Incident Response Fairness Analysis

This project focuses on analyzing the fairness of incident response processes within an IT company. The analysis is based on an event log extracted from the audit system of an instance of the ServiceNow™ platform used by the organization.

The dataset provides insights into the incident management process and includes the following details:

Number of instances: 141,712 events (24,918 incidents)

Number of attributes: 36 attributes

The data was anonymized for privacy, and information from a relational database was used to enrich the event log.

The data cleaning process involved several steps, including:

Handling missing values in date columns

Dropping columns with many "?" values

Removing rows with "?" values in specific columns

Extracting numeric parts from text columns

Filtering rows based on the 'incident_state' attribute

The exploratory data analysis phase involved visualizing various aspects of the dataset, such as incident priorities, impact, urgency, and resolution times.

The fairness analysis was conducted using the Chi-Square test, which is a statistical method used to determine if there is a significant association between categorical variables. The analysis revealed significant associations between the 'assigned_to' attribute and various factors, such as reassignment count, knowledge, priority confirmation, and resolution time. These findings suggest potential unfairness in the assignment or decision-making processes, highlighting areas for further investigation and improvement.

The project also involved a machine learning component to further explore the fairness of incident handling. Key steps included:

Calculating mean values for specific attributes ('assigned_to', 'reassignment_count', and 'resolution_time')

Selecting relevant columns for fairness analysis

Defining fairness thresholds based on mean values

Creating a binary target variable 'fairness' based on the defined criteria

Preprocessing the data using a ColumnTransformer for one-hot encoding and preserving numerical features

Training a RandomForestClassifier model on the training data

Evaluating the model's performance on the test data

The trained model achieved an accuracy of 86.67% on the test dataset.

demo.mp4

For further exploration and analysis, you can access the dataset and related notebooks on Kaggle.

Click the Kaggle logo above to access the dataset on Kaggle.