Feature: Implement series-specific model selection using meta-features (#1211)

Author: codeloop

ads/opctl/operator/lowcode/common/data.py

@@ -19,16 +19,21 @@
 
 
 class AbstractData(ABC):
-    def __init__(self, spec, name="input_data", data=None):
+    def __init__(self, spec, name="input_data", data=None, subset=None):
         self.Transformations = Transformations
         self.data = None
         self._data_dict = dict()
         self.name = name
         self.spec = spec
+        self.subset = subset
         if data is not None:
             self.data = data
         else:
             self.load_transform_ingest_data(spec)
+        # Subset by series if requested
+        # if self.subset is not None and hasattr(self, 'data') and self.data is not None:
+        #     subset_str = [str(s) for s in self.subset]
+        #     self.data = self.data[self.data.index.get_level_values(DataColumns.Series).isin(subset_str)]
 
     def get_raw_data_by_cat(self, category):
         mapping = self._data_transformer.get_target_category_columns_map()
@@ -72,7 +77,7 @@ def get_data_for_series(self, series_id):
     def _load_data(self, data_spec, **kwargs):
         loading_start_time = time.time()
         try:
-            raw_data = load_data(data_spec)
+            raw_data = load_data(data_spec, subset=self.subset if self.subset else None, target_category_columns=self.spec.target_category_columns)
         except InvalidParameterError as e:
             e.args = e.args + (f"Invalid Parameter: {self.name}",)
             raise e

