This disclosure describes graph-based privacy-aware platforms for serving features to machine learning models, optimized to meet constraints of privacy, model quality, multi-granularity, incremental update, overlapping features, scale, etc. Privacy is preserved by grouping features into a collection, subject to lower-bound constraints, such that features corresponding to individual users are rendered unrecoverable. Such feature aggregation does not prevent generation of personalized recommendations, as model quality is maintained by aggregating similar, rather than arbitrary features, subject to upper-bound constraints. Upper and lower bounds can vary to account for differing privacy sensitivities and feature importance with respect to a given model (multi-granularity). Features can belong to multiple clusters (overlapping features). Incremental feature grouping is used to enable efficient dynamic feature grouping. Feature serving, as described herein, is of low complexity, such that features can be served to a large number of clients at minimum latency without degraded performance.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.