ABOUT

In 2019, there are nearly 4 million women in the U.S. starting their journey of two-men-diet. According to one research, around 10% of the moms experienced nutrition deficiency during their pregnancy even after taking daily prenatal supplements. Nutrition deficiency are risky to both the mom and the baby, and adding more supplements also have side effects. So better planning each meal is the first option to solve the problem.

OUR MISSION
Make just-in-time nutrition choices easy for new moms.

OUR PRODUCT(MVP)
An on-demand and informative nutrition recipe recommendations Web application.

  • Target-Focused Needs: App focuses on needs of pregnant/lactating women
  • Customization: Allows customization based on individual needs and preferences
  • Minimize User Fatigue: Only two input pages and could be saved for next use

OUR METHODOLOGY

  • ALL
  • Data
  • Model
  • Evaluation
  • Infrastructure

We incorporated three types of data:
Recipe meta data, raw ingredient meta data, and special knowledge for women health and nutriology

We tried two options to solve each question and eventually chose the NER model for Recipe Reommendation step and chose W2V embedding for the Ingredient Suggestion step

Recipe Recommendation Option 1: Named Entity Recognition (NER) Model

Recipe Recommendation Option 2: Shingle(n-gram) Similarity Algorithm
We used the algorithm shingle(n-grams) similarity algo which converts strings into sets of n-grams then the distance are calculated between them. In our case, user ingredients list and recipe ingredients list are converted into two pre-calculated profiles and the cosine similarity between them are calculated. After overlaying with user preference and constraints, the result is ranked by nutrition and show the top three.

Ingredient Suggestion Option 1: Word2vec Embeddings (Gensim Model)
By using gensim model from word2vec, we got top 50 most similar ingredients for each input value. We apply hard constraints, like allergen, exclusion ingredient list, nutrition intake reference and user nutrition ranking preference, then got the final results of co-occurring ingredients in each nutrient category. And user will have a list for ingredients recommendation in the lens of nutrition.

Recipe Recommendation Option 1: Named Entity Recognition (NER) Model
The model extracted “turkey breast” rather than just “turkey” and extracted “imitation crabmeat” not just “crabmeat,” as well as extract “vegetable broth” rather than just “vegetable”. However, in some cases the model was not able to extract any core ingredients like anise or anchovy, or extracted the wrong information like “hot sausage” or “vanilla soy. These problems are most likely derived from insufficient tagged data for training.

Ingredient Suggestion Option 1: Word2vec Embeddings (Gensim Model)
Data visualization of ingredients clusters are generated to evaluate the similar ingredients from the clusters. For example: Mozzarella, Provolone and Pepperoni are in a cluster. From this data viz, we can confirm the clusters are valid.

Simple, Scalable Infrastructure using AWS Service.
The App is developped using Flutter.

MEET OUR TEAM

We are students of Master of Information and Data Science (MIDS) at UC Berkeley.
Please check our Project Webpage on MIDS website.

Suzy Choi

Revenue Management

Suzy works in Boeing Commercial Airplanes Revenue Management as a Pricing Analyst and Data Analytics focal. She supports airline customer negotiation activities and helps develop pricing strategies that enable the best financial outcome for The Boeing Company.

Wenqi Liu

Data Scientist

Wenqi works in the Enterprise Marketing department at Wells Fargo as a Data Scientist. She builds statistical/machine learning models to understand and predict customer behavior patterns, like new customer acquisition, customer churn and customer segmentation models.

Jennifer Middleton

VP

Jennifer is VP, InfoSec Data Engineering & Analytics at a Fortune 200 company. She oversees petabytes of data and the associated analytics required to support the firm's Information Security activities and is leading the firm's Data Protection Strategy.

Yaran Zhang

Data Operations

Yaran is a Data Operations Director at Moody's Analytics. She manages Data Quality, Data Engineering and Services functions and resources, to build and support the data pipelines and data contents for quantitative Credit Risk models delivered via various product suites.

CONTACT US

Copyright © 2020.MIDS.HealthMamaEats.