Epicurious Recipes

Machine Learning Study on a Dataset of Epicurious Recipes

Data provided by Hugo Darwood via Kaggle: https://www.kaggle.com/hugodarwood/epirecipes

Goals:

1. Recipe recommendation engine: classification model

• Will take some input (base ingredients, cuisine style, nutritional content) and return a classifier
• Collaborative filtering ( If a person A likes item 1, 2, 3 and B like 2,3,4 then they have similar interests and A should like item 4 and B should like item 1)

2. Predict user rating based on features of each recipe: Supervised Learning

• Higher or lower rating based on similar recipes (classifier)
• Estimated score (regressor)

Challenges

Features are a mess

Rating distribution is NOT normalized

Feature Engineering: 680 columns

• Ingredients are already label encoded
• Nutrition content useful, but not relevant in recommendations based on user taste preferences

Initial EDA yields some pretty strange outliers:

Linearly Separable Data?

• Binary classification and linear models will have a hard time if data is not linearly seperable
• Use tree based models if so

Danger of over fitting: if model is trained on specific inputs, it will not be applicable to others

• Clustering
• Build a 'user' similarity matrix

Process

Split and Clean Data

• Worth developing LSA to separate similar features (ingredients, nutrition, etc) Recommendation Engine
• Vectorize salient features
• build LSA models to train recommendations

Similarity Metrics

1. Cosine Similarity
2. Pearson Similarity
3. Jaccard Similarity

Rating Prediction

• Develop webscraper to collect new articles for rating prediction