Swiggy replaced its previous lifetime‑value predictor with a multi‑task multilayer perceptron (MLP) that consumes more than 350 pre‑order features and adds order count as an auxiliary output. The auxiliary task shrank the model by 63% and lifted prediction quality, feeding a Google Target ROAS bidding loop for customer acquisition.
Architecture Overview
The new model treats food orders and Instamart orders as separate but related tasks within a single MLP. Over 350 engineered features—derived before an order is placed—are fed into the shared hidden layers. A secondary head predicts order count, which acts as a regularizer and reduces the total parameter count by roughly two‑thirds while still improving the primary pLTV estimate.
Operational Implications
Training now requires a joint loss that balances the pLTV and order‑count objectives, which may affect data pipeline design and experiment tracking. The smaller parameter footprint lowers GPU/CPU demand for both training and inference, potentially reducing cloud spend and latency in real‑time scoring. Because the pLTV signal is consumed by Google Target ROAS bidding, the model must expose a stable API that downstream ad‑tech can poll or receive pushes from, demanding reliable versioning and monitoring to avoid bidding anomalies.
Security and Data Considerations
Handling >350 customer‑derived features means the feature store must enforce access controls and audit trails to protect personally identifiable information. Any pipeline that exports the pLTV score to an external bidding platform should ensure data is transmitted over encrypted channels and that only the minimal necessary fields are shared.
Related CloudNinjas coverage: DevOps.
What This Means For Practitioners
- Consider multi‑task learning when a related auxiliary prediction can act as a regularizer and shrink model size.
- Validate that feature engineering pipelines can sustain a high‑dimensional input set without compromising latency.
- Integrate model serving with external bidding or decision systems using robust, versioned APIs.
- Implement strict data‑handling policies around the large feature set, especially when exporting signals to third‑party services.