Scoring POI Sensitivity for Location Datasets

To let a pipeline decide which stops to hide first, assign each point-of-interest a sensitivity score that combines how revealing its category is, how long the person stayed, and how rare visits there are — then suppress in descending score order until the release meets its risk budget.

Core Calculation # permalink

Not every stop carries equal disclosure risk. A visit to a grocery store reveals almost nothing; a visit to a dialysis clinic, a domestic-violence shelter, or a place of worship can expose health status, victimization, or belief. A sensitivity score turns that intuition into a ranking a machine can act on, feeding directly into a privacy risk scoring framework for GIS.

Define the raw sensitivity of a stop i as a weighted product of three factors:

where:

• $w_{c(i)}$ is the category weight of the POI category c(i) the stop maps to;
• $\tau_i$ is the dwell time at the stop and $\tau_0$ a reference dwell (a short visit); the logarithm keeps very long dwells from dominating;
• $r_i$ is the visit rarity, defined from the visit frequency $f_i$ (share of individuals who ever visit that POI) as $r_i = -\log(f_i)$;
• $\alpha, \gamma \ge 0$ tune how much dwell and rarity matter relative to category.

Normalize to a comparable zero-to-one scale with a min-max transform across the dataset:

The suppression rule is then simply: suppress every stop with $\tilde{S}_i \ge \theta$, where the threshold θ is set by how much of the dataset your risk budget allows you to remove.

Category weights (illustrative tiers) # permalink

Worked example # permalink

Take four stops, with $\tau_0 = 5$ min, $\alpha = 0.5$, $\gamma = 0.3$:

After min-max normalization the shelter stop D lands at $\tilde{S} = 1.0$, the clinic and worship stops near 0.37–0.38, and the supermarket at 0.0. With a budget to suppress the top 25% by score, only D is removed; tighten θ and B and C follow, while the innocuous supermarket visit A is never suppressed. The score has correctly spent scarce suppression on the most damaging disclosures.

Python Implementation # permalink

The routine joins detected stops to a POI layer, computes the score, and returns a ranked suppression list. The stops are assumed to come from a stay-point detector; suppressing them is the enforcement step described in stop-location and POI suppression.

import numpy as np
import geopandas as gpd
import pandas as pd

# CRS: reproject both layers to a projected metric CRS (UTM zone 33N,
# EPSG:32633) so the nearest-POI join uses true metre distances.
CRS_METRIC = "EPSG:32633"

CATEGORY_WEIGHTS = {
    "supermarket": 1.0, "retail": 1.0, "park": 1.0,
    "bar": 2.5, "gym": 2.5, "political_office": 2.5,
    "clinic": 6.0, "pharmacy": 6.0, "worship": 6.0,
    "shelter": 9.0, "addiction_service": 9.0, "home": 9.0,
}
UNKNOWN_WEIGHT = 6.0  # conservative: absence of evidence != evidence of safety


def score_poi_sensitivity(
    stops: gpd.GeoDataFrame,
    pois: gpd.GeoDataFrame,
    max_join_m: float = 40.0,
    tau0_min: float = 5.0,
    alpha: float = 0.5,
    gamma: float = 0.3,
    threshold: float = 0.75,
) -> gpd.GeoDataFrame:
    """
    Score each stop's disclosure sensitivity and flag suppression candidates.

    Args:
        stops: GeoDataFrame with 'geometry' (points) and 'dwell_min'.
        pois:  GeoDataFrame with 'geometry' and 'category'.
        max_join_m: max metres to accept a nearest-POI match; beyond this the
            stop is treated as unknown-category (scored conservatively).
        tau0_min: reference dwell for the dwell term.
        alpha, gamma: dwell and rarity weightings.
        threshold: normalized-score cutoff at/above which to suppress.

    Returns:
        stops with 'category', 'sensitivity', 'suppress' columns, sorted by
        descending sensitivity (the ranked suppression list).
    """
    assert stops.crs == pois.crs == CRS_METRIC, "Reproject both layers to EPSG:32633"

    # 1. Nearest-POI spatial join to attach a category to each stop.
    joined = gpd.sjoin_nearest(
        stops, pois[["category", "geometry"]],
        how="left", max_distance=max_join_m, distance_col="poi_dist_m",
    )
    # Deduplicate ties from sjoin_nearest (keep the closest single match).
    joined = joined.sort_values("poi_dist_m").loc[~joined.index.duplicated(keep="first")]
    joined["category"] = joined["category"].fillna("unknown")

