Intelligent Business Zone Planning for Super Bus Service Using DBSCAN Clustering and Convex Hull

Background – Super Bus is an innovative transportation service that offers on‑demand routes such as airport pick‑up, drop‑off, and station transfers, providing a cost‑effective alternative to traditional buses while aiming to satisfy the “last‑mile” need. Defining service zones (business circles) that balance coverage and cost is a core challenge.

Evolution of Intelligent Zones – Initially, zones were manually assigned by suppliers, leading to subjectivity, redundancy, and low efficiency. To address these issues, a data center was built to collect historical user inquiries, hotel bookings, and ride‑hailing orders, amassing 30 million inquiry points, 6 million hotel orders, and 350 k ride‑hailing orders.

Algorithm Introduction

Business Circle Clustering

DBSCAN is chosen for its fast clustering speed, ability to handle noise, no need to pre‑specify the number of clusters, shape‑agnostic clusters, and tunable noise‑filtering parameters.

Evaluating Zone Compliance

Coverage is defined as the sum of clustered data points divided by total points. Only clusters covering at least 10 % of total points are considered. Desired coverage thresholds are 50 % for A‑type zones, 70 % for B‑type, and 80 % for C‑type. The eps parameter of DBSCAN starts at 0.1 km and is increased iteratively to meet these targets.

Process diagram:

Drawing the Zone Boundary

The convex‑hull (Graham scan) algorithm is used:

Identify the point with the smallest y‑coordinate (P0) as the hull start.

Translate coordinates so P0 becomes the origin.

Compute polar angles α of all points relative to P0, sort by α (break ties by distance).

Initialize a stack with P0 and the first sorted point P1.

Iterate through sorted points, checking the turn direction with the line formed by the top of the stack and the current point; pop the stack on right‑turns, push on left‑turns.

Continue until the last point is processed; the stack then contains the convex‑hull vertices.

Framework Design

Shard data by city and service type to reduce task size.

Use the QSchedule platform for load‑balanced task scheduling and periodic execution.

Cache edge points of zones in Redis for fast lookup.

Provide a web tool for visualizing zones and inquiry points.

Notify downstream consumers via qmq after computation completes.

Effect Demonstration

To date, 34 cities and nearly 200 business scenarios have been covered, producing A/B/C‑type zones that are 100 % compliant, eliminating manual bias, and dramatically improving efficiency and dynamic adaptability.

Conclusion – By applying unsupervised DBSCAN clustering and convex‑hull geometry, the Super Bus team can automatically generate compliant service zones based on user behavior data, with ongoing efforts to improve clustering accuracy and computational speed, inviting further collaboration.