Personalization feature vectors

Detects personalization feature vectors and user embeddings: a feature vector / user embedding / feature store export phrase combined with populated model data - numeric vectors, JSON records keyed by a user or customer ID, or labelled user identifiers. Phrase-only mentions (engineering guidance, course notes) stay discovery-only.

Type: regex
Engine: boost_regex
Confidence: medium
Confidence justification: Enforce tiers require populated vectors or identifier-keyed records beside the feature-vector phrase; 'user profile', 'machine learning' and 'behavioral data' were removed from the primary as too generic. Engineering docs and ML course notes stay at 65, as do header-only extracts. Medium because model-training fixtures can carry realistic vectors.
Detection quality: Topic false positive
Jurisdictions: global
Regulations: Privacy Act 1988 (Cth), Information Privacy Act 2009 (Qld), NDB Scheme (Cth), GDPR
Frameworks: ISO 27001, ISO 27701
Data categories: pii
Scope: wide
Risk rating: 8
Platform compatibility: Purview: Compatible, GCP DLP: Compatible, Macie: Compatible, Zscaler: Compatible, Palo Alto: Degraded, Netskope: Unsupported

Pattern

(?i)\b(?:personali[sz]ation\s+(?:feature\s+)?vectors?|feature\s+vectors?|user\s+embeddings?|customer\s+embeddings?|embedding\s+vectors?|feature\s+store\s+(?:exports?|extracts?|snapshots?)|recommendation\s+features)\b

Corroborative evidence keywords

category_affinity, price_sensitivity, affinity, propensity, inferred interest, model version, nearest neighbours, cosine similarity, sensitive inference, health interest

Proximity: 300 characters

Should match

Personalization feature vectors - feature store snapshot 2026-09-12 {"user_id": "u-2204471", "model": "recs-v7", "embedding": [0.1142, -0.5521, 0.0873, 0.3310, -0.0457, 0.2291], "category_affinity": {"baby": 0.91, "fertility": 0.77}, "price_sensitivity": 0.34} — Feature-store record - user_id, embedding vector and affinity evidence (85)
User embeddings export User ID: u-2204471 Sent to the recommendations team. — Single labelled user identifier (75)
Customer embeddings - nightly extract, model version recs-v8 customer_id,emb_1,emb_2,emb_3,emb_4,affinity_outdoor C-8812044,0.2210,-0.4471,0.0932,0.6105,0.88 C-8812051,-0.1317,0.2045,0.5512,-0.0716,0.12 Customer ID: C-8812044 nearest neighbours: [0.2210, -0.4471, 0.0932, 0.6105] — Embedding CSV with customer_id/emb columns, labelled customer ID and numeric vector plus model-version evidence (85)

Should not match

Engineering guidance note: feature vectors are numeric representations of behaviour used by the recommender. Never store raw identifiers in the feature store; use the pseudonymous key. — Engineering guidance
Feature vector export user_id,f_1,f_2,f_3,f_4 — Blank extract - column header only (65)
Course notes: personalization feature vectors let a model compare users by cosine similarity. In this lesson we compute embeddings for a toy dataset. — Course notes
unrelated generic text without domain phrases — Unrelated
customer invoice — Generic word pair

Known false positives

Privacy notices, help-centre articles, product pages and policies describing feature vectors and embeddings. Mitigation: They name the topic without populated rows, labelled values or identifiers, and guidance/help framing is excluded, so they stay at 65.
Analytics training material and documentation with worked examples of feature vectors and embeddings. Mitigation: Training and template framing is excluded; unframed worked examples with realistic rows can still reach 75.
News and industry commentary with aggregate statistics. Mitigation: Aggregate percentages are not keyed to a record identifier, so they do not satisfy the row or labelled-value structures.