Predictive model for Customer Loyalty

100 %
0 %
Information about Predictive model for Customer Loyalty

Published on November 21, 2016

Author: VamshiVennamaneni


1. Vamshi Vennameneni Sai Rajendra Wuppalapati Valli S Vadlamani Gautam Gite Customer Loyalty – Retail A Predictive Model CS 593 - Data Mining II Prof. Dehnad Khasha

2. Problem Statement Anticipated Business Value Data pertaining to a retail data set predicting customer loyalty. The goal is to get best possible classification accuracy in a data set with a binary target and several features. Active Customer Significance Value 1 LOYAL Value 0 NOT LOYAL

3. Data :  All data is anonymized.  Only behavioral attributes (purchase behavior)  26K observation for training data.  No Data Dictionary provided.  257 columns with 1. 41 Trans attributes 2. 160 Food attributes 3. 48 Promotional attributes

4. Data Analysis Dimension Reduction Logistic Regression Building Model: First Steps From the data, we identify that dependent variable is “Active_Customer” The model is a logistic regression function of the 256 variables As the number of variables are too high, dimensionality reduction techniques like PCA should be employed The Eigen vectors from PCA should be the input for logistic regression

5. Data Analysis Dimension Reduction Logistic Regression

6. Exploratory Data Analysis

7. Exploratory Data Analysis

8. Exploratory Data Analysis

9. Dimension Reduction Techniques • Business / Domain Knowledge Application Specific knowledge on the business would help us to eliminate a few variables or assign weightages to prioritize the important ones. In this scenario, lack of data dictionary does not give information on the type of variable. • Removed duplicate values., since no value is added to the model using the same. • Used PCA to reduce the number of variables to be used for analysis.

10. Dimension Reduction Techniques  Principal Component Analysis In this technique, variables are transformed into a new set of variables, which are linear combination of original variables. These new set of variables are known as principal components On running PCA for our data, 90% of variance explained by more than 10 components Scree plot is nearly flat after 10 components

11. Dimension Reduction Techniques  Principal Component Analysis

12. Logistic Regression  Done with top 10 Principal components

13. Logistic continued

14. Feature selection Random Forest  We used Random Forests to select the best features.  Random forest gives ‘Feature_importance_score’ to each variable  Variable Selection : Features with score greater than mean of scores is selected for further analysis

15. Random Forest - Scores

16. Bagging  Used the reduced dataset with important feature to run a Bagging Classifier.

17. Ada Boosting  Results of Ada Boosting

18. Ada Boost

19. Comparison of Classifiers Accuracy Classifier Ada-Boost 66.8% XG Boost 67.1%

20. Ideas tried to improve the model  Included first few PCs as extra features  Tried replacing missing values with median and mode  Modified algorithm parameters of the classifiers to get the best model  Tried a combined classifier which is a majority vote of all the strong classifiers

21. Conclusions  ‘Trans9’, ‘Trans4’ and ‘Trans8’ are the most important features affecting customer loyalty.  Ada Boost has model performed well  Identifying the loyal customers would help in increasing the customer lifetime value by introducing products like loyalty cards or promotions customized to loyal customers  Identifying the demographics of the customers who are not very loyal would help the company in designing their marketing strategy

Add a comment