311 Service Request Analysis

Reading lens

Info

A civic-data study over 1.6M+ Winnipeg 311 records spanning 16 years — association mining with FP-Growth, seasonal decomposition with STL, and MAD-based anomaly detection.

Role

University Project

Data mining

University of Manitoba

Stack

Python

FP-Growth

STL

MAD

Proof

FP-Growth association rules over request co-occurrence

STL decomposition separating seasonal cycles from trend

MAD-based anomaly detection at 1.6M-row scale

GitHub ↗

Sixteen years of 311 records hold operational patterns no one reads row by row.

01 — The Analysis04 Films

Film — overview.mp4

01 — Raw Data to First Chart

Sixteen years of Winnipeg 311 requests — over 1.6 million records — load, clean and resolve into a first picture of what a city complains about.

Raw 311 data to the first chart

Film — question1.mp4

02 — Where the Requests Come From

Which wards file the most requests, plotted — the geographic shape of demand a staffing plan could act on.

Which wards file the most requests

Film — question2.mp4

03 — How Long the City Takes

Response times, plotted across request types — the gap between filing and fixing, made visible.

Response-time distribution, plotted

Film — method.mp4

04 — The Cleaning, Run Live

The cleaning and method cells run on camera — FP-Growth for co-occurrence, STL for seasonality, MAD for anomalies. The pipeline is the evidence.

The cleaning step, run live

02 — Business ValueWhat Changes

Patterns

FP-Growth surfaces which request types travel together in the same neighbourhoods — co-occurrence argued from 1.6M rows.

Seasons

STL decomposition turns the Winnipeg-winter demand cycle into a measured curve a maintenance calendar could follow.

Anomalies

MAD thresholds flag the spikes that deserve a second look — robust statistics that extreme days can't drag around.

A data-mining study over public records — the signals are demonstrated, not deployed, and this page doesn't claim otherwise.

Engineering Deep DiveArchitecture · Decisions · Lessons

What does a city complain about, and when?

Sixteen years of Winnipeg 311 service requests — over 1.6 million records — mined for which problems travel together, which follow the seasons, and which spikes deserve a second look.

01 — MethodEngineering Lens

Association over Anecdote

FP-Growth association mining over request co-occurrence answers which request types cluster in the same neighbourhoods — patterns argued from 1.6M rows, not from memorable examples.

Seasonality Made Explicit

STL decomposition separates the seasonal structure from trend and noise, making the Winnipeg-winter demand cycle a measured curve instead of a folk assumption.

Anomalies by Median, Not Mean

Median-absolute-deviation thresholds flag the spikes — a robust statistic that a handful of extreme days can't drag around, which matters at 1.6M-row scale.

02 — Engineering & ArchitectureSystem Map

The pipeline is deliberately boring: load and clean 1.6M records in Python, then ask three questions with three tools — FP-Growth for co-occurrence, STL for seasonal structure, MAD for anomaly flags.

1.6M Records

FP-Growthco-occurrence

STLseasonality

MADanomalies

Method fan-out — three questions asked of the same dataset

Data

1.6M+ records16 years of Winnipeg 311

Methods

FP-GrowthSTL decompositionMAD thresholds

Output

Co-occurrence patternsSeasonal cyclesAnomaly flags

System map — three methods, one civic dataset

03 — What I Learned02 Lessons

Scale Changes the Tools

At 1.6M rows, fragile statistics break quietly — MAD's median-based thresholds hold where means get dragged by extreme days.

Ask the Data Three Ways

Association, seasonality and anomaly detection answer different questions about the same records — any one method alone would have left the other two answers invisible.

04 — Roadmap & CollaborationFuture Work · Contact

Open to Collaboration

Public operational data is full of staffing and maintenance signals — if you steward a dataset like this, I'd love to compare notes.

Email ↗LinkedIn ↗

Reading lens