lagged target features

lagged-feats.py

    # Group by product / timewindow and compute aggregate features
    print('[+] Generating weekly lagged product aggregation features ...')
    agg_week = demand_df.groupby(['product_id', 'year', 'weekofyear'])[
        'actual_raw'].agg(num_week_lag_aggregations)
    agg_week.columns = ["_week_lagged_".join(agg_feature)
                        for agg_feature in agg_week.columns.ravel()]
    agg_week.reset_index(drop=False, inplace=True)

    print('[+] Generating monthly lagged product aggregation features ...')
    agg_month = demand_df.groupby(['product_id', 'year', 'month'])[
        'actual_raw'].agg(num_month_lag_aggregations)
    agg_month.columns = ["_month_lagged_".join(agg_feature)
                         for agg_feature in agg_month.columns.ravel()]
    agg_month.reset_index(drop=False, inplace=True)

    # Lag features with 1 respective timewindow for each product
    print('[+] Processing lagged aggregation features ...')
    agg_lagged_week = agg_week.groupby(['product_id']).shift(1)
    agg_lagged_month = agg_month.groupby(['product_id']).shift(1)

    # Add back lost groupby columns
    agg_lagged_week[['product_id', 'year', 'weekofyear']
                    ] = agg_week[['product_id', 'year', 'weekofyear']]
    agg_lagged_month[['product_id', 'year', 'month']] = agg_month[['product_id', 'year', 'month']]

    # Impute now missing first observations per product by backfilling with first known observation
    agg_lagged_week = agg_lagged_week.groupby(['product_id']).fillna(method='backfill')
    agg_lagged_month = agg_lagged_month.groupby(['product_id']).fillna(method='backfill')

    # Add back lost groupby columns x2
    agg_lagged_week['product_id'] = agg_week['product_id']
    agg_lagged_month['product_id'] = agg_month['product_id']