Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value

Swiggy, a food delivery and quick commerce company, has created an in-house model to forecast the long-term worth of new customers. This predicted lifetime value (pLTV) model can be utilized for ad optimization prior to a customer's first purchase, rather than depending on short-term conversion metrics. The model architecture, as detailed in Swiggy's blog post, incorporates over 350 features from various sources such as acquisition channels, device and fraud signals, geographic behavior, customer complaints, Instamart category preferences, order behavior, payment habits, and socioeconomic data.

The challenge faced by Swiggy was dealing with a limited number of early signals and a highly uneven target distribution. To tackle this, the model employs more than 350 features that are available prior to a customer's initial order, addressing the "cold start" problem for advertising systems. Staff Data Scientist at Swiggy, Soumyajyoti Banerjee, revealed that integrating a secondary prediction task diminished the model size by 63% while elevating its accuracy.

The model utilizes a relatively uncomplicated multilayer perceptron rather than a more intricate structure. It contains three shared hidden layers that learn representations across the prediction tasks, followed by task-specific heads for Food and Instamart. Furthermore, the model integrates order count as an additional prediction task alongside lifetime value. This four-head model achieved marginally better outcomes while decreasing the parameter count from 363,000 to 135,000, a significant 63% reduction.

To evaluate the model, Swiggy opted for conventional regression metrics, such as mean absolute error and mean absolute percentage error, which were deemed insufficient due to the abundance of zero-value users and the long tail of customer value. Instead, the company assessed the model's capacity to rank customers by value, dividing predicted and actual values into ten quantiles and measuring the frequency of users falling into the corresponding predicted and actual groups.

In production, Swiggy employs the pLTV signal in conjunction with Google's target return on ad spend (tROAS) bidding. The objective is to rank customers based on their predicted value and enable the advertising system to differentiate bids depending on that ranking. Swiggy also conducted A/B tests comparing their in-house model with a third-party pLTV platform.

The results showed that Swiggy's model resulted in higher customer retention and gross order value per acquired user, without incurring additional platform costs, while the external platform attracted more users per dollar spent. The company's next objective is to transition from point estimates to probabilistic pLTV predictions, aiming to represent uncertainty in customer value and integrate it into bidding decisions.