Extending bandicoot

The user object

The User object is composed of a list of records and, optionally, a dictionary of attributes.


A record is stored in the class Record:

name type description
interaction string [required] 'call', 'text'
direction string whether the user was called/texted ('in') or was the one calling/texting ('out')
correspondent_id string identifier of the correspondent
datetime datetime timestamp of the record
call_duration interaction duration of the call in seconds or None for texts
position Position() a position object Position(antenna=13084)

Records are stored as a list, and can be accessed or modified with the property User.records.

User attributes

User attributes are loaded at the same time as the records. Attributes are stored in a dictionary that can be access by User.attributes:

>>> user.attributes['age'] = 42
>>> user.attributes['likes_trains'] = True

Object attributes

Object attributes are created by bandicoot when the user's records are loaded:

keys type description
has_call bool whether call records have been loaded
has_text bool whether text records have been loaded
has_antennas bool whether antennas have been loaded
has_recharges bool whether recharges have been loaded
has_gps bool whether gps locations have been loaded
start_time datetime time of the first record
end_time datetime time of the last record
antennas dict dictionary of antennas with antenna_id as keys and latlon tuples
home string the position (antenna id) the user spends the most time at during the night. Computed using recompute_home()

Writing a new indicator

A lot of the complexity of bandicoot is hidden from the user when writing a new indicator. For example, let's look at the method balance_of_contacts():

from bandicoot.helper.maths import summary_stats

def balance_of_contacts(records):
    Computes the balance of all interactions. For every tie, the balance is the
    number of outgoing interactions divided by the total number of interactions.

    counter_out = defaultdict(int)
    counter = defaultdict(int)

    for r in records:
        if r.direction == 'out':
            counter_out[r.correspondent_id] += 1
        counter[r.correspondent_id] += 1

    balance = [float(counter_out[c]) / float(counter[c]) for c in counter]

    return summary_stats(balance)

bandicoot's @grouping decorator manages the interaction and groupby keywords for you. It selects the right records (e.g. only calls) and groups them (e.g. by week). By default interaction=['call','text'] but this can be redefined in the decorator @grouping(interaction='call'). The function balance_of_contacts is then called for each group of records and the results are combined.

In this function, records is thus a subset of B.records (e.g. only the calls in a specific week). records is equal to B.records if the function is called with groupby='week' and interaction=['callandtext'].


The function executes the following operations:

  1. First, we initialize two empty int dictionaries using defaultdict from the collections module.
  2. The for loop then goes over each record passed by the decorator. It counts the total number of interactions and the number of outgoing interactions per contacts.
  3. We then compute, for each contact, the balance of interactions. Note that counter_out is a defaultdict, and counter_out[c] will return 0 even if c is not in the dictionary.
  4. balance is a list of the balance of interaction with each contact. We thus pass it to bandicoot's summary_stats() which will return the mean and std if summary=default; the mean, std, median, min, max, kurtosis, skewness if summary=extended; and the full distribution if summary=None.

Indicators using @grouping can return either a number (simply return the value) or a distribution (by calling summary_stats as shown); bandicoot automatically takes both values into account. For example, number_of_contacts() returns only one number.

Accessing the User object

A function to compute a new indicator might need to access more than just the list of records. A function might, for example, need to be able to access the GPS coordinate of an antenna or the first record we have available for this user. The method can ask the decorator to also pass the full user object using @grouping(user_kwd=True). It can then access all the records (user.records), the list of antennas (user.antennas), or other properties (see Object attributes).

def my_indicator(records, user):

Integrating your indicator

First, add it to bandicoot's test suite. bandicoot puts a strong emphasis on the correctness and consistency of its indicators. We thus require the values to be manually computed for the sample users located in bandicoot/tests/samples/manual/. These manually computed value can then be added to the JSON file also located in bandicoot/tests/samples/manual/ and tested using:

nosetests -w bandicoot/tests -v

The new metric can be integrated to the default bandicoot pipeline by adding it to all().


To run the unit tests with nose, use the following command:

nosetests -w bandicoot/tests -v

Note that running the tests requires additional modules: nose, numpy, scipy, and networkx.