1. Analysis and Prediction of Monthly City Gas Usage
Samchully is the biggest city gas provider in South Korea. Fittingly, Samchully possesses more than 2 billion usage data tuples in their database, each collected per month per account. Due to its sheer size, there has been difficulties properly analyzing this data so far. This project aimed to analyze that data, assess the factors determining the monthly demand, and predict future demand.
First, the data from Samchully SAP server was read and processed. Preliminary data analysis enabled distinguishing the data by usage patterns, then regression models were applied. The time dependency of the data was taken into account in forming the independent variables. Moreover, monthly weather data was employed to further increase the model accuracy as necessary.
As a result, the final model could reproduce the historical monthly usage with ~4% MAPE and predict the upcoming 3 months' demand. This output was then made available as a visual dashboard, contributing to easy interpretation and reporting.
2. Analysis of Debtor Accounts and Traits
Samchully has been monitoring account finance by storing monthly payment data and assigning a financial stability grade. While the stability grade is calculated based on number of missed/late payments, this project aimed to discover external factors affecting this grade.
From the data on Samchully SAP server, debtor accounts were identified and extracted. By employing EDA and decision tree models, the feature importance of each external factor was quantified.
Prediction of the final financial stability grade of an account was made possible based on the identified factors and their importances. If an account makes a late/missed payment, it is now possible to determine if it is a one-time event or the beginning of habitual late/missed payments.
3. Analysis and Visualization of Customer Service Phone Calls
Samchully performs various customer interactions regarding city gas, safety inspection, bills, etc. This project focuses on phone call interactions, of which Samchully receives hundreds a day. Despite its sheer amount, the calls can be grouped under a handful of themes. This project arose to satisfy the need for efficient phone call process through automatically analyzing the call contents and visualizing the results.
The phone call files saved on Samchully database were loaded then processed into text with an STT(Speech To Text) model. The generated call scripts were then analyzed and visualized.
Samchully is now able to monitor phone call theme distribution and trends. In addition, phone call themes and contents can be appropriately
categorized based on date, extension number, incoming/outgoing, etc. Moreover, phone call logs can be filtered by
specific themes or keywords for review. This allows easy statistics extraction and call detail monitoring.
Automatic scheduling was set up to update the monitoring result daily.