Data Science for Dinner: Recommending Recipes
summary
Gousto implements a pragmatic hybrid recommender: a failed attempt with LightFM matrix factorization leads to a content-based system built on a recipe taxonomy in neo4j graph, Jaccard similarity and image autoencoder features to solve a 40-item weekly menu with ~20–25% new recipes, plus a simple collaborative layer from k‑means (five clusters) to add variety. Pipelines run weekly via Airflow on Docker/AWS Batch. Content ranking beats random, LightFM proves too heavy, the hybrid improves discovery and conversion despite a drop in a basket‑match metric.
AI-written summary · about
chapters (25)
- 0:00Goose and meal kits intro
- 1:57Customer journey and data science
- 3:48Why recommendations matter
- 5:54Early matrix factorization
- 6:57Recipe recommendation challenges
- 8:44Recipe taxonomy with graphs
- 10:30Recipe similarity via attributes
- 12:40Visual similarity in recipes
- 13:25Content-based recommender
- 14:51Chicken conundrum explained
- 15:53Collaborative filtering intro
- 17:19Combining content and collaborative
- 19:32Balancing metrics and UX
- 20:13Tech infra for recommendations
- 21:14Future plans for rec systems
- 23:01UI/UX impact on conversion
- 24:52Key learnings from journey
- 27:06Q&A: Recipe filters and preferences
- 28:48Q&A: Early warehouse optimization
- 29:31Q&A: User feedback and rules
- 30:14Q&A: Behavioral segmentation
- 31:08Q&A: Weekly planning constraints
- 32:15Q&A: Evaluating rec impact
- 32:37Q&A: Metrics and recipe properties
- 33:42Q&A: Generative recipe dev
related talks
27mOur experiments with food recommendations
31mCreating recommender systems with Spark, Scala and Prediction.Io
37mFood is personal! Design and impact of personalized recommendations in e-grocery
45mContent recommendation at iFood scale
39mDiscover your latent food graph with this 1...
18mLearning How To Cook...
43mBuilding your first recommender
35mHybrid Recommender Systems in Python