Ingredients based - Recipe recommendation engine
- 1. Recipe Recommendation ENGINE By For: Data Science in the Wild - Spring 2020 Infinite Players
- 2. HOWDIDWE ARRIVEHERE? FINDCALLING 01 FINDDATA We found over 20+ usable data repositories & analyzed them 02 FINDRECIPE Upon cleaning, we tried various models to get the best possible result across various models. 03 COOK&SERVE! We collated the best results around an intuitive workflow 04 Infinite Players We foodies found that no-one has collated multiple datasets & made a good recipe recommendation engine
- 3. BACKGROUND Coronavirus has brought about an interesting fact about young working professionals and university students. All of them relied on take-aways and dine-ins skipping cooking. Most people do not know what to cook despite so many options available in the grocery store in this globalized world. Therefore, we want to answer: What do I cook, given I have these ingredients available? Food is the gateway to a new culture and so many cultures can be explored by what is in your fridge. Our engine enables this cross cultural exchange by telling you what is possible! Infinite Players
- 4. DATA-FROM THEWILD DATAsources: Infinite Players RECIPE_INGR_REVIEW (12K) YUMMLY CLEAN (6K) FOOD.COM DATA (231K) EPICURIOUS (20K) RECIPE_INGR (56K) Dataset Name SOURCE FIELDS TAKEN FOOD_COM LINK Ingredients, Recipe Name EPICURIOUS LINK Ingredients, Recipe Name, Ratings, Description YUMMLY CLEAN LINK Ingredients, Recipe Name, Cuisines RECIPE_INGR LINK Ingredients, Cuisines RECIPE_INGR_REV LINK Ingredients, Recipe Name, Ratings Total:~270k
- 5. DATA-FROM THEWILD FORINGREDIENTS We had ingredients ranging from ubiquitous wheat flour to the most exotic such as Saffron. In total, we had more than 100K+ ingredients in our datasets FORCUISINE We started with more than 35 unique cuisines, studied the differences, and commonalities among all. And finally mapped them to a superset of 7 FORUSERRATINGS Certain datasets had user reviews for the recipes. We utilised these reviews by defining a rating scale from 1-5 as a basis for our item-item based collaborative filtering model. Infinite Players FORREcipeNAMES All datasets have recipe names except recipe ingredients which has only cuisine names & ingredients. This is our desired output.
- 6. DATACLEANING- overview Basic Common text preprocessing techniques 01 No“quantities” “OZ”, “KG”, “POUNDS”, “TSP”, “LITTLE”, “PINCH” 02 Extractnouns POS tagging, extract ingredients from recipe instructions 03 Removerarewords AVG term frequency is 600, remove words occurring < 30 times 04 Infinite Players Iterate! 05 Continue cleaning as we see results
- 7. FEATUREENGINEERING WHY? ● Multiple datasets - different data formats ● Cleaning to 100% is hard, doesn’t scale to new data ● Ingredient related tokens > 2.5 MIL across 270K recipes Infinite Players
- 8. CUISINES- inthewild Which cuisine does the recipe belong to? Which cuisines should we narrow it down to? We tried to narrow down cuisines from this -> Infinite Players PROBLEM GROCERYINSPIRATION
- 9. CUISINESDEMYSTIFIED Infinite Players Confusion Matrix - Using Neural Network Upon refining further, we combined many cuisines to achieve the highest accuracy for our cuisine classifier while maintaining distinctive flavors and favoring numbers. Final List of Cuisines (7): American, Italian, European, Asian, Mexican, French, Indian
- 10. CUISINEs- AMERICA! Infinite Players Confusion Matrix - Using Ensembling methods Some patterns can be clearly noticed: French cuisine is very similar to America’s Cajun & Creole (Louisiana). Mexico influenced Texan food. Italy has a great influence on Northeast food with Pizza etc. European cuisine (Spanish,British & German) has a great influence too Asians & Indian cuisines have minimal collision This is really similar to the ethnicity of immigrants in the US *Indian & Mexican cuisines also share a lot of flavors.
