Skip to content

Kabilash01/Distributed-Hyperparameter-Optimization-Analytics-on-HDFS

Repository files navigation

HPO-B Big Data Analytics Project

Overview

A distributed hyperparameter optimization analytics system built on HDFS + Hive + PySpark + MLflow, ingesting 3.27 million real training evaluations from the HPO-B NeurIPS 2021 benchmark.

Key Results

Metric Value
Real HPO evaluations 3,279,050
Search spaces 16
Datasets 97
Bayesian TPE best accuracy 0.9750
Trials to 90% accuracy 10 vs 24 (2.4x faster)
Recommendation hit rate 99%
Surrogate R² test 0.8652
MLflow runs logged 8

Tech Stack

  • HDFS (Hadoop 3.x pseudo-distributed)
  • Apache Hive 4.0 + HiveQL
  • PySpark + MLlib
  • Optuna (TPE + CMA-ES)
  • MLflow experiment tracking
  • scikit-learn RandomForest surrogate

Architecture

HDFS → Hive → PySpark → Bayesian Optimizer → Meta-Learner → MLflow

Dataset

https://github.com/sebastianpinedaar/hpo-data

run this project

ScriptWhat it starts ~/start_project.sh -> Hadoop + HiveServer2 + MLflow ~/start_hive_fast.sh -> HiveServer2 only ~/stop_project.shS -> stops everything

About

Distributed HPO analytics system on HDFS + Hive + PySpark + MLflow ingesting 3.27M real evaluations (HPO-B NeurIPS 2021). Bayesian meta-learning optimizer achieves 2.4x faster convergence with 99% recommendation hit rate across 97 datasets.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages