A South-East Asian Airline Turned Around Maintenance Activity, Based On Customer Feedback & Classification






Predictive Classification Model

Categorization of feedback into desired issue categories achieved with a 90% accuracy.

5x organization-wide time-savings in NLP categorization project time, as our data scientists didn’t have to re-work on codes and metrics but rather on driving results. Previously, categorizing 200 feedbacks took the Airline about 10-12 hrs which is now reduced to 1-2 hrs. Simply put, codes and flow were modularized, so it was reused across the organization.

The engineering team, now, receives a live report to identify the aircrafts or facilities/services which need immediate attention based on complaints being made.

A South-East Asian Airline needed help in addressing customer complaints and improving the overall travel experience. Typically, the team would have to manually go through each feedback to identify a specific line of maintenance for each aircraft, which was a very time-consuming process.

TheMathCompany partnered with the airline to reclassify issues, establish multi-level categorization to help the engineering team with the automation of the complaint classification process and also with prediction. Natural Language Processing (NLP) techniques to identify issue categories and fast track this process was suggested as a solution.

Multi-label feedback was a challenge that the airline was yet to tackle. We used multiple classification techniques to build a predictive classification model for each category.

Data cleaning - Ensured all text was cleaned to remove unnecessary characters (stop words, punctuation, proper nouns, etc.), lemmatization and stemming of words.

Feature creation - Document term matrix on text corpus and n-grams


  • Two incongruous levels (L1& L2)
  • Built a model for each L2 category, post considering L1 classification
  • We added synthetic feedback based on words, for categories with a very low count of feedback (class imbalance).
  • For feedback with low confidence scores, we decided to consider them under the unclassified category which would be manually classified later and fed back to the model as input to train further.
  • Deployment - The model runs monthly post collection of all feedback from various sources like emails, social media, website, etc. The UI gives the business a view of the feedbacks and categories they fall under. Users can review and re-classify incorrect feedback, which would be fed into the model to re-train the same.

Related resources

Subscribe to Authentic Intelligence – our monthly newsletter

Stay up to date with the latest marketing, sales, and service tips and news.