Gilian Ponte - Private targeting strategies | PyData Amsterdam 2024

Differential privacy provides mathematical guarantees for privacy protection while still enabling data analysis and targeting

Two key targeting strategies were presented:

• DPK strategy that focuses on predicted uplift
• DP policy that uses probabilistic targeting decisions

Privacy risk is quantified using epsilon (ε) parameter:

• Lower ε = stronger privacy protection but less accuracy
• Higher ε = more privacy risk but better targeting precision

There’s an inherent tradeoff between privacy protection and profits:

• Adding noise reduces targeting accuracy and profits
• Larger sample sizes help maintain profits even with stronger privacy protection

Key findings from simulations:

• For large datasets (100k-1M customers), private targeting strategies can approach non-private profit levels
• DPK strategy outperforms random targeting even with strong privacy protection
• Location of privacy implementation matters significantly for maintaining utility

No perfect solution exists for complete privacy protection, but differential privacy allows:

• Quantifiable privacy risk management
• Controlled information release
• Mathematical proof of privacy guarantees

Implementation considerations:

• Privacy protection can be added at different stages (data collection, analysis, targeting)
• Neural networks can be used to represent treatment effect functions
• Sample size impacts effectiveness of privacy protection

Business implications:

• Privacy-aware strategies may attract privacy-conscious customers
• GDPR compliance needs to be balanced with profit objectives
• Companies need to carefully consider privacy risk vs utility tradeoffs