A data minimization strategy limits personal and sensitive data to what an AI system strictly needs for a defined purpose. More inputs can improve model quality or user experience, but unnecessary data increases privacy, compliance, and trust risks. section 4.1.2 section 11.2.2
Apply the strategy across the data lifecycle:
- Define the purpose and exclude sensitive fields that do not contribute enough value; Nina removed personal or physical traits from clustering inputs. section 4.1.2
- Obtain explicit consent for training, insights, or analytics, explain how data will be used, and provide retention opt-outs. section 4.1.2
- Remove or anonymize confidential data that flows through training, evaluation, or production without authorization. section 11.1.1
- Anonymize or redact sensitive data by default, expose only inputs needed for a specific recommendation, and securely delete data after its purpose ends. section 11.2.2
- Revisit the choice as training and evaluation reveal data problems, because early decisions can create downstream failures. section 4.1.2
Clustering customer segments
A team preparing customer data for clustering can use purchase history, activity timing, and visit counts while excluding unnecessary personal traits. If later recommendations need customer-specific information, the team can offer transparent inclusion choices rather than collecting it by default. section 4.1.2 section 11.2.2
Table 4.1 Examples of data points for clustering
| Features |
|||||
| user_id |
purchased_items |
purchase_value |
last_active |
n_visits |
|
| Examples |
abhj3k |
2 |
908 |
2024-04-30 08:36:24 |
48 |
| shj67d |
0 |
0 |
2023-12-26 12:56:24 |
24 |
|
| i963gh |
12 |
673 |
2024-05-15 23:22:11 |
156 |
|
Use the clustering data-point table to distinguish potentially useful behavioral fields from personal or sensitive attributes that may not be necessary. section 4.1.2