- 11. HOWDOESITWORK? Infinite Players OUTPUT INPUT : Ingredients feature vectors EnsembleTechniques Neuralnetwork Logistic Regression K Neighbors Classifier Decision Tree Classifier Random Forest Classifier Layer 1: Linear + Leaky ReLU Layer 2: Linear + LeakyReLU + Dropout Layer 3: Linear + LeakyReLU + Dropout Layer 4: Linear + Softmax :Cuisine type for a list of ingredients
- 12. HOWTOGETARECOMMENDATION Infinite Players OUTPUTINPUT Ingredient (s) Choice of Cuisine (if any) COLLABFILTER CONTENTBASEDRECOMMENDATION AlternativeINPUT Name of Recipe ONE List of recipes according to user preferences & Another, List of recipes closest to the ingredients mentioned. Ingredient to features using word2vec model Cosine Similarity for calculating distance KNN with Means for recipe ratings Cosine Similarity for calculating distance ITEM - ITEM based filter Ingredients taken as input from recipe
- 13. MODEL- COLLABORATIVEFILTERing INPUT We build a recommender system in which the user inputs the ingredients they have on hand. Based on these inputs we will generate a short list of recipes that fit the users preferences. MODEL KNN with Means has been chosen for the recommender, which is a basic collaborative filtering algorithm, taking into account the mean ratings of each user. Compute the cosine similarity DATACLEANINg We use only one rating per user. Further we define a rating scale for the recipe.This is determined by the lowest and highest rating possible given by the users. Infinite Players EVALUATE We use the Surprise lib to test our recsys. Using cross validation we evaluate the model using a few metrics like MSE and RMSE. OUTPUT Finally get a recommendation based on an input string of ingredients
- 14. COLLABORATIVEFILTERing-RESULTS Infinite Players Input:User_ID,Ingredients User_id: 2043209 Ingredients: ‘chicken,egg,milk’ RECIPE INGREDIENTS INSTRUCTIONS Chicken Lasagna with White Sauce Recipe mozzarella,mushroom,milk,spinach,egg,ricotta,n… Preheat oven to 350 degrees F (175 degrees C).... Swedish Meatballs egg,milk,ground beef,cereal,onion,chicken,mush. Preheat oven to 350 degrees F (175 degrees C).... Mushroom Chicken Piccata Recipe flour,salt,paprika,egg,milk,chicken,butter,mus… In a shallow dish or bowl, mix together flour,... User_id: 700 Ingredients:: ‘Cheese,onion’ RECIPE INGREDIENTS INSTRUCTIONS Tuna Noodle Casserole II Recipe noodle,mushroom,milk,tuna,cheese,onion,potato,... In a large pot with boiling salted water cook ... Hamburger Cheese Bake Recipe pasta,ground beef,onion,tomato sauce,white sauce.. In a large pot cook with boiling salted water.. I
- 15. MODEL- CONTENTBASEDRECOMMENDER SYSTEM RAWINPUT Ingredient list for every recipe. All ingredients are kept through the pre- processing pipeline MODEL:WorD2vec Ingredients to features. 200 dimensions (with PCA, negligible difference in cuisine results, hence unused), context window of 12 (based on experiments), downsampling threshold of 1e-3 Recommendation Take input ingredients and use w2v on it. Use cosine similarity to compare distance with recipes in dataset Infinite Players evaluation Based on performance of downstream task: cuisine classification Eyeballing results of recipes recommended
- 16. CONTENTBASEDRECOMMENDER -RESULTS Infinite Players Bananas hot custard and bananas : 0.99 grilled bananas platanos asados : 0.9998 banana fudge pie : 0.9998 Chicken, spinach spinach chicken caesar wrap : 0.999906 szechuan shrimp stir fry : 0.999901 crunchy low fat summer chicken salad : 0.999901 Bananas, strawberry raspberry cream smoothie : 0.9999254 ensure smoothie : 0.9999226 got milk vanilla banana wellness smoothie : 0.9999215 Chicken, peas, coconut - “Asian” cuisine filter jade soup : 0.999903 malay style curry puff : 0.9999025 joanne s hawaiian chicken : 0.999898 Recipe similar to “healthy oatmeal raisin cookie muffins” oatmeal blueberry muffins : 0.9999919 1 gram fat pumpkin spice muffins low fat : 0.999988 whole wheat apple banana oatmeal muffins : 0.999984 ww core ginormously big breakfast cookie : 0.9999819 Ingredientsonly Ingredientsw/ cuisinefilter “Similarto” recipes
- 17. POST RECOMMENDATION /FUTURESCOPE Personalization As a part of improving the recommendations, users can be prompted to rate the recipes they were recommended. ! The tool can be integrated into smart devices such as refrigerators. Integrationintosmart devices Infinite Players ! User can evaluate recommendation quality to improve the models LEARNREGULAR !
