In the current world, data on user preferences and behaviour is crucial in creating and promoting products with the goal of improving the experience of customers. However, user decisions concerning a product (e.g. buying a subscription or changing carriers) depend on exogenous factors, rendering them inherently difficult. The question becomes how to model the essential challenge of customer understanding.
Current efforts lie in optimising each task separately, e.g. a model for detecting customers who wish to upgrade a subscription and another model for a different behaviour. This approach neglects the fact that user decisions share an underlying structure, e.g. what is the relationship between the tasks of upgrading a subscription and changing carriers? By casting the problem as multi-task learning, we are able to extract the common elements in the user behaviour across their usage of different products and consrtuct a unifying understanding of the user base. The hypothesis underlying this approach is that the joint training will yield results superior to individual models.
In multi-task learning, each task serves as a different perspective to the problem. This is particularly useful in data concerning users, since they have considerable amounts of noise. Hence, each task is characterised by different noise patterns, the multi-task formulation can reap the benefits of the multifaceted understanding of the problem, yielding better performance and generalisation properties.
Overall, the project focuses on:
- creating a common embedding for tasks. In this context, we aim for the improvement of all tasks involved via joint optimization
- understanding the positive transfer dynamics of tasks, i.e. understanding which tasks should be learned together
- detecting user behaviour via the different perspective each task provides