ads/opctl/operator/lowcode/common/transformations.py

@@ -294,3 +294,210 @@ def get_target_category_columns_map(self):
     def _fill_na(self, df: pd.DataFrame, na_value=0) -> pd.DataFrame:
         """Fill nans in dataframe"""
         return df.fillna(value=na_value)
+
+    def build_fforms_meta_features(self, data, target_col=None, group_cols=None):
+        """
+        Build meta-features for time series based on FFORMS paper and add them to the original DataFrame.
+
+        Parameters
+        ----------
+        data : pandas.DataFrame
+            Input DataFrame containing time series data
+        target_col : str, optional
+            Name of the target column to calculate meta-features for.
+            If None, uses the target column specified in dataset_info.
+        group_cols : list of str, optional
+            List of columns to group by before calculating meta-features.
+            If None, calculates features for the entire series.
+
+        Returns
+        -------
+        pandas.DataFrame
+            Original DataFrame with additional meta-feature columns
+
+        References
+        ----------
+        Talagala, T. S., Hyndman, R. J., & Athanasopoulos, G. (2023).
+        Meta-learning how to forecast time series. Journal of Forecasting, 42(6), 1476-1501.
+        """
+        if not isinstance(data, pd.DataFrame):
+            raise ValueError("Input must be a pandas DataFrame")
+
+        # Use target column from dataset_info if not specified
+        if target_col is None:
+            target_col = self.target_column_name
+        if target_col not in data.columns:
+            raise ValueError(f"Target column '{target_col}' not found in DataFrame")
+
+        # Check if group_cols are provided and valid
+        if group_cols is not None:
+            if not isinstance(group_cols, list):
+                raise ValueError("group_cols must be a list of column names")
+            for col in group_cols:
+                if col not in data.columns:
+                    raise ValueError(f"Group column '{col}' not found in DataFrame")
+
+        # If no group_cols, get the target_category_columns else treat the entire DataFrame as a single series
+        if not group_cols:
+            group_cols = self.target_category_columns if self.target_category_columns else []
+
+        # Calculate meta-features for each series
+        def calculate_series_features(series):
+            """Calculate features for a single series"""
+            n = len(series)
+            values = series.values
+
+            # Basic statistics
+            mean = series.mean()
+            std = series.std()
+            variance = series.var()
+            skewness = series.skew()
+            kurtosis = series.kurtosis()
+            cv = std / mean if mean != 0 else np.inf
+
+            # Trend features
+            X = np.vstack([np.arange(n), np.ones(n)]).T
+            trend_coef = np.linalg.lstsq(X, values, rcond=None)[0][0]
+            trend_pred = X.dot(np.linalg.lstsq(X, values, rcond=None)[0])
+            residuals = values - trend_pred
+            std_residuals = np.std(residuals)
+
+            # Turning points
+            turning_points = 0
+            for i in range(1, n-1):
+                if (values[i-1] < values[i] and values[i] > values[i+1]) or \
+                   (values[i-1] > values[i] and values[i] < values[i+1]):
+                    turning_points += 1
+            turning_points_rate = turning_points / (n-2) if n > 2 else 0
+
+            # Serial correlation
+            acf1 = series.autocorr(lag=1) if n > 1 else 0
+            acf2 = series.autocorr(lag=2) if n > 2 else 0
+            acf10 = series.autocorr(lag=10) if n > 10 else 0
+
+            # Seasonality features
+            seasonal_strength = 0
+            seasonal_peak_strength = 0
+            if n >= 12:
+                seasonal_lags = [12, 24, 36]
+                seasonal_acfs = []
+                for lag in seasonal_lags:
+                    if n > lag:
+                        acf_val = series.autocorr(lag=lag)
+                        seasonal_acfs.append(abs(acf_val))
+                seasonal_peak_strength = max(seasonal_acfs) if seasonal_acfs else 0
+
+                ma = series.rolling(window=12, center=True).mean()
+                seasonal_comp = series - ma
+                seasonal_strength = 1 - np.var(seasonal_comp.dropna()) / np.var(series)
+
+            # Stability and volatility features
+            values_above_mean = values >= mean
+            crossing_points = np.sum(values_above_mean[1:] != values_above_mean[:-1])
+            crossing_rate = crossing_points / (n - 1) if n > 1 else 0
+
+            # First and second differences
+            diff1 = np.diff(values)
+            diff2 = np.diff(diff1) if len(diff1) > 1 else np.array([])
+
+            diff1_mean = np.mean(np.abs(diff1)) if len(diff1) > 0 else 0
+            diff1_var = np.var(diff1) if len(diff1) > 0 else 0
+            diff2_mean = np.mean(np.abs(diff2)) if len(diff2) > 0 else 0
+            diff2_var = np.var(diff2) if len(diff2) > 0 else 0
+
+            # Nonlinearity features
+            if n > 3:
+                X = values[:-1].reshape(-1, 1)
+                y = values[1:]
+                X2 = X * X
+                X3 = X * X * X
+                X_aug = np.hstack([X, X2, X3])
+                nonlinearity = np.linalg.lstsq(X_aug, y, rcond=None)[1][0] if len(y) > 0 else 0
+            else:
+                nonlinearity = 0
+
+            # Long-term trend features
+            if n >= 10:
+                mid = n // 2
+                trend_change = np.mean(values[mid:]) - np.mean(values[:mid])
+            else:
+                trend_change = 0
+
+            # Step changes and spikes
+            step_changes = np.abs(diff1).max() if len(diff1) > 0 else 0
+            spikes = np.sum(np.abs(values - mean) > 2 * std) / n if std != 0 else 0
+
+            # Hurst exponent and entropy
+            lag = min(10, n // 2)
+            variance_ratio = np.var(series.diff(lag)) / (lag * np.var(series.diff())) if n > lag else 0
+            hurst = np.log(variance_ratio) / (2 * np.log(lag)) if variance_ratio > 0 and lag > 1 else 0
+
+            hist, _ = np.histogram(series, bins='auto', density=True)
+            entropy = -np.sum(hist[hist > 0] * np.log(hist[hist > 0]))
+
+            return pd.Series({
+                'ts_n_obs': n,
+                'ts_mean': mean,
+                'ts_std': std,
+                'ts_variance': variance,
+                'ts_cv': cv,
+                'ts_skewness': skewness,
+                'ts_kurtosis': kurtosis,
+                'ts_trend': trend_coef,
+                'ts_trend_change': trend_change,
+                'ts_std_residuals': std_residuals,
+                'ts_turning_points_rate': turning_points_rate,
+                'ts_seasonal_strength': seasonal_strength,
+                'ts_seasonal_peak_strength': seasonal_peak_strength,
+                'ts_acf1': acf1,
+                'ts_acf2': acf2,
+                'ts_acf10': acf10,
+                'ts_crossing_rate': crossing_rate,
+                'ts_diff1_mean': diff1_mean,
+                'ts_diff1_variance': diff1_var,
+                'ts_diff2_mean': diff2_mean,
+                'ts_diff2_variance': diff2_var,
+                'ts_nonlinearity': nonlinearity,
+                'ts_step_max': step_changes,
+                'ts_spikes_rate': spikes,
+                'ts_hurst': hurst,
+                'ts_entropy': entropy
+            })
+
+        # Create copy of input DataFrame
+        result_df = data.copy()
+
+        if group_cols:
+            # Calculate features for each group
+            features = []
+            # Sort by date within each group if date column exists
+            date_col = self.dt_column_name if self.dt_column_name else 'Date'
+            if date_col in data.columns:
+                data = data.sort_values([date_col] + group_cols)
+
+            for name, group in data.groupby(group_cols):
+                # Sort group by date if exists
+                if date_col in group.columns:
+                    group = group.sort_values(date_col)
+                group_features = calculate_series_features(group[target_col])
+                if isinstance(name, tuple):
+                    feature_row = dict(zip(group_cols, name))
+                else:
+                    feature_row = {group_cols[0]: name}
+                feature_row.update(group_features)
+                features.append(feature_row)
+
+            # Create features DataFrame without merging
+            features_df = pd.DataFrame(features)
+            # Return only the meta-features DataFrame with group columns
+            return features_df
+        else:
+            # Sort by date if exists and calculate features for entire series
+            date_col = self.dt_column_name if self.dt_column_name else 'Date'
+            if date_col in data.columns:
+                data = data.sort_values(date_col)
+            features = calculate_series_features(data[target_col])
+            # Return single row DataFrame with meta-features
+            return pd.DataFrame([features])
+
+        return result_df