- 18. INFINITEPLAYERS SALONI PRASHANT pj263@cornell.edu CM’21 DALE BHARAT bg445@cornell.edu CM’21 Infinite Players sg2452@cornell.edu CM’21 dm846@cornell.edu CS’20
- 20. APPENDIX Infinite Players American: Potato and Fennel Soup Hodge American: Banana-Chocolate Chip Cake With Peanut Butter Frosting Asian: Korean Marinated Beef Asian: Spicy Noodle Soup European: Lentil, Apple, and Turkey Wrap European: Ham Persillade with Mustard Potato Salad and Mashed Peas European: Mozzarella-Topped Peppers with Tomatoes and Garlic French: Boudin Blanc Terrine with Red Onion Confit Indian: Crisp Braised Pork Shoulder Indian: Jeweled Rice Indian: Pork Chops with Sweet-and-Sour Cider Glaze Italian: Pancetta and Taleggio Lasagna with Treviso CUISineprediction onrecipes
- 21. APPENDIX Infinite Players Pretrained*: 'crude_oil', 0.771960, 'petroleum', 0.76550, 'gas', 0.7105979, 'Oil', 0.69080603, 'natural_gas', 0.6823650, 'crude', 0.6720212, 'hydrocarbon', 0.66379529, 'oilfields', 0.6539833, 'hydrocarbons', 0.629402399, 'oilfield', 0.6252065 Word2vecmodel: Pretrained*vs.custom Ours: 'cooking oil', 0.8903055, 'vegetable oil', 0.85543, 'corn oil', 0.74846285, 'canola oil', 0.7239809, 'peanut oil', 0.7191789, 'salad oil', 0.63877767, 'olive oil', 0.6366288, 'sunflower oil', 0.5754506, 'safflower oil', 0.54821264, 'lite olive oil', 0.5148879 *GoogleNews-vectors-negative300
- 23. References andRelevantWork 1. https://www.kaggle.com/c/whats-cooking/data (P) 2. https://www.kaggle.com/shuyangli94/food-com-recipes-and-user-interactions (D) 3. https://www.kaggle.com/hugodarwood/epirecipes (B) 4. https://www.kaggle.com/kaggle/recipe-ingredients-dataset (S) 5. https://www.kaggle.com/kanaryayi/recipe-ingredients-and-reviews (P) 6. https://data.world/datafiniti/food-ingredient-lists (D) 7. https://link.springer.com/article/10.1007/s10844-017-0469-0 8. http://foodb.ca/ (B) 9. https://github.com/lingcheng99/Flavor-Network (S) 10. https://www.nature.com/articles/srep00196 11. https://www.foodpairing.com/ 12. https://www.wired.com/2013/11/a-new-kind-of-food-science/ 13. https://www.prescouter.com/2019/05/flavor-discovery-big-data-ai/ 14. https://waterfootprint.org/media/downloads/Mekonnen-Hoekstra-2011-WaterFootprintCrops.pdf 15. https://www.footprintnetwork.org/licenses/public-data-package-free/ 1. A New Kind of Food Science: How IBM Is Using Big Data to Invent Creative Recipes ● The study develops an algorithm that generates a list of recipes ranked using three categories: surprise, pleasantness of odor, and flavor pairings 1. Flavor network and the principles of food pairing ● The study introduces a flavor network that captures the flavor compounds shared by culinary ingredients. Given the increasing availability of information on food preparation, their data-driven investigation also opens new avenues towards a systematic understanding of culinary practice. 1. How healthy is the meal: an analysis of recipe data ● The study looks into the interconnection between ratings, nutrients, ingredients, meals, seasons, holidays and cooking techniques. Infinite Players
- 24. CUISINE Infinite Players GROCERYINSPIRATION To these classifications. We looked at our own grocery store experiences and saw that we all could identify items in the supermarket from these cuisines. Therefore, people could recognize most these cuisines On the other hand, some cuisines had very distinctive flavors and classifications. Such as Jamaican & Moroccan. Therefore we tried keeping a small sample & building a model around it.
- 25. Fonts& colors used This presentation has been made using the following fonts: Staatliches (https://fonts.google.com/specimen/Staatliches) Roboto Condensed (https://fonts.google.com/specimen/Roboto+Condensed) #4c1130 #ff5864 #df183d#20124d #76a5af #134f5c#ffd966 OTHER RESOURCEs: Inspiration from across SlidesGo Infinite Players
Editor's Notes
- #3: We were looking for inspiration and we foodies who are students found that there is no good recipe recommendation engine which would suggest recipes with the ingredients available at our hand and at the same time satisfy of our specific taste. Most of the existing solutions suggest very common recipes. Data Try best models We collated the best results through an intutive workflow
- #4: COVID-19 has had an adverse impact on the populations health and finances. We want to provide people with a way to make food at home easily and quickly by recommending recipes depending on their preferences and what ingredients they have available on hand. Our proposed solution bodes well in the current ongoing pandemic wherein access to restaurant food is becoming increasingly difficult. Anyone interested in saving money on eating out while simultaneously becoming more independent and healthier will benefit from this. People using the service can save money while simultaneously honing a skill everyone should have which is the ability to cook food for one's self. Also, they will have an informed choice for eating healthy food.
- #5: We identified the fields in every dataset which would provide some sort of value to our recommendation engine Fussiness Ratings
- #7: INGREDIENT CLEANING: We wanted to transform all our datasets to use the “Ingredient only” format. We used the following techniques across our datasets: Removal of punctuation, numeric quantities and extra spaces Removal of quantity strings such as “ounces”, “pounds” etc. and their variations Splitting ingredients by “and” and “with” into individual ingredients. For example, “tomato sauce with basil and garlic” would become “tomato sauce”, “basil”, “garlic” Use of POS tagging to identify noun phrases. For example, “Whisk some eggs” -> “eggs” Removal of words ending in ‘ed’.
- #12: We applied the cuisine prediction on recipe datasets (>270K) Key differences: NN distributes recipes from Ensembling’s “American” and distributes in mexican, and other cuisines.
- #15: Explain one result