1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
#
# Tag Campaign if ROAS performance is abnormally low within Account
#
#
# Author: Michael S. Huang
# Date: 2023-03-24

RPT_COL_CAMPAIGN = 'Campaign'
RPT_COL_DATE = 'Date'
RPT_COL_ACCOUNT = 'Account'
RPT_COL_CAMPAIGN_ID = 'Campaign ID'
RPT_COL_CAMPAIGN_TYPE = 'Campaign Type'
RPT_COL_CAMPAIGN_STATUS = 'Campaign Status'
RPT_COL_PUB_COST = 'Pub. Cost $'
RPT_COL_COST_PER_CONV = 'Cost/Conv. $'
RPT_COL_ROAS = 'ROAS'
RPT_COL_AVG_CPC = 'Avg. CPC $'
RPT_COL_CONV_RATE = 'Conv. Rate %'
RPT_COL_CTR = 'CTR %'
RPT_COL_CONV = 'Conv.'
RPT_COL_REVENUE = 'Revenue $'
RPT_COL_SEARCH_LOSTTOPISBUDGET = 'Search Lost Top IS (Budget) %'
RPT_COL_SEARCH_LOSTTOPISRANK = 'Search Lost Top IS (Rank) %'
RPT_COL_LOST_IMPRSHAREBUDGET = 'Lost Impr. Share (Budget) %'
RPT_COL_LOST_IMPRSHARERANK = 'Lost Impr. Share (Rank) %'
RPT_COL_DAILY_BUDGET = 'Daily Budget'
RPT_COL_AVG_BID = 'Avg. Bid $'
RPT_COL_HIST_QS = 'Hist. QS'
RPT_COL_IMPR = 'Impr.'
RPT_COL_CLICKS = 'Clicks'
BULK_COL_ACCOUNT = 'Account'
BULK_COL_CAMPAIGN = 'Campaign'
BULK_COL_AUTOMATION_INFO = 'AUTOMATION - INFO'

outputDf[BULK_COL_AUTOMATION_INFO] = numpy.nan

## Data Prep

print(inputDf[RPT_COL_DATE].min(), inputDf[RPT_COL_DATE].max())

# 30-day lookback without most recent 3 days due to conversion lag
start_date = pd.to_datetime(datetime.date.today() - datetime.timedelta(days=33))
end_date = pd.to_datetime(datetime.date.today() - datetime.timedelta(days=3))

df_reduced = inputDf[ (inputDf[RPT_COL_DATE] >= start_date) & (inputDf[RPT_COL_DATE] <= end_date) ]

if (df_reduced.shape[0] > 0):
    print("reduced dates\\n", min(df_reduced[RPT_COL_DATE]), max(df_reduced[RPT_COL_DATE]))
else:
    print("no more input to process")

# reduce to needed columns
df_reduced = df_reduced[[RPT_COL_ACCOUNT, RPT_COL_CAMPAIGN, RPT_COL_DATE, RPT_COL_PUB_COST, RPT_COL_CONV, RPT_COL_REVENUE, RPT_COL_CLICKS]].copy()

# sum metics across dates
df_campaign_perf = df_reduced.groupby([RPT_COL_ACCOUNT, RPT_COL_CAMPAIGN]).sum()

# remove rows without cost 
df_campaign_perf = df_campaign_perf[(df_campaign_perf[RPT_COL_PUB_COST] > 0)]

# index by account
df_campaign_perf = df_campaign_perf.reset_index().set_index([RPT_COL_ACCOUNT]).sort_index()

# calculate features
df_campaign_perf[RPT_COL_COST_PER_CONV] = (df_campaign_perf[RPT_COL_PUB_COST] / df_campaign_perf[RPT_COL_CONV])
df_campaign_perf[RPT_COL_ROAS] = df_campaign_perf[RPT_COL_REVENUE] / df_campaign_perf[RPT_COL_PUB_COST]
df_campaign_perf[RPT_COL_CONV_RATE] = df_campaign_perf[RPT_COL_CONV] / df_campaign_perf[RPT_COL_CLICKS]
df_campaign_perf[RPT_COL_AVG_CPC] = (df_campaign_perf[RPT_COL_PUB_COST] / df_campaign_perf[RPT_COL_CLICKS])


