end_time = time.time()
-    print(f"Time taken for neighbors: {end_time - start_time:.4f} seconds")
-    start_time = time.time()
-    prox_df_all = prox.all_neighbors()
-    end_time = time.time()
-    print(f"Time taken for all_neighbors: {end_time - start_time:.4f} seconds")
-
-    # Now compare the two dataframes
-    print("Neighbors DataFrame:")
-    print(prox_df)
-    print("\nAll Neighbors DataFrame:")
-    print(prox_df_all)
-    # Check for any discrepancies
-    if prox_df.equals(prox_df_all):
-        print("The two DataFrames are equal :)")
-    else:
-        print("ERROR: The two DataFrames are not equal!")
-
-    # Test querying without the id_column
-    df_no_id = df.drop(columns=["foo_id"])
-    print(prox.neighbors(query_df=df_no_id, include_self=False))
-
-    # Test duplicate IDs
-    data = {
-        "foo_id": ["a", "b", "c", "d", "d"],  # Duplicate ID (d)
-        "Feature1": [0.1, 0.2, 0.3, 0.4, 0.5],
-        "Feature2": [0.5, 0.4, 0.3, 0.2, 0.1],
-        "target": [1, 0, 1, 0, 5],
-    }
-    df = pd.DataFrame(data)
-    prox = Proximity(df, id_column="foo_id", features=["Feature1", "Feature2"], target="target", n_neighbors=3)
-    print(df.equals(prox.df))
-
-    # Test with a categorical feature
-    from workbench.api import FeatureSet, Model
-
-    fs = FeatureSet("abalone_features")
-    model = Model("abalone-regression")
-    df = fs.pull_dataframe()
-    prox = Proximity(df, id_column=fs.id_column, features=model.features(), target=model.target())
-    print(prox.neighbors(query_df=df[0:2]))
+        return result