We use content-based filtering (item attributes), popularity/trending, and demographic or onboarding data. For new items, we rely on metadata and similar-item logic. As users interact, we gradually blend in collaborative signals. Hybrid approaches typically solve cold-start within a few interactions.

"Customers who bought this" is item-based collaborative filtering: find similar items by co-purchase patterns. User-based CF finds similar users and recommends what they liked. Both use interaction data; item-based scales better. We often combine both with matrix factorization for richer personalization.

Yes. With sparse data, we use content-based (item features), session-based (current browse behavior), or popularity/trending. We've built engines that perform well with as few as 10–20 interactions per user by leveraging item metadata and transfer learning.

Exploitation shows what the model thinks is best; exploration tries new items to gather data. We use bandit algorithms (e.g., Thompson sampling) or inject a small percentage of random/diverse items. The exploration rate can decay over time or adapt based on engagement.

Trending works for cold-start users, new visitors, or when personalization isn't ready. It's simpler and faster to implement. We often blend trending into personalized feeds—e.g., 20% trending, 80% personalized—to balance discovery with relevance.

We add diversity and serendipity: diversify results by category, inject unexpected items, or use MMR. We avoid over-weighting past clicks and balance short-term engagement with long-term satisfaction. Regular audits and A/B tests help detect and fix filter bubbles.

Batch precomputes recommendations (e.g., nightly) and serves from cache—good for email and homepage. Real-time computes on each request using latest session data—essential for "add to cart" suggestions. We often use hybrid: batch for base scores, real-time for session boost.