Research question

How well can a “good for kids” rating of restaurant and food businesses be predicted from business features such as Yelp category, service attributes, city and key words in the business's name?

Data Source

Yelp Dataset Challenge 6

Application

• Parent-centric recommendation engines
• Restauration market research
• Review fraud detection

Preprocessing

• Extraction of all businesses identificaton, evaluation, categories, attributes, location
• Outcome variable: attribute “good for kids” (binary)
• Filtering (no NA in outcome, category “Restaurant”, “Food”)
• k-means clustering of GPS coordinates
• Inclusion of top 300 words in business names
• Creation of dummy variables

Model building

• Impute NAs with variable median
• Train three individual classifiers and one ensemble
• Validate and test
• select against multiple metrics

Outcome variable is skewed

More than 80% of all “good for kids” entries are positive:

	Not good for kids	Good for kids	Total
n	3667	16424	20091
% Total	18.3%	81.7%	100%

Key Metrics

	RF	LB	NN	GAM
AUROC	0.91	0.87	0.83	0.56
Balanced Accuracy	0.75	0.71	0.72	0.53
Sensitivity	0.97	0.97	0.92	0.90
Specificity	0.54	0.46	0.53	0.15
Kappa	0.57	0.50	0.46	0.06

RF=Random Forest, LB=LogitBoost, NN=Neural Net, GAM=General Additive Model

Random Forest emerges as clear winner. GAM shows surprisingly low performance, despite being the ensemble model.

ROC Curves (GAM curve shows performance against testing data, others against validation data)