anmilleriii / multilabel-lsvm

Linear Support Vector Machine (LVSM) multilabel classifier to identify trends and deficiencies in utility equipment failure mitigation strategies

Disclaimer

Utility specific information regarding equipment and maintenance programs has been redacted, rendering this view-only.

Purpose

The use case is classification of electric plant (utility) equipment failure mitigation strategies into general categories to identify trends and deficiencies.

Background

This model classifies recods in the free-text 'Mitigation' field of a plant equipment database (71,000 records) into based on 'Mitigation' strategy type. Records are binned into one or multiple of the following classes:

• maintenance
• operational
• physical_barrier
• design_and_engineering
• supply_chain
• unknown

The NLP classifier applies a Linear Support Vector Classification (SVC) algorithm (https://scikit-learn.org/stable/modules/generated/sklearn.svm.LinearSVC.html).

LVSC was selected based on trials using a number of text-classification algorithms, including:

• Random Forrest
• Naive Bayes
• Linear Regression

Moreover, this classifier applies a multi-label, "One vs Rest" (also known as 'binomial classifiication') strategy, which iteratively applies a seperate LVSM classifier for each label.

Classifier Setup

1. Clone the repository

git clone https://github.com/ANMillerIII/LVSM.git

2. Initialize and activate vitual environment

py -m venv venv

./venv/Scripts/activate

3. Install dependencies

py -m pip install requirements.txt -r

4. Switch to "LVSM" directory

cd LVSM

Run Classifier

To run the 'Mitigation' classifier

py ./1_mitigation/mitigation_model.py

Prediction output will be in the respective 'out' directories

Limitations

1. 'train_set' data set is used based on naive string matching with some oversight. This should be improved by manually classifying a greater sample of 'Mitigation' fields by hand.
2. Accuracy is not a meaningful metric for the 'apply_set' data, since the model is used to make predictions (i.e., must spot check manually)
3. Foreign-language entries are classified with low fidelity due to lack of manual classification.
4. 'train_set' data (3,569 records) are not classified by NLP, but rather string-matching/manually with inherent biases.