Making Recommendations without Data
summary
Recommendation system for Six Tribes bootstraps an interest space from Facebook posts and likes, iPhone music and photo/GPS location rather than historical behavioral logs, starting with 10 users for the first 4 months and about 100 users after launch. It uses prismatic and Alchemy APIs for keyword extraction, decays old plays and ignores the pushed U2 album, then runs a Elasticsearch-based similarity search with age/gender/location scoring heuristics, served by a Flask microservice and tuned by rapid user feedback; APIs prove useful but costly so some lists are hard‑coded.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:56Why recommendations are easy
- 3:22Challenge of interest spaces
- 5:50Extracting location and music preferences
- 8:01Using Facebook likes and posts to infer interests
- 10:27APIs impractical for production
- 11:31Building user profiles
- 12:57Designing tribes with filters
- 14:42Scoring with Elasticsearch
- 16:03Microservice for real-time recommendations
- 17:47Deployment and improvement
- 19:53Lessons from minimal data
- 22:20Q&A
related talks
43mBuilding your first recommender
48mCollective Intelligence
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
23mFrom Data to Recommendations: Building an Intelligent System with ROR
41mRecommender Systems: Crafting Effective Personalised Experiences
39mRecommendations at XING: From theory to production
55mBuilding a Recommendation Engine with Machine Learning Techniques