This text provides a brief, practical overview of the recommendation strategies and personalization models I have been working on in Customer Marketing at Booking.com.

The Algorithm

In the modern e-commerce landscape, a broad inventory is both an asset and a liability. While a diverse catalog can satisfy a wide range of customer needs, it introduces significant friction: the “paradox of choice”. When customers are overwhelmed by options, the likelihood of abandonment increases. The challenge, therefore, is not just having the inventory, but surfacing the right match at the right time to the right customer. Building a robust recommendation system requires a layered strategy that evolves from simple heuristics to sophisticated, autonomous learning. Let’s consider them from the simplest to the most advanced, tracing the evolution of Recommendation Models.

The Global Popularity Top

When a customer is new, you have zero historical data. In the absence of any information, the safest bet is global popularity. This zero-factor model showcases what is broadly trending, providing a reliable starting point. As soon as we capture a single customer data point—such as a customer’s country of origin—we can pivot to a regional top, offering inventory that resonates with local preferences or seasonal relevance.

Behavioral Targeting

Once a customer begins to interact, we move from passive observation to active engagement. By identifying the customer’s focus area (e.g., a specific travel destination or product category), we can build popularity models within that category. This increases the relevance of our models, but as we progress into more relevant models, the data to build them is getting increasingly sparse.

Funnel Modeling

Impression-to-transaction model. This model analyzes the paths of successful conversions. If a significant cohort of users who viewed item X eventually purchased item Y, we create a recommendation bridge between these items, helping prospective buyers bridge the gap between interest and commitment.

Transaction-to-transaction model. For returning customers, we pivot to purchase history. By identifying sequential chains of behavior (e.g., purchasing X leads to the future need for Y, and eventually Z), we can predict and offer the next logical item in the customer’s lifecycle.

Feature Engineering

The model’s precision is heavily dependent on context. Incorporating customer factors such as average past transaction cost, allows the model to shift from “what is generally popular” to what is affordable and relevant for this particular customer.

Reinforcement Learning

Traditional static models struggle with ‘concept drift’—when user preferences change or market trends evolve. Reinforcement Learning (RL) allows the engine to learn in real-time through an ’explore/exploit’ loop. The system continuously experiments with new recommendation strategies (exploration) while prioritizing conversion based on proven successes (exploitation), ensuring the model remains adaptive to shifting dynamics without requiring constant manual retraining.

The Model of Models

Rather than attempting to manually determine which recommendation algorithm performs best, we can implement an automated decision layer using Multi-Armed Bandits. This architecture treats different recommendation models as competing ‘arms,’ dynamically reallocating traffic to the specific model currently delivering the highest performance. This effectively eliminates the overhead of manual model selection, allowing the engine to self-optimize autonomously.

Performance and Optimization Metrics

Moving beyond traditional A/B testing is essential for high-velocity systems. While A/B testing is effective for static validation, it is slow and requires significant, steady traffic. RL enables continuous, automated optimization. However, success depends heavily on metric alignment: whether you optimize for surface-level engagement (views/clicks) or core business outcomes (checkouts/margins) determines the behavior of your system.

Catalogs vs Chatbots

While conversational AI and chatbots are trending, they have never proved to be a successful sales interface. For large inventories where visual comparison is essential, catalogs remain superior.