2024

Dallas Police Incident Analysis

Built a reproducible cleaning, feature-reduction, and supervised classification pipeline turning the City of Dallas police-incident dataset of 86 raw columns into modeled predictions of offense status, enabling pattern discovery across location and demographics. Solo data science project.

Headline outcome

Turned an unusable 86-column extract into modeling-ready data and established a reproducible scoring baseline comparing KNN, Decision Tree, and Random Forest on fixed splits.

Stack

Problem → solution → result

1. Problem. High-cardinality categoricals with severe imbalance exploded into thousands of one-hot columns.

   Solution. I engineered mappings collapsing compound and ambiguous category values, dropped categories below a 0.1% frequency threshold, and removed attributes missing more than 50% of values.

   Result. A tractable encoded feature space and reduced label noise, giving more stable precision and recall on tree-based models.

2. Problem. The raw 86-column extract blocked any modeling — sparse columns, text-typed dates, incomplete rows.

   Solution. I automated a cleaning pipeline dropping majority-missing columns, converting date fields to datetime, removing incomplete rows, and emitting two cleaned datasets with and without rare categories.

   Result. Modeling-ready data, and rare-category pruning measurable as a controlled comparison.

3. Problem. Ad hoc splits made scores incomparable between runs.

   Solution. I implemented stratified sampling across both imbalanced and balanced distributions and exported fixed train/test splits to disk.

   Result. Every model evaluated against identical data, so score differences reflected the model rather than the split.

4. Problem. No baseline existed for which classifier suited this categorical-heavy dataset.

   Solution. I evaluated KNN, Decision Tree, and Random Forest, tuning via grid search and selecting K empirically at 18 for KNN.

   Result. Identified the best-performing configuration and improved accuracy over untuned baselines.

5. Problem. Accuracy alone hid minority-class failures.

   Solution. I standardized evaluation on accuracy, full precision/recall/F1 reports, and confusion matrices in tabular and graphical form.

   Result. Exposed per-class performance and established a reproducible scoring baseline.

6. Problem. Inline preprocessing couldn't be reused.

   Solution. I implemented modular functions for sampling, binning, encoding, column filtering, and visualization.

   Result. The pipeline is reusable on other categorical-heavy datasets without rewriting preprocessing.