Feature Abstraction

Optimize Your Model

Pyhon software coming soon to GitHub

The software package provides end users an innovative way to fine tune feature hyper parameters and optimize model performance for multi-classification problems. Read through the following documentation to find how to get started building better performing models today.

Introduction

use case & beyond

Fine tune time series filter hyper parameters to maximize feature performance. The following methodology can be used for other  feature optimization methods. The following criteria must be met.

  • The model must be classification

  • Data must continuous

Augment Feature

Optimize Feature

Distribution Analysis

Signal Processing

Calculate Performance

 

Time Series Filter
ema_optimization

The time series filter consists of three hyper parameters every parameter has the no linear domain [0,1]  explained below.

  • Decay  a,  How rapidly the the filter exponentially decreases

  • Shift    b,  The duration of the domain with the maximum weight

  • Delta   c,   Smoothing factor between decay and shift 

Augment Feature

Optimize Feature

Distribution Analysis

Signal Processing

Calculate Performance

 

Time Series Filter Optimization

ema_optimization

To optimize the time series filter hyper parameters bayesian optimization and Gaussian processes are applied to reduce algorithm run time and maximize performance.  In addition the algorithm utilizes a batch design from later works which mimics the optimal sequential policy. 

Decay

Shift

Delta

Bayesian Optimization

Optimize Score

Augment Feature

Optimize Feature

Distribution Analysis

Signal Processing

Calculate Performance

 

Distribution Analysis

distribution_optimizer

Transform raw data to a gaussian distributed space using a performance metric. If the data has been found to be gaussian distributed it can be used to:

  • Score feature performance in multi-classification models

  • Dimensionality reduction via gaussian processes

Augment Feature

Optimize Feature

Distribution Analysis

Signal Processing

Calculate Performance

 

Signal Processing

distribution_distillation

The signal processing class uses Bayesian optimization to filter noise and isolate key areas of discontinuity between multi-class distributions. 

 

 

 

 

The area under the curve is then used to weight the importance of the probability the correct multi-class distribution was identified.

 

weighted by the  of the optimization heuristic is derived from discontinuity between classification distributions being able to correctly identify the appropriate classification.