From 204efc8bdcac7df240009e0394985789f9e7b788 Mon Sep 17 00:00:00 2001
From: Edrin Basha <edrin.basha@zalando.de>
Date: Fri, 22 May 2020 15:43:13 +0200
Subject: [PATCH] =?UTF-8?q?Add=20a=20new=20CustomDateFeatureSet=20class=20?=
 =?UTF-8?q?to=20calculate=20holiday=20features=20ba=E2=80=A6=20(#12)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* Add a new CustomDateFeatureSet class to calculate holiday features based on an array of pandas timestamps

Add a new CustomDateFeatureSet class to calculate holiday features based on an array of pandas timestamps

* Added test case for the CustomDateFeatureSet

Added test case for the CustomDateFeatureSet

* Added new class in the init file

Added new class in the init file
---
 pts/feature/__init__.py      |  2 +-
 pts/feature/holiday.py       | 74 ++++++++++++++++++++++++++++++++++++
 test/feature/test_holiday.py | 18 +++++++++
 3 files changed, 93 insertions(+), 1 deletion(-)

diff --git a/pts/feature/__init__.py b/pts/feature/__init__.py
index c85f7f7..7d921fe 100644
--- a/pts/feature/__init__.py
+++ b/pts/feature/__init__.py
@@ -1,4 +1,4 @@
-from .holiday import SPECIAL_DATE_FEATURES, SpecialDateFeatureSet
+from .holiday import SPECIAL_DATE_FEATURES, SpecialDateFeatureSet, CustomDateFeatureSet
 from .lag import get_lags_for_frequency, get_fourier_lags_for_frequency
 from .time_feature import (
     DayOfMonth,
diff --git a/pts/feature/holiday.py b/pts/feature/holiday.py
index 5d5e69c..e23ba52 100644
--- a/pts/feature/holiday.py
+++ b/pts/feature/holiday.py
@@ -219,3 +219,77 @@ class SpecialDateFeatureSet:
                 for feat_name in self.feature_names
             ]
         )
+    
+class CustomDateFeatureSet:
+    """
+    Implements calculation of holiday features. The CustomDateFeatureSet is
+    applied on a pandas Series with Datetimeindex and returns a 2D array of
+    the shape (len(dates), num_features), where num_features are the number
+    of holidays.
+
+    Note that for lower than daily granularity the distance to the holiday is
+    still computed on a per-day basis.
+
+    Example use:
+
+        >>> import pandas as pd
+        >>> cfs = CustomDateFeatureSet([pd.to_datetime('20191129', format='%Y%m%d'), pd.to_datetime('20200101', format='%Y%m%d')], kernel)
+        >>> date_indices = pd.date_range(
+        ...     start="2019-11-24",
+        ...     end="2019-12-31",
+        ...     freq='D'
+        ... )
+        >>> cfs(date_indices)
+        array([[1., 0., 0., 0., 0., 0., 0., 0.],
+               [0., 1., 0., 0., 0., 0., 0., 0.]])
+
+    Example use for using a squared exponential kernel:
+
+        >>> kernel = squared_exponential_kernel(alpha=1.0)
+        >>> cfs = CustomDateFeatureSet([pd.to_datetime('20191129', format='%Y%m%d'), pd.to_datetime('20200101', format='%Y%m%d')], kernel)
+        >>> cfs(date_indices)
+        array([[1.00000000e+00, 3.67879441e-01, 1.83156389e-02, 1.23409804e-04,
+                1.12535175e-07, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00],
+               [3.67879441e-01, 1.00000000e+00, 3.67879441e-01, 1.83156389e-02,
+                1.23409804e-04, 1.12535175e-07, 0.00000000e+00, 0.00000000e+00]])
+
+    """
+
+    def __init__(
+        self,
+        reference_dates: List[pd.Timestamp],
+        kernel_function: Callable[[int], int] = indicator,
+    ):
+        """
+        Parameters
+        ----------
+        reference_dates
+            list of panda timestamps for which features should be created.
+        kernel_function
+            kernel function to pass the feature value based
+            on distance in days. Can be indicator function (default),
+            exponential_kernel, squared_exponential_kernel or user defined.
+        """
+        self.reference_dates = reference_dates
+        self.kernel_function = kernel_function
+
+    def __call__(self, dates):
+        """
+        Transform a pandas series with timestamps to holiday features.
+
+        Parameters
+        ----------
+        dates
+            Pandas series with Datetimeindex timestamps.
+        """
+        return np.vstack(
+            [
+                np.hstack(
+                    [
+                        self.kernel_function((index - ref_date).days)          
+                        for index in dates
+                    ]
+                )
+                for ref_date in self.reference_dates
+            ]
+        )
diff --git a/test/feature/test_holiday.py b/test/feature/test_holiday.py
index b45dd25..64b1ec0 100644
--- a/test/feature/test_holiday.py
+++ b/test/feature/test_holiday.py
@@ -39,6 +39,7 @@ from pts.feature.holiday import (
     CYBER_MONDAY,
     SpecialDateFeatureSet,
     squared_exponential_kernel,
+    CustomDateFeatureSet
 )
 
 test_dates = {
@@ -253,4 +254,21 @@ def test_special_date_feature_set_daily_squared_exponential():
     sfs = SpecialDateFeatureSet([CHRISTMAS_EVE, CHRISTMAS_DAY], squared_exp_kernel)
     computed_features = sfs(date_indices)
     np.testing.assert_almost_equal(computed_features, reference_features, decimal=6)
+    
+def test_custom_date_feature_set():
+
+    ref_dates = [pd.to_datetime('20191129', format='%Y%m%d'), pd.to_datetime('20200101', format='%Y%m%d')]
+
+    kernel = exponential_kernel(alpha=1.0)
+
+    cfs = CustomDateFeatureSet(ref_dates, kernel)
+    sfs = SpecialDateFeatureSet([BLACK_FRIDAY, NEW_YEARS_DAY], kernel)
+
+    date_indices = pd.date_range(
+            start=pd.to_datetime('20191101', format='%Y%m%d'),
+            end=pd.to_datetime('20200131', format='%Y%m%d'),
+            freq='D')
+
+    assert(np.sum(cfs(date_indices) - sfs(date_indices)) == 0), "Features don't match"
+