In [1]:
import warnings
warnings.filterwarnings('ignore')
import numpy as np
import pandas as pd
%matplotlib inline
import matplotlib.pyplot as plt
Then, load the data (takes a few moments):
In [2]:
# Load data
df = pd.read_csv("./aws-data/firence_foreigners_3days_past_future.csv", header=None)
df.columns = ['lat', 'lon', 'date_time_m', 'home_region', 'cust_id', 'in_florence']
df.head()
# np.max(data.date_time_m) # max date : '2016-09-30
# np.min(data.date_time_m) # min date: 2016-06-07
Out[2]:
This create a calls-per-person frequency distribution, which is the first thing we want to see.
In [3]:
fr = df['cust_id'].value_counts().to_frame()['cust_id'].value_counts().to_frame()
fr.columns = ['frequency']
fr.index.name = 'calls'
fr.reset_index(inplace=True)
fr = fr.sort_values('calls')
fr['cumulative'] = fr['frequency'].cumsum()/fr['frequency'].sum()
fr.head()
Out[3]:
Plot this distribution. This shows that 19344 people made 1 call over the 4 months, 36466 people made 2 calls over the 4 months, 41900 people made 3 calls over the 4 months, etc.
In [4]:
fr.plot(x='calls', y='frequency', style='o-', logx=True, figsize = (10, 10))
plt.axvline(5,ls='dotted')
plt.ylabel('Number of people')
plt.title('Number of people placing or receiving x number of calls over 4 months')
Out[4]:
It might be more helpful to look at a cumulative distribution curve, from which we can read off quantiles (e.g., this percentage of the people in the data set had x or more calls, x or fewer calls). Specifically, 10% of people have 3 or fewer calls over the entire period, 25% have 7 of fewer, 33% have 10 or fewer, 50% have 17 of fewer calls, etc., all the way up to 90% of people having 76 or fewer calls.
In [5]:
fr.plot(x='calls', y='cumulative', style='o-', logx=True, figsize = (10, 10))
plt.axhline(1.0,ls='dotted',lw=.5)
plt.axhline(.90,ls='dotted',lw=.5)
plt.axhline(.75,ls='dotted',lw=.5)
plt.axhline(.67,ls='dotted',lw=.5)
plt.axhline(.50,ls='dotted',lw=.5)
plt.axhline(.33,ls='dotted',lw=.5)
plt.axhline(.25,ls='dotted',lw=.5)
plt.axhline(.10,ls='dotted',lw=.5)
plt.axhline(0.0,ls='dotted',lw=.5)
plt.axvline(max(fr['calls'][fr['cumulative']<.90]),ls='dotted',lw=.5)
plt.ylabel('Cumulative fraction of people')
plt.title('Cumulative fraction of people placing or receiving x number of calls over 4 months')
Out[5]:
Now, we want to look at temporal data. First, convert the categorical date_time_m to a datetime object; then, extract the date component.
In [6]:
df['datetime'] = pd.to_datetime(df['date_time_m'], format='%Y-%m-%d %H:%M:%S')
df['date'] = df['datetime'].dt.floor('d') # Faster than df['datetime'].dt.date
In [7]:
df2 = df.groupby(['cust_id','date']).size().to_frame()
df2.columns = ['count']
df2.index.name = 'date'
df2.reset_index(inplace=True)
df2.head(20)
Out[7]:
In [9]:
df3 = (df2.groupby('cust_id')['date'].max() - df2.groupby('cust_id')['date'].min()).to_frame()
df3['calls'] = df2.groupby('cust_id')['count'].sum()
df3.columns = ['days','calls']
df3['days'] = df3['days'].dt.days
df3.head()
Out[9]:
In [17]:
plt.scatter(np.log(df3['days']), np.log(df3['calls']))
plt.show()
In [ ]:
In [41]:
Out[41]:
In [77]:
fr.plot(x='calls', y='freq', style='o', logx=True, logy=True)
Out[77]:
In [97]:
x=np.log(fr['calls'])
y=np.log(1-fr['freq'].cumsum()/fr['freq'].sum())
plt.plot(x, y, 'r-')
Out[97]:
In [ ]:
# How many home_Regions
np.count_nonzero(data['home_region'].unique())
In [ ]:
# How many customers
np.count_nonzero(data['cust_id'].unique())
In [ ]:
# How many Nulls are there in the customer ID column?
df['cust_id'].isnull().sum()
In [ ]:
# How many missing data are there in the customer ID?
len(df['cust_id']) - df['cust_id'].count()
In [3]:
df['cust_id'].unique()
Out[3]:
In [17]:
data_italians = pd.read_csv("./aws-data/firence_italians_3days_past_future_sample_1K_custs.csv", header=None)
data_italians.columns = ['lat', 'lon', 'date_time_m', 'home_region', 'cust_id', 'in_florence']
regions = np.array(data_italians['home_region'].unique())
regions
Out[17]:
In [22]:
'Sardegna' in data['home_region']
Out[22]: