Finding Singapore’s next business hubs

01 Overview

The brief

For CP3403 Data Mining at JCU Singapore, our group set out to identify which areas of Singapore are most likely to become economic and innovation hubs. The only raw material was business registration data: postal code, street name and block, plus each entity’s type, status and industry code (SSIC).

The analysis combined supervised learning, to classify and rank locations, with unsupervised clustering, to describe what different kinds of areas look like.

02 Problem & users

A useful answer is a ranked, explainable shortlist

We wrote for urban planners, investors and policymakers: people deciding where infrastructure, capital or support programmes should go. They need more than a yes/no prediction per street. They need a ranking they can question.

03 My role

My part of the work

This was a group project (Group 34). I worked on feature engineering and model development across the classification and clustering stages: shaping the region-level features, setting up the models and interpreting their outputs.

04 Approach

Decisions that shaped the analysis

1. 01

   Sample before modelling

   The raw file had more than 280,000 rows. We drew a 10% random sample (28,000 rows, fixed seed) so models could be iterated on quickly. The trade-off is fewer records for sparsely populated areas.

2. 02

   Define a “region” as postal code + street name

   Street names alone merge very different places; the combination gave a stable unit to label, score and cluster.

3. 03

   Try two proxies for “hub”

   The SVM used a density label: regions with more registrations than the median. The Random Forest used a growth-oriented label: regions in the top 10% of new registrations in their most recent year.

4. 04

   Rank, don’t just classify

   A location with three businesses and a 99% hub probability shouldn’t outrank one with 900. Each region’s score is its average predicted probability × log(1 + number of businesses), which rewards confidence backed by volume without letting size dominate.

5. 05

   Use K-Prototypes for clustering

   The data mixes numeric location fields with categorical business attributes. K-Means only handles numbers; K-Prototypes handles both directly.

05 Analysis

From raw records to models

• Cleaning: registration dates converted to datetimes (invalid values coerced to missing), block and postal code converted to numbers, text fields trimmed. Rows with missing values were dropped; the block column alone had 360 gaps.
• SVM baseline: block, postal code and registrations-per-region, standardised, with a sigmoid kernel and 3-fold stratified cross-validation.
• Random Forest: 100 trees in a scikit-learn pipeline. Block and postal code were scaled; entity type, company type, SSIC code and status were one-hot encoded (ignoring unseen categories). Evaluated with 3-fold stratified cross-validation, then on a holdout set.
• Clustering: regions aggregated with their business count and most common entity type, company type, SSIC code and status; K-Prototypes with k = 3; PCA to plot the result in two dimensions.

06 Results

What we found

Random Forest mean accuracy, 3-fold stratified CV

~91%

Random Forest accuracy on the 6,192-record holdout set

89%

SVM mean accuracy (folds: 80.2%, 79.9%, 77.6%)

~79%

SVM ROC AUC

0.866

The ranking at the top of this page is the main output. Its top four, Paya Lebar Road, Anson Road, Robinson Road and Circular Road, were the same under both classifiers, and they are recognisable commercial areas. That agreement is a useful sanity check, not proof.

07 Limitations & learning

What I would do differently

• The labels come from the data the models learn from. The SVM’s label is defined by registrations per region, and registrations per region is also one of its features. The accuracy figures therefore show how well each model reproduces our proxy, not how well it forecasts growth.
• Location features can leak the answer. A region is defined by postal code + street name, and the Random Forest also uses postal code and block as features. Because the split was by record rather than by region, businesses from the same region sit in both training and test data, so the model can partly memorise each region’s label. The ~91% is therefore an upper bound, not an estimate of performance on unseen areas.
• 79% vs 91% is not a like-for-like comparison. The two classifiers were trained on different label definitions, so their accuracies answer different questions.
• k = 3 was chosen from business logic. An elbow or silhouette check would make the cluster count easier to defend, and the clusters largely follow registration volume.

08 Links

Notebook and report

• GitHub repositoryJupyter notebook with preprocessing, classification and clustering, plus the full project report (PDF).github.com/thantthuaung/Data-Mining-Project (opens in a new tab)