    # 2. Category weight (unknown -> conservative high default).
    w_c = joined["category"].map(CATEGORY_WEIGHTS).fillna(UNKNOWN_WEIGHT)

    # 3. Visit rarity: rarer POIs -> smaller equivalence class -> more identifying.
    freq = joined["category"].map(joined["category"].value_counts(normalize=True))
    rarity = -np.log(freq.clip(lower=1e-6))

    # 4. Dwell term (log-damped so long stays don't swamp the score).
    dwell_term = 1.0 + alpha * np.log1p(joined["dwell_min"] / tau0_min)

    # 5. Combine, then min-max normalize to [0, 1].
    raw = w_c * dwell_term * (1.0 + gamma * rarity)
    span = raw.max() - raw.min()
    joined["sensitivity"] = (raw - raw.min()) / span if span > 0 else 0.0

    # 6. Rank + flag. Suppression here means "hide before release".
    joined["suppress"] = joined["sensitivity"] >= threshold
    return joined.sort_values("sensitivity", ascending=False)

Verification # permalink

Confirm the ranking behaves and the suppression budget is respected before wiring the output into a release pipeline.

ranked = score_poi_sensitivity(stops, pois, threshold=0.75)

# Verification checklist:
# [ ] every sensitivity is a normalized value in [0, 1]
assert ranked["sensitivity"].between(0, 1).all()
# [ ] the most sensitive category present outranks the least sensitive
top_cat = ranked.iloc[0]["category"]
assert CATEGORY_WEIGHTS.get(top_cat, UNKNOWN_WEIGHT) >= 6.0, "High-tier stop should rank first"
# [ ] suppression set size stays within the risk budget (e.g. <= 30% of stops)
assert ranked["suppress"].mean() <= 0.30, "Suppression budget exceeded; raise threshold"
# [ ] no unknown-category stop was silently scored as safe
assert (ranked.loc[ranked["category"] == "unknown", "sensitivity"] > 0).all()
print(ranked[["category", "dwell_min", "sensitivity", "suppress"]].head())

The four checks catch the failure modes that matter: an un-normalized score, an inverted ranking, a runaway suppression set that destroys utility, and — most importantly — an unknown-category stop leaking through as if it were safe.

Edge Cases & Adjustments # permalink

• Mixed-use buildings. A single address may host a pharmacy above a café. A nearest-POI join picks one category and understates risk. Where the join returns multiple POIs within max_join_m, take the maximum category weight, not the closest, so the sensitive tenant sets the floor.
• Unknown category. Default unknown stops toward the sensitive tier and flag them for review; treating them as ordinary is the most common silent leak. Better still, infer a provisional weight from the categories of neighboring POIs before defaulting.
• Inference from neighbors. Even a suppressed stop can be reconstructed if the surrounding stops pin it geometrically. When a high-score stop is removed, check whether its position is implied by adjacent retained points, and generalize those neighbors if so.
• Population-relative rarity. Visit frequency must be computed over the release population, not a sample; a place that looks rare in a 1% extract may be common overall, and vice versa. Recompute rarity against the exact cohort you publish.

FAQ # permalink

Why weight by category rather than treating all stops equally?

Because disclosure harm depends on what the place reveals. A supermarket visit exposes little; a visit to an oncology clinic, a shelter, or a place of worship can disclose health, victimization, or belief. Category weighting directs a limited suppression budget to the disclosures that hurt most.

Why include visit rarity in the score?

Rarity is an identifiability signal. A POI visited by few people yields a small equivalence class, so one visit there is far more re-identifying than a visit somewhere everyone goes. Weighting by inverse visit frequency raises the score of exactly the stops that narrow an attacker’s candidate set, complementing a formal privacy risk matrix for municipal GIS.

How should unknown-category stops be scored?

High, not low. An unknown category is an absence of evidence of safety, not evidence of absence of harm. Assign a weight near the sensitive tier or infer one from nearby POIs, and flag the stop for human review before release.

Does a high score mean the stop must be deleted?

No. Suppression is one response; spatial generalization, timestamp cloaking, and region merging are others. The score sets priority order — what you do at each level depends on the utility you must keep and the applicable k-anonymity or regulatory threshold.

Related

• Privacy Risk Scoring Frameworks for GIS — the parent framework this score plugs into
• Building a Privacy Risk Matrix for Municipal GIS — placing POI scores on a likelihood-by-impact grid
• Stop-Location & POI Suppression — enforcing the ranked suppression list on trajectories
• Spatial Privacy Fundamentals & Threat Modeling — where POI sensitivity fits the overall threat model

← Back to Privacy Risk Scoring Frameworks for GIS