FIX: Use rolling window for outlier detection

+ Rename find_outliers to split_outliers.
+ Make remove_ouliers call split_outliers internally.
+ Add "window" argument to both functions.
+ Add TODOs to call sites not setting window size explicitly.
+ Improve docstrings.

Change-Id: I24961e2859ddbfa62b543031284517c7389a2abb
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
(cherry picked from commit 4f5872c1bb23873b3a93cb471aae8700d5ca029d)

3 files changed, 55 insertions, 40 deletions

diff --git a/resources/tools/presentation/generator_CPTA.py b/resources/tools/presentation/generator_CPTA.py
index 25be09f959..3817ea9043 100644
--- a/resources/tools/presentation/generator_CPTA.py
+++ b/resources/tools/presentation/generator_CPTA.py
@@ -25,7 +25,7 @@ import numpy as np
 import pandas as pd
 
 from collections import OrderedDict
-from utils import find_outliers, archive_input_data, execute_command
+from utils import split_outliers, archive_input_data, execute_command
 
 
 # Command to build the html format of the report
@@ -247,7 +247,8 @@ def _generate_trending_traces(in_data, build_info, period, moving_win_size=10,
 
     data_pd = pd.Series(data_y, index=data_x)
 
-    t_data, outliers = find_outliers(data_pd, outlier_const=1.5)
+    t_data, outliers = split_outliers(data_pd, outlier_const=1.5,
+                                      window=moving_win_size)
     results = _evaluate_results(data_pd, t_data, window=moving_win_size)
 
     anomalies = pd.Series()
diff --git a/resources/tools/presentation/generator_tables.py b/resources/tools/presentation/generator_tables.py
index 74579b0a9d..9b9f09f4be 100644
--- a/resources/tools/presentation/generator_tables.py
+++ b/resources/tools/presentation/generator_tables.py
@@ -25,7 +25,7 @@ from math import isnan
 from xml.etree import ElementTree as ET
 
 from errors import PresentationError
-from utils import mean, stdev, relative_change, remove_outliers, find_outliers
+from utils import mean, stdev, relative_change, remove_outliers, split_outliers
 
 
 def generate_tables(spec, data):
@@ -405,14 +405,16 @@ def table_performance_comparison(table, input_data):
         item = [tbl_dict[tst_name]["name"], ]
         if tbl_dict[tst_name]["ref-data"]:
             data_t = remove_outliers(tbl_dict[tst_name]["ref-data"],
-                                     table["outlier-const"])
+                                     outlier_constant=table["outlier-const"])
+            # TODO: Specify window size.
             item.append(round(mean(data_t) / 1000000, 2))
             item.append(round(stdev(data_t) / 1000000, 2))
         else:
             item.extend([None, None])
         if tbl_dict[tst_name]["cmp-data"]:
             data_t = remove_outliers(tbl_dict[tst_name]["cmp-data"],
-                                     table["outlier-const"])
+                                     outlier_constant=table["outlier-const"])
+            # TODO: Specify window size.
             item.append(round(mean(data_t) / 1000000, 2))
             item.append(round(stdev(data_t) / 1000000, 2))
         else:
@@ -594,14 +596,16 @@ def table_performance_comparison_mrr(table, input_data):
         item = [tbl_dict[tst_name]["name"], ]
         if tbl_dict[tst_name]["ref-data"]:
             data_t = remove_outliers(tbl_dict[tst_name]["ref-data"],
-                                     table["outlier-const"])
+                                     outlier_const=table["outlier-const"])
+            # TODO: Specify window size.
             item.append(round(mean(data_t) / 1000000, 2))
             item.append(round(stdev(data_t) / 1000000, 2))
         else:
             item.extend([None, None])
         if tbl_dict[tst_name]["cmp-data"]:
             data_t = remove_outliers(tbl_dict[tst_name]["cmp-data"],
-                                     table["outlier-const"])
+                                     outlier_const=table["outlier-const"])
+            # TODO: Specify window size.
             item.append(round(mean(data_t) / 1000000, 2))
             item.append(round(stdev(data_t) / 1000000, 2))
         else:
@@ -708,7 +712,8 @@ def table_performance_trending_dashboard(table, input_data):
             name = tbl_dict[tst_name]["name"]
 
             median = pd_data.rolling(window=win_size, min_periods=2).median()
-            trimmed_data, _ = find_outliers(pd_data, outlier_const=1.5)
+            trimmed_data, _ = split_outliers(pd_data, outlier_const=1.5,
+                                             window=win_size)
             stdev_t = pd_data.rolling(window=win_size, min_periods=2).std()
 
             rel_change_lst = [None, ]
diff --git a/resources/tools/presentation/utils.py b/resources/tools/presentation/utils.py
index 8365bfad5c..bc62268937 100644
--- a/resources/tools/presentation/utils.py
+++ b/resources/tools/presentation/utils.py
@@ -67,59 +67,68 @@ def relative_change(nr1, nr2):
 
     return float(((nr2 - nr1) / nr1) * 100)
 
+def remove_outliers(input_list, outlier_const=1.5, window=14):
+    """Return list with outliers removed, using split_outliers.
 
