Swiggy Develops Multi-Task MLP Model with 350+ Features for Customer Lifetime Value Prediction

Predicting Customer Lifetime Value

Swiggy has implemented an in-house predicted lifetime value (pLTV) model to forecast the long-term value of new customers across its food delivery and Instamart quick commerce platforms. This model aims to generate a value signal before a customer's initial order, allowing Swiggy to optimize advertising bids using predicted value instead of short-term conversion metrics.

Feature Engineering and Cold Start Solution

The model incorporates more than 350 features, categorized by acquisition channel, device and fraud signals, geographic behavior, complaint history, Instamart category affinity, order behavior, payment patterns, and socioeconomic indicators. These features are derived from information available prior to a customer's first order, addressing the cold start problem for advertising systems.

Multi-Task Learning for Efficiency and Accuracy

Swiggy's model utilizes a multilayer perceptron with three shared hidden layers and task-specific heads for Food and Instamart. An auxiliary prediction task for order count was introduced alongside lifetime value. This four-head model achieved marginally better results while reducing the parameter count from 363,000 to 135,000, a 63% reduction.

Evaluation Methodology

Due to a high number of zero-value users and a long tail of customer value, Swiggy found conventional regression metrics like mean absolute error insufficient. Instead, the model's performance is evaluated by its ability to correctly rank customers by value. This involves dividing predicted and actual values into ten quantiles and measuring the alignment between predicted and actual customer groups.