Machine learningMSc Data Science · University of Salford

Lifestyle and Obesity Segmentation

PythonUnsupervised learningK-MeansHierarchical clusteringPCAscikit-learn

Can lifestyle data alone reveal natural groups related to weight? Using 16 survey features, I built a preprocessing pipeline, chose the number of clusters with the elbow method, compared K-Means with Ward hierarchical clustering and judged both with three internal metrics and a PCA projection. The scores are candid: the segments overlap, which is itself a finding about how continuous lifestyle behaviour is.

2,111Records

31Features after encoding

k = 5Clusters (elbow method)

3 of 3Metrics won by K-Means

The problem

Weight is shaped by many habits at once: diet, water intake, activity, screen time, transport. Rather than predicting a label, this project asks whether people fall into distinct lifestyle profiles, and how clear those profiles really are.

The data

2,111 records and 17 columns, 16 of them features (gender, age, height, weight, family history, eating habits, smoking, water intake, activity, screen time, alcohol, transport). No missing values; 24 duplicate rows noted.
Height and weight are the most correlated numeric pair (r = 0.46). The behavioural features are mostly weakly correlated with each other (|r| of 0.3 or less), so they add complementary information.

Approach

Audit and exploreDescriptive statistics, histograms, boxplots, categorical counts, a correlation heatmap and a pairplot.
Keep it unsupervisedThe obesity-level label is removed from the clustering input.
PreprocessStandardise numeric features and one-hot encode categorical ones, giving 31 dimensions.
Choose kElbow curve of within-cluster sum of squares for k = 1 to 10, pointing to five clusters.
Cluster two waysK-Means (k-means++ initialisation) and agglomerative clustering with Ward linkage, guided by a dendrogram on a 400-record sample.
ValidateSilhouette, Davies–Bouldin and Calinski–Harabasz scores, plus a 2-D PCA projection of each result.

Results

Internal validity metrics (k = 5)
Method	Silhouette ↑	Davies–Bouldin ↓	Calinski–Harabasz ↑
K-Means	0.141	1.987	260.2
Hierarchical (Ward)	0.120	2.066	220.4

Elbow curve of within-cluster sum of squares — Elbow curve used to choose five clusters.

K-Means clusters in PCA space — K-Means clusters in a 2-D PCA projection.

Hierarchical clusters in PCA space — Hierarchical clusters in the same projection.

Key findings

K-Means wins on all three internal metrics (silhouette 0.141 against 0.120, Davies–Bouldin 1.99 against 2.07, Calinski–Harabasz 260 against 220).
Silhouette scores near 0.14 indicate overlapping clusters. The PCA views show a continuous cloud rather than separate islands, which suggests that lifestyle varies along a continuum and that the five segments are best read as soft profiles, not hard categories.
The elbow curve bends gradually rather than breaking sharply, another sign that there are no strongly separated groups. Five clusters is a reasonable choice, not a definitive one.

Next steps

Profile each cluster against the withheld obesity label to see whether segments line up with clinical categories.
Try Gaussian mixture models, which give soft assignments for overlapping groups.
Use Gower distance or k-prototypes to handle mixed numeric and categorical data natively.