Lifelong Learning for Dynamic Churn Prediction

Learning continually, accumulating knowledge, and using it to learn new tasks was the characteristic of lifelong learning. Alternatively, also known as Continual learning, the concept takes benefit from the one that called the previous tasks to solve the new task-this schema can learn on a series of tasks across diverse sources (or domains or even modality) through the selection of the right mechanism. Elastic weight consolidation (EWC) method proposed by Google Deepmind, provides a way of calculating the importance of weight preserving the previously acquired information then selectively adjusts the plasticity. The term plasticity is the main reason for catastrophic forgetting since the weight that previously learned in the first task can be easily modified or even abruptly lost. The probability perspective underlies the formation of the EWC; the method assumes that the second task's solution is closely related to the solution space for the first task. In this paper, the EWC exploited to tackle the sequence tasks of predicting churn activity. The tasks involve two distinct datasets as two sequential tasks from the domain of Telecom. Learning sequential tasks without applying the lifelong learning concept shows the worst result with no sign of recovery. However, the experiment that applies the EWC method shows enhancement (maximizing accuracy while suppressing the loss value) on both task performance. Lifelong learning offers a more flexible way of learning to further research in dynamic learning.