City of Lethbridge data projectCan one simple rule spot Lethbridge’s arterial roads, and what does it miss?
The City of Lethbridge publishes its street network with each segment’s type, posted speed, lanes, signals and sanding priority. Our students hide the type, let a program learn which single yes-or-no question does most to pick out arterials, and then measure it three ways. Every figure here is our own result, not the City’s.
Our results on the City of Lethbridge Street Network (4,830 Local, Collector and Arterial segments fetched 22 October 2026; 3,622 to learn from, 1,208 held back, 247 of them arterial)| Rule | Accuracy | Precision (when it says arterial, how often it is right) | Recall (share of real arterials it finds) |
|---|
| Never say arterial | 79.6% | none (it never says arterial) | 0% |
| Learned rule: speed limit over 50 | 88.7% | 99.1% | 44.9% |
| Hand rule: sanding priority 1 or 2 | 86.8% | 63.7% | 83.0% |
Below is everything, written with Python’s standard library only. The program pages through the City’s street service 1,000 segments at a time, keeps only local, collector and arterial streets, shuffles them with a fixed seed, and tries seven candidate questions. For each it measures how mixed the two resulting groups are (the Gini impurity used in decision trees) and keeps the question that separates arterials most cleanly. Run on 22 October 2026, it printed the three rows above:import json, random, urllib.request, urllib.parse
URL = ("https://gis.lethbridge.ca/gisopendata/rest/services/OpenData/"
"odl_streets/MapServer/0/query?")
ask = {"where": "FDESC IN ('Local','Collector','Arterial') AND SPEED IS NOT NULL",
"outFields": "FDESC,SPEED,NumOfLanes,RedLight,SandPriority",
"returnGeometry": "false", "orderByFields": "OBJECTID", "f": "json",
"resultOffset": 0, "resultRecordCount": 1000}
roads = []
while True:
batch = json.load(urllib.request.urlopen(URL + urllib.parse.urlencode(ask)))["features"]
if not batch:
break
for f in batch:
a = f["attributes"]
roads.append({"speed": a["SPEED"], "lanes": int(a["NumOfLanes"] or 0),
"signal": a["RedLight"] is not None, "sand": a["SandPriority"],
"arterial": a["FDESC"] == "Arterial"})
ask["resultOffset"] += len(batch)
random.seed(2026)
random.shuffle(roads)
cut = len(roads) * 3 // 4
train, test = roads[:cut], roads[cut:]
QUESTIONS = {"speed limit over %d" % t: (lambda r, t=t: r["speed"] > t) for t in (30, 40, 50, 60)}
QUESTIONS.update({"more than %d lane" % t: (lambda r, t=t: r["lanes"] > t) for t in (1, 2)})
QUESTIONS["has a traffic signal"] = lambda r: r["signal"]
def impurity(group):
p = sum(r["arterial"] for r in group) / max(1, len(group))
return len(group) * 2 * p * (1 - p)
def split_cost(rule):
yes = [r for r in train if rule(r)]
return impurity(yes) + impurity([r for r in train if not rule(r)])
def report(name, rule):
hit = [r for r in test if rule(r)]
tp = sum(r["arterial"] for r in hit)
every = sum(r["arterial"] for r in test)
right = sum(rule(r) == r["arterial"] for r in test)
print("%s: accuracy %.1f%%, precision %.1f%%, recall %.1f%%"
% (name, 100 * right / len(test), 100 * tp / max(1, len(hit)), 100 * tp / every))
winner = min(QUESTIONS, key=lambda q: split_cost(QUESTIONS[q]))
print(len(roads), "street segments;", len(train), "to learn from,", len(test), "to test")
report("never say arterial", lambda r: False)
report("learned rule '" + winner + "'", QUESTIONS[winner])
report("sanding priority 1 or 2", lambda r: r["sand"] in ("1", "2"))
1. Accuracy can lie
Only about one segment in five is an arterial, so a rule that never says arterial is still right 79.6 per cent of the time while finding none of them. That is why we never judge a classifier by accuracy alone on an uneven dataset.
2. Precision against recall
The learned rule, a posted speed above 50, is almost never wrong when it fires, at 99.1 per cent precision, but it finds fewer than half the arterials, because 532 arterial segments are posted at 50. Sanding priority finds 83 per cent of them, at the cost of many false alarms.
3. Which mistake is worse?
There is no single right rule. If a wrong label is expensive, use the precise one; if missing an arterial is worse, use the one with more recall. Students write down a situation for each and defend their choice.
A deeper tree did not help
We also let the program ask a second question inside each branch, making a two-level decision tree. On this split it made exactly the same test predictions as the single rule. More model is not always more learning; sometimes the data simply does not hold the extra information. And notice that sanding priority is itself a decision someone made about each road, so a rule built on it partly copies an earlier human label rather than discovering anything new.
Some details for careful readers. The layer holds 5,833 segments in all; we kept the 4,830 typed Local, Collector or Arterial with a speed recorded (2,133 local, 1,645 collector, 1,052 arterial) and left out ramps, private, provincial, county, park and gravel segments. Only 8 of the kept non-arterial segments are posted above 50, which is why the speed rule is so precise. Posted speed is a reasonable clue, not a definition: the City decides street types for its own planning reasons. The data is published under the City of Lethbridge Open Data License, Version 2.0, which grants “a worldwide, royalty-free, non-exclusive license to use the information” and lets users use it “in any medium, mode or format for any lawful purpose”. Source: City of Lethbridge, Street Network, Lethbridge Open Data Catalogue. The City does not endorse this page or our classes.
