top of page

What makes a complaint different to a non-complaint: A tale of a complaint detector

A food redistribution platform needed a way to analyse and quantify customer complaints, knowing that complaint handling strategies are key drivers for creating satisfying customer relationships.


Given the urgency of climate change and pandemics our clients services are needed more than ever. However, more customers also brought more complaints on their app - therefore, a company approached Kubik to help them find a way to detect complaints in a timely manner in order to alleviate some of the main consumer pain points. However, their text data was unstructured, high in volume and a novel approach was needed in order to create an effective machine learning model that successfully distinguishes between complaints and non-complaints.


Data used in this challenge was in the form of short text data, similar to tweets. We first analysed a corpus of text documents from the company's discussion forum, in order to obtain a better understanding of the reasons behind complaints. We then developed a machine learning model to identify complaints based on linguistic and behavioural features.

The texts were written in a casual language, often using abbreviations and emojis - however, natural language techniques were able to detect some of the main features that makes complaints different to non-complaints: For example, complaints were slightly longer (usually because people would describe an issue and how they feel about it) and naturally, complaints were also more negative in nature and contained more 'angry' words. In total, 15 features were used in this predictive modelling and Logistic Regression and Random Forest had the most successful prediction accuracy. Following these results, the remaining data set was labelled as either complaints or non-complaints, which provided us with the opportunity to further examine the reasons behind complaints (e.g. rudeness). More importantly, this work provided the company with a great basis to deploy our model and develop a software to 'capture' complaints more timely and efficiently, based on the model and features we used.


Complaints can be seen as a significant aspect of customer relationship management and a 'mirror' in which a company may see the impacts of its performance and receive beneficial feedback. Additionally, being able to respond to customers in an effective and timely manner, via such mechanisms, can help move services nearer to their customers and cultivate a positive reputation for the companies themselves. Following the implementation of the complaints’ classifier on their platform, our clients can improve delivery of their service and minimise issues of collective action specific to food redistribution.

If you would like to find out more about this case study please contact us.


bottom of page