Building a recommendation system for EPAM systems
summary
Enterprise video platform for thousands of employees uses a hybrid recommender: candidate generation with LightFM over a ~100,000-video corpus (≈300 candidates/user), a LightGBM LambdaRank reranker, and distilled BERT embeddings for content similarity. Targets label relevance by watched fraction (e.g., 3h of 5h → relevant; 8m of 10m for short videos) and weight recency; pipelines run weekly in Airflow with real‑time triggers via Kafka. Metrics include watch time from recommendations, launches long CTR (80% watch threshold); tooling: Amplitude, Power BI, Metabase. Cold‑start uses popularity and similarity; system is live and improving.
AI-written summary · about
chapters (10)
- 0:00Recommendation system intro
- 3:16Business goals and context
- 8:19Data collection and modeling
- 11:15Hybrid recommendation pipeline
- 14:27Cold start and context
- 17:18Pipeline orchestration
- 18:42Quality measurement
- 21:14Key lessons
- 22:58Other EPAM use cases
- 24:53Q&A
related talks
23mML-Powered Search & Recommendation 101
29mBuilding Generative Content Understanding Systems
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
43mRecommendation Systems in Market Research
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
43mBuilding your first recommender
30mHybrid Recommendation Systems using Probabilistic Graphical Models