## Define Anomaly Fuctions

# Finds anomalies using a certain function (e.g. sigma rule, IRQ etc.)
# data: DataFrame
#     Dataset with features
# func: func
#     Function to use to find anomalies
# features: list
#     Feature list
# thresh: int
#     Threshold value (e.g. 2/3 * sigma, 2/3 * IRQ)
# Returns: tuple
def get_feature_anomalies(data, func, features=None, thresh=3):

    if features:
        features_to_check = features
    else:
        features_to_check = data.columns 
        
    outliers_over = pd.Series(data=[False] * data.shape[0], index=data[features_to_check].index, name='is_outlier')
    outliers_under = pd.Series(data=[False] * data.shape[0], index=data[features_to_check].index, name='is_outlier')

    anomalies_summary = {}
    for feature in features_to_check:
        anomalies_mask_over, anomalies_mask_under, upper_bound, lower_bound = func(data, feature, thresh=thresh)
        anomalies_mask_combined = pd.concat([anomalies_mask_over, anomalies_mask_under], axis=1).any(1)
        anomalies_summary[feature] = [upper_bound, lower_bound, sum(anomalies_mask_combined), 100*sum(anomalies_mask_combined)/len(anomalies_mask_combined)]
        outliers_over[anomalies_mask_over[anomalies_mask_over].index] = True
        outliers_under[anomalies_mask_under[anomalies_mask_under].index] = True
        
        
    anomalies_summary = pd.DataFrame(anomalies_summary).T
    anomalies_summary.columns=['upper_bound', 'lower_bound', 'anomalies_count', 'anomalies_percentage']
    
    anomalies_ration = round(anomalies_summary['anomalies_percentage'].sum(), 2)
    
    return anomalies_summary, outliers_over, outliers_under

# Finds outliers/anomalies using IRQ 
# data: DataFrame
# col: str
# thresh: int
#     Number of IRQ to apply 
# Returns: Series 
#     Boolean Series Mask of outliers 
def is_anomaly_irq(data, col, thresh):

    IRQ = data[col].quantile(0.75) - data[col].quantile(0.25)
    upper_bound = data[col].quantile(0.75) + (thresh * IRQ)
    lower_bound = data[col].quantile(0.25) - (thresh * IRQ)
    anomalies_mask_over = data[col] > upper_bound
    anomalies_mask_under = data[col] < lower_bound
#     print("Anomalies mask: ", (anomalies_mask_over, anomalies_mask_under))
    
    return anomalies_mask_over, anomalies_mask_under, upper_bound, lower_bound

def find_peer_anomaly(df_slice, features, irq_threshold=1.8, outliers_desired=(True, True)):
    
    (want_outliers_over, want_outliers_under) = outliers_desired
   
    if (df_slice.shape[0] < 3):
        return
    
    idx = df_slice.index.unique()
    
    df_slice.reset_index(inplace=True)
    
    anomalies_summary_irq, outlier_over_irq, outlier_under_irq = get_feature_anomalies( \
                df_slice, \
                func=is_anomaly_irq, \
                features=features, \
                thresh=irq_threshold)
    
    median_cost = df_slice[RPT_COL_PUB_COST].median()
    
    
    # include over/under outliers as desired
    is_outlier_irq = np.logical_or(
                        np.logical_and(want_outliers_over, outlier_over_irq),
                        np.logical_and(want_outliers_under, outlier_under_irq)
    )
    
    
    # ignore anomaly from low spend adgroups (greater than campaign median)
    is_outlier_irq = np.logical_and(is_outlier_irq, df_slice[RPT_COL_PUB_COST] > median_cost)
    
    if sum(is_outlier_irq) > 0:
        print(">>> ANOMALY", idx)
        print(anomalies_summary_irq)
        cols = [RPT_COL_CAMPAIGN, RPT_COL_PUB_COST, RPT_COL_CONV, RPT_COL_REVENUE] + features
        print(df_slice.loc[is_outlier_irq, cols])
        
    return is_outlier_irq

## Find ROAS Anomalies

print("input shape:", df_campaign_perf.shape)
df_anomalies = pd.DataFrame()

# annotate via Marin Dimensions
def rowFunc(row):
    return '{}: ROAS {:,.2f} is much lower than account avg {:,.2f}'.format(
        row[RPT_COL_ROAS], \
        row[RPT_COL_ROAS + '_median'],
        datetime.date.today()
    )

for account_idx in df_campaign_perf.index.unique():
    df_account = df_campaign_perf.loc[[account_idx]].copy()
    df_account[RPT_COL_ROAS + '_median'] = df_account[RPT_COL_ROAS].mean()
    df_account[BULK_COL_AUTOMATION_INFO] = np.nan
    # dump data used for anomaly detection
    print("checking account: ", account_idx, tableize(df_account))
    outliers = find_peer_anomaly(df_account, [RPT_COL_ROAS], irq_threshold=0.75, outliers_desired=(False,True))

    if outliers is not None and sum(outliers) > 0:
        df_outliers = df_account.loc[outliers].copy()
        df_outliers[BULK_COL_AUTOMATION_INFO] = df_outliers.apply(rowFunc, axis=1)
        print(df_outliers)
        df_anomalies = pd.concat([df_anomalies, df_outliers], axis=0)

## Prepare Output
if not df_anomalies.empty:
    print(tableize(df_anomalies))
    outputDf = df_anomalies[[RPT_COL_ACCOUNT, RPT_COL_CAMPAIGN, BULK_COL_AUTOMATION_INFO]]
else:
    print("No anomalies found!")
    outputDf = outputDf.iloc[0:0]