Reduction of Concept Drift Effect on Mobile Network Fraud Detection Using Deep Learning Markov Transition Encoding Dynamic LSTM Approach

Background: Mobile network infrastructure has exponentially advanced in the past 20 years, resulting in tremendous evolution of the capabilities/services that the network can deliver. With these advancements, fraud activities have in equal measure metamorphosed into a complex web involving and spanning multiple touchpoints. Although different machine learning approaches have been developed and adopted, mobile network fraud remains dynamic and continually changes in form and means with time, resulting in degraded efficiency of these ML models. The phenomenon resulting in this degradation of efficiency with time is known as concept drift, which can be described as changes in the conditional distribution of the target variable, which is also referred to as output given the input features or the inputs, while the distribution of the inputs may still stay unchanged. This study aimed to analyze the effect of concept drift in relation to mobile network fraud detection, propose and develop an approach that reduces the concept drift effect when classifying and detecting mobile network fraud. Method: A quantitative research methodology was adopted using experiments as the core method. The dataset used for the study was an extract of DataStream from a Kenyan telco company and contains about 400,000 events or records generated by a process that randomly selected 65000 events for each month (Jan 2023 to June 2023) from the complete dataset. A deep learning model known as MTED-LSTM was developed and trained, then evaluated using the 6-month data stream. The other four commonly used approaches, i.e., Random Forest, Naïve Bias, Support Vector Machine, and Logistic Regression, were also developed and evaluated using the same data stream. Eight evaluation criteria were used to evaluate the effectiveness of the models. The data stream was used to evaluate whether the models were affected by the drift and how good they were in handling the concept drift effect. Results: Using each evaluation criterion, the models were tested through each month's data stream without retraining the model. The result indicated a gradual reduction of the effectiveness of the models with time, attributed to the concept drift effect. MTED-LSTM model showed to be more effective through the six-month data stream and least affected by concept drift as compared to other models developed in the study. Conclusion: The MTED-LSTM model effectively managed the effects of concept drift better due to its ability to store the previous model state on the time series DataStream, which is also referred to as memory. This is evidenced by the results, where all the performance indicators show that the model is least affected by the drift with time. It is worth noting that between the 4th and 5th months, all the other models drastically reduce their effectiveness, but the MTED-LSTM model recorded improved effectiveness during this time frame, resulting in a stable and smooth performance.