P.R.E.G.G.E.Rs

Pregnancy Resources Encouraging Good Growth & Eating Right. A Capstone project for the Galvanize Data Science Immersive.

Business Understanding

There are a ton of resources available for new mothers. Overwhelmed by this abundance, new mothers frequently turn to friends and family for advice rather than sifting through the deluge to find evidence-based nutrition evidence.

This project scraped over a dozen pregnancy and nutrition websites to aggregate resources in one place. Mothers then answer a few short questions and the resource recommender uses Kmeans clustering to sort through the resources for them, providing a tailored list of information that provides a spectrum of resources that supports mothers nutritional information needs. It also provides a psuedosience detector that allows users to copy and paste nutrition-related information into a form and provides the probability that the information is science-based or psuedoscience.

Data Understanding

There are be three inputs to this project:

1. A collection of nutrition-related documents and resources from trusted parenting resource websites.
2. A collection of pregnancy-related documents and resources from trusted health and nutrition websites.
3. Psuedoscience articles scraped from disreputable health and nutrition websites.

A lot of the work on this project was related to web scraping online resources and cleaning them up enough to belong in the repository. I saved the title, link and content of each article. I did not save pictures, videos or infographics. In total there were 2757 articles.

Pregnancy websites:

Baby Center, Kellymom, What to Expect, The Bump, Parenting.com and Babble. Nutrition-specific keywords: nutrition, food, eat, meal, formula, nutrients, vitamins, supplements, diet, health, wellness,

Nutrition websites:

Authority Nutrition, Eat Right, Weight and Wellness, Food Insight, Food and Nutrition, Mayo Clinic Pregnancy-specific keywords: pregnant, baby, breast-feeding, birth, ovulation, postpartum, placenta, trimester,

Psuedoscience websites:

Dr Perl Mutter, FoodBabe, beachbody, Bodybuilding, Weston A Price, Mercola

Data Preparation

Once I had my corpus of scientific and pseudoscientific articles, I combined everything into one overarching corpus and then stemmed the words from each article using snowball stemmer.
I also created 53 text features. NLTK's Part of Speech tagging made up the vast majority of these, but were complemented with additional features I thought might have some correlation with science or psuedoscience (for instance, capital letters and exclamation points in the title).

Modeling

Resource Repository Clustering

Using only the trusted scientific documents, I vectorized the articles using TF-IDF and then clustered the articles into categories using a KMeans clustering algorithm. Below is an summary of the top 10 features from documents clustered together into 20 different clusters.

Pseudoscience Filtering

Using 80% of my psuedoscience and evidence-based science documents as training data, I ran a Naive Bayes classifier to determine the likelihood that new articles are likley to be evidence-based or full of false claims. I combined this with a Gradient Boosted Classifier built on the 53 text features.

Evaluation

This ensemble model was very accurate in discerning science from pseudoscience in the remaining 20% of the testing data. Below is the ROC curve demonstrating that at a probability threshold of about .9 the true positive rate is around .9 and the false positive rate is low, only about .04.

We will ask several mothers that we know to take the quiz and rate the quality of resources the recommender provides.

Deployment

Eventually the tool will live on a nonprofit’s website, and will be launched later this year as a soft launch for the nonprofit itself.