Decyphering Recipes: Mapping ontologies for personalization
2 likes220 views
This document discusses how Gousto, an online recipe box service, uses Neo4j and a recipe ontology graph to power personalized recipe recommendations. It outlines Gousto's data challenges with siloed databases and lack of customer journey tracking. The ontology graph captures recipe attributes and ingredient relationships to calculate similarity scores beyond basic ingredients. This allows personalizing recommendations based on cuisines, dish types, and customer goals. Benchmarking with human ratings helps validate the model, while future uses could include substitution for diets and AI-generated recipes.
Decyphering Recipes: Mapping ontologies for personalization
- 1. De-cyphering recipes with neo4j Irene Iriarte Carretero GraphConnect NY 2017
- 3. - UK-based online recipe box service - Leaders in choice, offering 22 meals a week - Deliver exactly proportioned ingredients with step-by-step recipe cards in 2-3 days - No planning, no supermarkets and no food waste – you just cook and eat!
- 4. Data Science Team Dejan Marc Manuel Cat Irene
- 5. MySQL Data Warehouse Excel reports Google Analytics - Started our data journey early on - Many external data sources - Lots of ad-hoc analyses - Lack of being able to track customer journeys
- 10. Personalisation
- 11. Personalisation
- 12. What does personalisation look like? “For you”
- 13. What does personalisation look like?
- 14. What does personalisation look like? - Default customers – on a subscription and automatically get allocated recipes - Better retention when they are happy with their allocated choices - Personalise their recipes
- 15. To recommend recipes to our customers: Collaborative Filtering
- 16. To recommend recipes to our customers: Collaborative Filtering
- 17. To recommend recipes to our customers: Collaborative Filtering
- 18. To recommend recipes to our customers: Collaborative Filtering Pros: - Several data points for each recommendation Cons: - Cold Start - Sparsity
- 19. To recommend recipes to our customers: Content- based Filtering
- 20. To recommend recipes to our customers: Content- based Filtering
- 21. To recommend recipes to our customers: Content- based Filtering Pros: - Recipe cold start is not a problem Cons: - No information sharing across users - Serendipity
- 22. To recommend recipes to our customers: Collaborative Filtering Content- based Filtering+
- 23. To recommend recipes to our customers: Hybrid
- 24. To recommend recipes to our customers: Hybrid Hybrid model recommender developed by Maciej Kula (Lyst) https://arxiv.org/pdf/1507.08439.pdf
- 25. How to overcome cold start problem For new users: recipe battles vs
- 26. How to overcome cold start problem For new recipes: similarity from recipe properties -Italian -Pasta based -Non-vegetarian
- 27. Challenge
- 28. Brazilian Black Beans and Limey Chicken with Rice Cambodian Chicken Samla Curry with Rice
- 29. Brazilian Black Beans and Limey Chicken with Rice Cambodian Chicken Samla Curry with Rice ADVENTURE
- 30. Brazilian Black Beans and Limey Chicken with Rice Cambodian Chicken Samla Curry with Rice WHOLESOME
- 31. Beany Tacos with Sweetcorn and Chorizo and Sweet Potato Fries Pork, Pineapple and Red Onion Tacos
- 32. Beany Tacos with Sweetcorn and Chorizo and Sweet Potato Fries Pork, Pineapple and Red Onion Tacos KID-FRIENDLY
- 33. Beany Tacos with Sweetcorn and Chorizo and Sweet Potato Fries Pork, Pineapple and Red Onion Tacos CONVENIENCE
- 34. Recipe Similarity - Ingredients in common offer basic recipe similarity score - Not good enough for our purposes - We want to take into account subjective aspects: • Cuisines • Type of dishes • Presentation • Why is the customer using our service?
- 35. Solution
- 36. Ontology in neo4j Ontology: is a formal naming and definition of the types, properties, and interrelationships of the entities that fundamentally exist for a particular domain
- 37. Why neo4j? - Recipe & ingredient attributes are highly interconnected - In order to capture the different point of views, it was vital that we were able to easily explore relations between the data
- 38. Why neo4j? - It allowed for flexibility in terms of describing recipe and ingredients attributes - We can easily create inferences from data attributes and relations
- 39. Calculating Similarities - Supervised Use tagged data to calculate weights of different attribute to fit to training data - Unsupervised Use tagged data to validate our unsupervised model - We will be using what customers are and are not ordering as feedback
- 40. Benchmarking - In order to benchmark our similarity scores with those coming from humans - We set up a RecipeBot on Slack that asked Gousto employees to rate the similarity of certain recipes - Gathered thousands of answers
- 41. Future
- 42. Future - We could take personalisation one step further with Snowplow data - Where do people click? - What does this tell us about the user?
- 43. Future - Ontology could help us when substituting problematic ingredients for dietary requirements - AI recipe development