ads/opctl/operator/lowcode/common/utils.py

@@ -124,6 +124,14 @@ def load_data(data_spec, storage_options=None, **kwargs):
         data = data[columns]
     if limit:
         data = data[:limit]
+    # Filtering by subset if provided
+    subset = kwargs.get('subset', None)
+    if subset is not None:
+        target_category_columns = kwargs.get('target_category_columns', None)
+        mask = False
+        for col in target_category_columns:
+            mask = mask | data[col].isin(subset)
+            data = data[mask]
     return data
 
 

ads/opctl/operator/lowcode/forecast/__main__.py

@@ -3,17 +3,20 @@
 # Copyright (c) 2023, 2025 Oracle and/or its affiliates.
 # Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/
 
+import copy
 import json
 import os
 import sys
 from typing import Dict, List
 
+import pandas as pd
 import yaml
 
 from ads.opctl import logger
 from ads.opctl.operator.common.const import ENV_OPERATOR_ARGS
 from ads.opctl.operator.common.utils import _parse_input_args
 
+from .const import AUTO_SELECT_SERIES
 from .model.forecast_datasets import ForecastDatasets, ForecastResults
 from .operator_config import ForecastOperatorConfig
 from .whatifserve import ModelDeploymentManager
@@ -24,9 +27,56 @@ def operate(operator_config: ForecastOperatorConfig) -> ForecastResults:
     from .model.factory import ForecastOperatorModelFactory
 
     datasets = ForecastDatasets(operator_config)
-    results = ForecastOperatorModelFactory.get_model(
-        operator_config, datasets
-    ).generate_report()
+    model = ForecastOperatorModelFactory.get_model(operator_config, datasets)
+
+    if operator_config.spec.model == AUTO_SELECT_SERIES and hasattr(
+        operator_config.spec, "meta_features"
+    ):
+        # For AUTO_SELECT_SERIES, handle each series with its specific model
+        meta_features = operator_config.spec.meta_features
+        results = ForecastResults()
+        sub_results_list = []
+
+        # Group the data by selected model
+        for model_name in meta_features["selected_model"].unique():
+            # Get series that use this model
+            series_groups = meta_features[meta_features["selected_model"] == model_name]
+
+            # Create a sub-config for this model
+            sub_config = copy.deepcopy(operator_config)
+            sub_config.spec.model = model_name
+
+            # Create sub-datasets for these series
+            sub_datasets = ForecastDatasets(
+                operator_config,
+                subset=series_groups[operator_config.spec.target_category_columns]
+                .values.flatten()
+                .tolist(),
+            )
+
+            # Get and run the appropriate model
+            sub_model = ForecastOperatorModelFactory.get_model(sub_config, sub_datasets)
+            sub_result_df, sub_elapsed_time = sub_model.build_model()
+            sub_results = sub_model.generate_report(
+                result_df=sub_result_df,
+                elapsed_time=sub_elapsed_time,
+                save_sub_reports=True,
+            )
+            sub_results_list.append(sub_results)
+
+            # results_df = pd.concat([results_df, sub_result_df], ignore_index=True, axis=0)
+            # elapsed_time += sub_elapsed_time
+        # Merge all sub_results into a single ForecastResults object
+        if sub_results_list:
+            results = sub_results_list[0]
+            for sub_result in sub_results_list[1:]:
+                results.merge(sub_result)
+        else:
+            results = None
+
+    else:
+        # For other cases, use the single selected model
+        results = model.generate_report()
     # saving to model catalog
     spec = operator_config.spec
     if spec.what_if_analysis and datasets.additional_data:

ads/opctl/operator/lowcode/forecast/const.py

@@ -89,4 +89,5 @@ class ForecastOutputColumns(ExtendedEnum):
 PROPHET_INTERNAL_DATE_COL = "ds"
 RENDER_LIMIT = 5000
 AUTO_SELECT = "auto-select"
+AUTO_SELECT_SERIES = "auto-select-series"
 BACKTEST_REPORT_NAME = "back_test.csv"