400 Labels to 94% Accuracy: Validating Grocery Store Data

Google Places API returns thousands of results for "grocery store" across California counties. The results include supermarkets like Safeway and Trader Joe's, but also 7-Eleven locations, gas station minimarts, liquor stores, and restaurants with incidental grocery items.

For food access research, this distinction matters. A convenience store does not provide equivalent food security value to a full-service supermarket. Manual validation of thousands of locations is not feasible. Automated classification becomes necessary.

The Training Data

400 labeled locations (200 confirmed grocery stores, 200 confirmed non-grocery: convenience stores, liquor stores, miscategorized restaurants) provide enough training data for a binary classifier. The labeling takes about two hours: checking each location against its business website and Google Street View imagery.

Google Places provides several features for each location:

Our hypothesis: real grocery stores tend to have certain type tags, higher review counts, and names containing words like "market" or "foods" rather than "liquor" or "gas."

Model Iteration

With a labeled dataset, Claude Code can iterate through classifier specifications rapidly. Each cycle (write code, run it, evaluate accuracy, propose modifications) takes minutes rather than hours.

Iteration 1: Logistic regression, raw features
→ 78% accuracy

Iteration 2: Added text features from business names
  (name_has_market, name_has_liquor, name_has_gas)
→ 84% accuracy

Iteration 3: Parsed Google type tags into binary indicators
→ 88% accuracy

Iteration 4: Added review count interaction terms
→ 91% accuracy

Iteration 5: XGBoost with tuned hyperparameters
→ 94% balanced accuracy

The jump from 78% to 84% came from a simple observation: business names contain signal. "Safeway" and "Trader Joe's" differ systematically from "7-Eleven" and "Chevron Food Mart." Extracting indicator variables for common substrings captures this.

The final jump to 94% came from switching to XGBoost, which handles feature interactions better than logistic regression. Review count matters more for locations tagged as "convenience_store" than for those tagged as "supermarket." A large convenience store with hundreds of reviews might actually be a grocery store miscategorized by Google.

What the Model Learned

Feature importance analysis reveals which signals drive predictions:

The type tags dominate, which makes sense: Google's categorization captures real information. But name-based features add predictive power, especially for edge cases where Google's tags are ambiguous or missing.

Validation

When we spot-check 50 predictions from each category against business information, we find:

• Predicted grocery, actually grocery: 47/50 (94%)
• Predicted non-grocery, actually non-grocery: 48/50 (96%)

The three false negatives were ethnic markets with non-English names that lacked the typical "market" or "grocery" substrings. The two false positives were large convenience stores (a Walgreens and a CVS) that arguably do provide meaningful food access; the classification boundary is genuinely ambiguous here.

Results

Applied to the full 6,613 locations: 4,847 classify as grocery stores (73%), 1,766 as non-grocery (27%).

Without classification, grocery access would be overstated by approximately 27%. The spot-check suggests roughly 6% misclassification remains, but this is acceptable for research purposes: the alternative is either manual review of 6,613 locations or no validation at all.

Time Investment

Without agent-assisted iteration, this work would require 2-3 days: learning the scikit-learn API, debugging feature engineering code, manual hyperparameter tuning. The agent compresses the implementation bottleneck; the human time investment is almost entirely labeling and validation.

All code and data acquisition instructions are available in the GitHub repository.