Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

  • Summary

    This document provides the release note for federated ml applicatons at edge.


    what is released 

    components of the release (Akraino new)

    • Refactored Hetero FTL with optional communication-efficiency mechanism, with 4x time efficiency improvement
    • Hetero SecureBoost supports complete secure mode
    • Hetero SecureBoost now can reduce time consumption over highly sparse data by using sparse matrix
      computation on histogram aggregations.
    • Hetero SecureBoost optimization: the communication round in prediction is reduced to no larger than tree depth,
      prediction speed is improved by 32 times in a 100-tree model.
    • Addition of Hetero FastSecureBoost module, whose mixed/layered modeling method makes it twice as efficient as SecureBoost
    • Improved Hetero Federated Binning with 30%~50% time efficiency improvement
    • Better GLM: >10% improvement in time efficiency
    • FATE first unsupervised learning algorithm: Hetero KMeans
    • Upgraded Hetero Feature Selection: add PSI filter and SecureBoost feature importance filter
    • Add Data Split module: splitting data into train, validate, and test sets inside FATE modeling workflow
    • Add DataStatistic module: compute min/max, mean, median, skewness, kurtosis, coefficient of variance, percentile, etc.
    • Add PSI module for computing population stability index
    • Add Homo OneHot module for one-hot encoding in homogeneous scenario
    • Evaluation module adds metrics for clustering
    • Optional FedProx mechanism for Homo LR, useful for training with non-iid data
    • Add Oblivious Transfer Protocol and OT-based module Secure Information Retrieval
    • Random Iterative Affine protocol, providing additional security

    dependencies of the release (upstream version, patches)

         still the same

    differences from previous version

        add several new modules in FATE
    • SecureBoost: more efficient computation with GOSS, histogram subtraction, cipher compression, 2-4x faster
    • Hetero GLM: improved communication efficiency, adjustable floating point precision, 2x faster
    • Hetero NN: adjustable floating point precision, support SelectiveBackPropagation and dropOut on interaction layer, 2x faster
    • Hetero Feature Binning: improved algorithm with cipher compression, 2x faster
    • Intersect: add split calculation option and adjustable random base fraction, 30% faster
    • Homo NN: restructure torch backend and enhanced grammar; train and predict with raw image data
    • Intersect supports SM3 hashing method
    • Hetero SecureBoost: L1 penalty & adjustable min_child_weight to prevent overfitting
    • NEW SecureBoost Transformer: feature engineering module that encodes instances with leaf nodes from SecureBoost model
    • Hetero Pearson: support local VIF computation
    • Hetero Feature Selection: support selection based on VIF and Pearson
    • NEW Homo Feature Binning: support virtual/recursive binning strategy
    • NEW Sample Weight: set sample weights based on label or from feature column, Hetero GLM & Hetero SecureBoost support weighted training
    • NEW Data Transformer: case-insensitive on data schema
    • Local Baseline supports prediction task
    • Cross Validation: output fold split history
    • Evaluation: add multi-result-unfold option which unfolds multi-classification evaluation result to several binary evaluation results in a one-vs-rest manner

              accelerate the computing for hetero, horizontal, and feature binning method