BUYER CONVERSION MODEL
WHY
​​
-
ASPAC JJ Vision Care has the need to apply machine learning for predictive and prescriptive analytics - identifying customer segments with high likelihood of conversion and optimal customer engagement strategies – in order to increase customer conversion
​
​
​
-
The Buyer Conversion Model produces recommendations for target group, content, channel and frequency. To do this, machine learning techniques are applied on data including customer demographic and behavioral traits from Salesforce CRM (e.g., age, gender, months since registration), customer transactions, Salesforce Marketing Cloud (e.g., interactions through push notifications, SMS, email), MyAcuvue app activity data (volume & frequency of logins, pages viewed). Below is the overview diagram of the modelling phases.
HOW
WHERE
J&J
WHERE
WHAT
Machine learning predictive model utilising Pyspark
J&J
(2021)
This is in line with some challenges faced by JJVC at present:
-
Significant % of registered users have no purchases​
-
Lack of customized marketing campaigns targeting inactive buyers​
Benefits
-
Increased Profit: Increase in revenue from more conversions as well as decrease in cost of customer acquisition will result in higher profit. By engaging customers with higher likelihood to convert, this will allow for top-line growth with reduction & optimization of marketing spend
-
Increased Customer Engagement: Optimized targeting would result in reduced customer fatigue and higher engagement, hence increasing customer loyalty in the long run
-
Scalability & Re-usability: This model may be refreshed monthly, hence it can provide up-to-date insights. The code is optimized to be scaled to other markets and other use cases requiring customer insights too
​
Methodology
Phase 1 – Customer Segmentation
Customer segmentation is done using K-means clustering based on (A) Volume and amount of purchase (B) Months since last purchase and (C) Membership points. These customer segments will represent “non-buyers”, “new buyers”, “frequent buyers” and more.
Phase 2 – Buyer Conversion Recommendation Model
The 2 segments, “buyers with some purchases” and “non-buyers” from the previous phase are compared to determine the relative importance of features which influences whether a user is a buyer.
Light GBM algorithm is used for modelling and here is a brief summary of how it works:
We examine the SHAP values which represents the contribution of each feature to the prediction. Hence, here are the recommendations for Target Group & Content:
For each feature, we can obtain deeper insights through partial dependence plots, for instance, age:
Below are the recommendations for channel and frequency. We found that push notification & inbox is the most effective channel, followed by inbox, SMS, push notification only then email.
In addition, we observe the optimal number of interactions as well, such as less than 5 messages for inbox and less than 3 for email.
