A product team has identified useful customer segments, but new users keep arriving. A classifier can assign each new user to one of those predefined segments from their behavior, making tailored experiences possible without manually sorting profiles. Unlike clustering, which discovers groups in unlabeled data, classification learns to predict known labels from labeled examples. section 4.2 section 4.2.4
Table 4.2 A sample of labeled examples
| Features |
Label |
|||||
| user_id |
purchased_items |
purchase_value |
last_active |
n_visits |
search_queries |
segment |
| abhj3k |
2 |
908 |
2024-04-30 08:36:24 |
48 |
3 |
Seekers |
| shj67d |
0 |
0 |
2023-12-26 12:56:24 |
24 |
45 |
Conservatives |
| i963gh |
12 |
673 |
2024-05-15 23:22:11 |
156 |
25 |
Indecisives |
| ty54df |
20 |
1250 |
2024-05-10 14:21:07 |
190 |
5 |
Champions |
The sample pairs user activity and purchase features with a segment label. Those pairs provide the learning signal: which label should the classifier predict for a similar user? section 4.2.1
A practical path from segments to decisions:
- Prepare input–label pairs. Existing cluster labels can be a starting point; otherwise, collect or annotate examples. Check proxy data for distribution differences and bias, and plan human oversight for difficult labeling tasks. section 4.2.1
- Start with a baseline, then test whether added complexity earns its place. Logistic regression can provide segment probabilities and show influential features; a neural network may improve accuracy while making decisions harder to interpret. section 4.2.2
- Evaluate on users excluded from training. Precision matters when assigning someone to the wrong segment would undermine personalization; recall matters when missing an at-risk user would forfeit a useful intervention. Tune against both requirements, not accuracy alone. section 4.2.3
From prediction to action
A predicted segment can trigger a style quiz and recommendations, a purchase nudge, or a loyalty offer. The team should keep monitoring performance as behavior changes: automated assignments are useful only while they remain reliable enough for the decisions built on them. section 4.2.4
