Optimizing Merchandising: Machine Learning for Real-Time App Experience and LTV Maximization

At PhonePe Limited (“PhonePe”), the top banner on our platform’s homepage (the “masthead”) isn’t just an ad slot—it’s the important real estate we have. However, delivering real-time, personalized recommendations at a scale of hundreds of millions of users presents a complex problem to solve for two reasons:

• Operational Scalability: Relying on manual intervention to curate category recommendations is operationally intensive and may result in sub-optimal user experiences.
• Contextual Relevance: Recommendations should be relevant considering the user’s specific journey within the PhonePe ecosystem.

Whether a user is brand new or a power user, the goal is to show the single next best action to take.

To achieve this goal, we built a sophisticated ML-driven decision-making system. Regardless of where a customer is in their PhonePe journey, the system aims to show the next best suggested action by showing the single highest-value opportunity for each individual user.

This system optimizes the user lifecycle across two core journeys:

1. Acquire: Discovering and introducing users to new categories.
2. Repeat: Nudging users to transact again at exactly the right time

As the ML model learns to prioritize the propensity of a relevant category, it simultaneously enhances the Lifetime Value (LTV) of the customer.

Here is how we built our Category Recommendation Model to predict the next best action for every user at scale, improving both the business metrics such as Categories Per Customer (CPC) and LTV.

From Promotion to Recommendation: Moving Beyond Manual Bias

While manual merchandising is effective for solving early-stage cold start problems, it cannot scale to hundreds of millions users. However, at our current scale, relying on human curation will inevitably create bottlenecks. We identified three major limitations with the manual approach 

• Operational Inefficiency: Campaign collation, execution, and prioritization often required significant manual overhead.
• Sub-optimal Personalization: Manual prioritization can lead to fragmented user experiences, where users might not see the product categories most relevant to their immediate needs.
• Missed Revenue: Focusing solely on a few top-performing categories can lead to missed revenue opportunities in niche or emerging categories.

The Solution: A Self-Improving Engine

We developed a system that operates as a Self-Improving Recommendation Engine. Think of it as below:

Ingestion: The model consumes massive amounts of data, analyzing over 5,000 distinct user features and quantified transaction behavioral patterns.

Ranking: System ranks selected business categories for our entire user base.

Feedback: The user’s reaction, whether they convert to the recommended category or ignore the recommendation is captured. This signal is fed back into the system to refine the model for the next cycle.

Infrastructure Components

The workflow leverages several key components within the PhonePe ecosystem:

• Raw Data Ingestion: The central storage where all transactional and demographic data is collected.
• Feature Store: A repository storing over 5,000 pre-computed user features, aggregated data ensuring consistency between training and serving.
• Distributed Compute Cluster: A large-scale cluster dedicated to ML computation, model training, and scoring.
• Model Serving Layer: The service responsible for hosting models and serving recommendations to the app.

Under the Hood: The Modeling Workflow

Our recommendation engine runs a multi-stage pipeline on our compute cluster.

Phase 1: Feature Engineering and Target Definition

Stored features from Feature Store are pre-computed at a set frequency to ensure low latency during retrieval.

We define a “success event” based on the final completed transaction. Every success or failure is recorded and fed back into the system, allowing the model to learn and improve continuously.

Phase 2: Model Training

The acquisition and repeat ML models are designed to learn complex, non-linear patterns from the data. The main technical challenges were:

• Generalization at Scale: Extracting robust patterns from a massive PhonePe dataset representing our diverse user base.
• Training Time: Ensuring the model training completes within an acceptable time window.
• Model Interpretability: Maintaining explainability so we understand why a specific recommendation was made.
  
  The model training process is designed to prevent bias toward only the most popular categories.

Phase 3: Scoring & Ranking

Once the model is trained, we run a scoring job that executes the following steps:

1. Transform: The model outputs a probability score for each user across all potential categories.
2. Filter: We apply critical business rules to filter out specific users (e.g., risk checks or fraud prevention).
3. Rank: We rank the remaining recommendations based on a utility function that balances Propensity and LTV.
4. Serve: The final ranked list is pushed to the serving layer for the app to consume.

Validation: Measuring Success

We validate the model using a strict and fair A/B testing framework, comparing a Treatment Group (TG) against a Control Group (CG).

The incremental impact is measured using two key metrics:

1. New Acquisitions & Conversions: The uplift in users converting in a category covering both repeat and acquisition journeys
2. Incremental Revenue: The additional value generated specifically due to the improved relevance of the recommendations.

This comparison ensures that the uplift we see is genuinely driven by merchandising intelligence and not just organic growth. 

To prove the system’s effectiveness further, we subjected the model to strict evaluation protocols. We compared its performance against both random recommendations and traditional rule-based methods to ensure it was providing a measurable uplift in conversions and LTV

Closing Thoughts

By transitioning to this ML-driven ecosystem, PhonePe is curating a personalized journey for every user, one interaction at a time. Moving from manual curation to an ML-driven ecosystem was a necessity, not just an optimization. At PhonePe’s scale, only a self-improving engine can handle the complexity of millions of users with distinct journeys.

By automating the feedback loop between user action and model training, we have solved the dual challenge of operational scalability, real-time relevance and LTV maximisation