ASSOCIATION RULES MINING

Correlation Graph

Let's start by observing the correlation in our dataset thanks to python.

Thanks to this graph we can see strong correlations between the time spend to call and the charge. It seems logical, the more you call the more the charge will be expensive. We can also see that the churn part has little correlations with some other part like 'international plan', 'total day minutes', 'total day charge' and 'customer service calls'. So in the next part we are going to study that more precisely.

Association Rule Mining

What are the association rule mining?

An association rule mining allow us to figure out a pattern that occurs frequently in a dataset. This method is often used for basket data analysis, cross-marketing, catalog design, sale campaign analysis for example.

Why and how to use it?

Thanks to association rule, we will be able to identify underlying relations between different items and to do a correlation analysis of our data.

Different statistical algorithms have been developed to implement association rule mining, and Apriori is one such algorithm. Apriori algorithm is based on 3 keys metrics to consider when evaluating association rule :

- Support : Given an item A and an item B, it's the probability to have A and B on the same row

- Confidence : Given an item A and an item B, it's the probability that if A is present in a row then the B will be present also.

- Lift : Given two items, A and B, lift indicates whether there is a relationship between A and B, or whether the two items are occuring together in the same orders simply by chance.

Application

First of all we need to use descritize data to apply Apriori Algorithm. We had already done that during the preprocessing.

Then we have to transform data in order to be suitable to use our algorithm. Finally we can import a method which perform the apriori algorithm. The detail of these operations are below.

Results (International Plan/Churn)

This result show that in our dataset, False and no are on the same row 8 times out of 10. We can also said that the probability to stay subscribe when a customer don't have a international plan is very high (93%).

Results (Total time per day/Churn)

We see that on our fisrt result, the lift is smaller than 1. It's a negative correlation which means that the most of the time the customer who have a total time of call weak or normal keep their subscription. The two next result seem to prove this interpretation. However the support is really weak so we can't be really be sure about this interpretation.

Results (Customer service call/Churn)

The first and the last results are not revelant because of their weak measure. But with the second result we can say that a customer who call rarely the customer service will keep his subscription.

Conclusion :

Despite the results that we found, we can't take out a stronger association rule with this data.

The first reason may be that there isn't strong corellation with the colulmn churn. The results are better when we study two features with a good correlation. The proof just on the left shows when we applied the Apriori Algorithm for two features with a strong correlation : 'total day minutes' and 'total day charge' .

We see easily that the result are widely better there.

Another reason maybe the lack of churn in our dataset. There are only about 500 exemples with a churn against almost 3000 without churn. The results will be more accurate and more revelant if we had more data with churn

Google Sites

Report abuse