-def remove_outliers(input_data, outlier_const):
-    """
-
-    :param input_data: Data from which the outliers will be removed.
+    :param input_list: Data from which the outliers will be removed.
     :param outlier_const: Outlier constant.
-    :type input_data: list
+    :param window: How many preceding values to take into account.
+    :type input_list: list of floats
     :type outlier_const: float
+    :type window: int
     :returns: The input list without outliers.
-    :rtype: list
+    :rtype: list of floats
     """
 
-    data = np.array(input_data)
-    upper_quartile = np.percentile(data, 75)
-    lower_quartile = np.percentile(data, 25)
-    iqr = (upper_quartile - lower_quartile) * outlier_const
-    quartile_set = (lower_quartile - iqr, upper_quartile + iqr)
-    result_lst = list()
-    for y in data.tolist():
-        if quartile_set[0] <= y <= quartile_set[1]:
-            result_lst.append(y)
-    return result_lst
+    input_series = pd.Series()
+    for index, value in enumerate(input_list):
+        item_pd = pd.Series([value, ], index=[index, ])
+        input_series.append(item_pd)
+    output_series, _ = split_outliers(input_series, outlier_const=outlier_const,
+                                      window=window)
+    output_list = [y for x, y in output_series.items() if not np.isnan(y)]
+
+    return output_list
 
 
-def find_outliers(input_data, outlier_const=1.5):
+def split_outliers(input_series, outlier_const=1.5, window=14):
     """Go through the input data and generate two pandas series:
-    - input data without outliers
+    - input data with outliers replaced by NAN
     - outliers.
     The function uses IQR to detect outliers.
 
-    :param input_data: Data to be examined for outliers.
+    :param input_series: Data to be examined for outliers.
     :param outlier_const: Outlier constant.
-    :type input_data: pandas.Series
+    :param window: How many preceding values to take into account.
+    :type input_series: pandas.Series
     :type outlier_const: float
-    :returns: Tuple: input data with outliers removed; Outliers.
-    :rtype: tuple (trimmed_data, outliers)
+    :type window: int
+    :returns: Input data with NAN outliers and Outliers.
+    :rtype: (pandas.Series, pandas.Series)
     """
 
-    upper_quartile = input_data.quantile(q=0.75)
-    lower_quartile = input_data.quantile(q=0.25)
-    iqr = (upper_quartile - lower_quartile) * outlier_const
-    low = lower_quartile - iqr
-    high = upper_quartile + iqr
+    list_data = list(input_series.items())
+    head_size = min(window, len(list_data))
+    head_list = list_data[:head_size]
     trimmed_data = pd.Series()
     outliers = pd.Series()
-    for item in input_data.items():
-        item_pd = pd.Series([item[1], ], index=[item[0], ])
-        if low <= item[1] <= high:
+    for item_x, item_y in head_list:
+        item_pd = pd.Series([item_y, ], index=[item_x, ])
+        trimmed_data = trimmed_data.append(item_pd)
+    for index, (item_x, item_y) in list(enumerate(list_data))[head_size:]:
+        y_rolling_list = [y for (x, y) in list_data[index - head_size:index]]
+        y_rolling_array = np.array(y_rolling_list)
+        q1 = np.percentile(y_rolling_array, 25)
+        q3 = np.percentile(y_rolling_array, 75)
+        iqr = (q3 - q1) * outlier_const
+        low, high = q1 - iqr, q3 + iqr
+        item_pd = pd.Series([item_y, ], index=[item_x, ])
+        if low <= item_y <= high:
             trimmed_data = trimmed_data.append(item_pd)
         else:
-            trimmed_data = trimmed_data.append(pd.Series([np.nan, ],
-                                                         index=[item[0], ]))
             outliers = outliers.append(item_pd)
+            nan_pd = pd.Series([np.nan, ], index=[item_x, ])
+            trimmed_data = trimmed_data.append(nan_pd)
 
     return trimmed_data, outliers