A content-based recommender system for financial news
summary
Content-based recommender for financial news that solves cold-start by modeling clicks as a per-user classification problem using implicit feedback and balanced negative sampling. Pipeline computes a 14k sparse features vector from article metadata plus NLP enrichments (spaCy, fastText), user reading history, and explicit user–article overlap (authors/tags) signals. Production uses a gradient-boosted decision tree with partial_fit_grow (daily add-trees on seven-day windows), batch morning predictions, cached scores, and ongoing online evaluation of accuracy, diversity and coverage.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:08Cold start problem
- 3:19Training data collection
- 5:29Designing features
- 8:19User features without demographics
- 10:29Combining features
- 11:13Model selection
- 13:22Gradient boosted decision trees
- 14:29Evaluating ranking metrics
- 17:22Facebook ensemble model
- 18:28Training data recency
- 19:53Feature ablation results
- 21:17Key findings
- 21:38Implementation pipeline
- 23:04FD website recommendations
- 24:31Online testing beyond clicks
- 25:54Coverage and diversity
- 28:01Lessons and conclusion
related talks
29mRecommender Systems for Personalized Financial Advice
43mBuilding your first recommender
20mBuilding Recommender system
35mEnhancing User Engagement Through Personalized Recommendation at Dow Jones
30mParticle Swarm Classification for Next-Gen Recommendation Engines
03mServing Billions of Personalized News Feeds with AI
36mMachine Learning in News Media: Case study 24sata.hr
31mNLP text recommender system journey to automated training pipeline with Spark and Sagemaker