多元时间序列特征工程的指南

import pandas as pd
 
 # skipping second row, setting time column as a datetime column
 # dataset available here: https://github.com/vcerqueira/blog/tree/main/data
 buoy = pd.read_csv('data/smart_buoy.csv',
                    skiprows=[1],
                    parse_dates=['time'])
 
 # setting time as index
 buoy.set_index('time', inplace=True)
 # resampling to hourly data
 buoy = buoy.resample('H').mean()
 # simplifying column names
 buoy.columns = [
    'PeakP', 'PeakD', 'Upcross',
    'SWH', 'SeaTemp', 'Hmax', 'THmax',
    'MCurDir', 'MCurSpd'
 ]

 import pandas as pd
 
 from sklearn.model_selection import train_test_split
 from sklearn.metrics import mean_absolute_percentage_error as mape
 from sklearn.multioutput import MultiOutputRegressor
 from lightgbm import LGBMRegressor
 
 # https://github.com/vcerqueira/blog/blob/main/src/tde.py
 from src.tde import time_delay_embedding
 
 target_var = 'SWH'
 
 colnames = buoy.columns.tolist()
 
 # create data set with lagged features using time delay embedding
 buoy_ds = []
 for col in buoy:
    col_df = time_delay_embedding(buoy[col], n_lags=24, horizon=12)
    buoy_ds.append(col_df)
 
 # concatenating all variables
 buoy_df = pd.concat(buoy_ds, axis=1).dropna()
 
 # defining target (Y) and explanatory variables (X)
 predictor_variables = buoy_df.columns.str.contains('\(t\-')
 target_variables = buoy_df.columns.str.contains(f'{target_var}\(t\+')
 X = buoy_df.iloc[:, predictor_variables]
 Y = buoy_df.iloc[:, target_variables]
 
 # train/test split
 X_tr, X_ts, Y_tr, Y_ts = train_test_split(X, Y, test_size=0.3, shuffle=False)
 
 # fitting a lgbm model without feature engineering
 model_wo_fe = MultiOutputRegressor(LGBMRegressor())
 model_wo_fe.fit(X_tr, Y_tr)
 
 # getting forecasts for the test set
 preds_wo_fe = model_wo_fe.predict(X_ts)
 
 # computing the MAPE error
 mape(Y_ts, preds_wo_fe)
 # 0.238

import numpy as np
 
 SUMMARY_STATS = {
    'mean': np.mean,
    'sdev': np.std,
 }
 
 univariate_features = {}
 # for each column in the data
 for col in colnames:
    # get lags for that column
    X_col = X.iloc[:, X.columns.str.startswith(col)]
 
    # for each summary stat
    for feat, func in SUMMARY_STATS.items():
        # compute that stat along the rows
        univariate_features[f'{col}_{feat}'] = X_col.apply(func, axis=1)
 
 # concatenate features into a pd.DF
 univariate_features_df = pd.concat(univariate_features, axis=1)

import itertools
 
 import pandas as pd
 
 from scipy.spatial.distance import jensenshannon
 from scipy import signal
 from scipy.special import rel_entr
 
 from src.feature_extraction import covariance, co_integration
 
 BIVARIATE_STATS = {
    'covariance': covariance,
    'co_integration': co_integration,
    'js_div': jensenshannon,
 }
 
 BIVARIATE_TRANSFORMATIONS = {
    'corr': signal.correlate,
    'conv': signal.convolve,
    'rel_entr': rel_entr,
 }
 
 # get all pairs of variables
 col_combs = list(itertools.combinations(colnames, 2))
 
 bivariate_features = []
 # for each row
 for i, _ in X.iterrows():
    # feature set in the i-th time-step
    feature_set_i = {}
    for col1, col2 in col_combs:
        # features for pair of columns col1, col2
 
        # getting the i-th instance for each column
        x1 = X.loc[i, X.columns.str.startswith(col1)]
        x2 = X.loc[i, X.columns.str.startswith(col2)]
 
        # compute each summary stat
        for feat, func in BIVARIATE_SUMMARY_STATS.items():
            feature_set_i[f'{col1}|{col2}_{feat}'] = func(x1, x2)
 
        # for each transformation
        for trans_f, t_func in BIVARIATE_TRANSFORMATIONS.items():
 
            # apply transformation
            xt = t_func(x1, x2)
 
            # compute summary stat
            for feat, s_func in SUMMARY_STATS.items():
                feature_set_i[f'{col1}|{col2}_{trans_f}_{feat}'] = s_func(xt)
 
    bivariate_features.append(feature_set_i)
 
 bivariate_features_df = pd.DataFrame(bivariate_features, index=X.index)

# concatenating all features with lags
 X_with_features = pd.concat([X, univariate_features_df, bivariate_features_df], axis=1)
 
 # train/test split
 X_tr, X_ts, Y_tr, Y_ts = train_test_split(X_with_features, Y, test_size=0.3, shuffle=False)
 
 # fitting a lgbm model with feature engineering
 model_w_fe = MultiOutputRegressor(LGBMRegressor())
 model_w_fe.fit(X_tr, Y_tr)
 
 # getting forecasts for the test set
 preds_w_fe = model_w_fe.predict(X_ts)
 
 # computing MAPE error
 print(mape(Y_ts, preds_w_fe))
 # 0.227

# getting the importance of each feature in each horizon
 avg_imp = pd.DataFrame([x.feature_importances_
  
 
 # getting the top 100 features
 n_top_features = 100
 
 importance_scores = pd.Series(dict(zip(X_tr.columns, avg_imp)))
 top_features = importance_scores.sort_values(ascending=False)[:n_top_features]
 top_features_nm = top_features.index
 
 # subsetting training and testing sets by those features
 X_tr_top = X_tr[top_features_nm]
 X_ts_top = X_ts[top_features_nm]
 
 # re-fitting the lgbm model
 model_top_features = MultiOutputRegressor(LGBMRegressor())
 model_top_features.fit(X_tr_top, Y_tr)
 
 # getting forecasts for the test set
 preds_top_feats = model_top_features.predict(X_ts_top)
 
 # computing MAE error
 mape(Y_ts, preds_top_feats)
 # 0.229