Online Transfer Learning for Concept Drifting Data Streams

Bespoke applications provide personalised functionalities to user but can be expensive to produce. To reduce costs, off-the-shelf applications deployed in real-world environments that evolve and change over time use manual input to tailor functionalities to user's personal preferences. Manual inputs must be frequently updated and maintained to ensure they effectively represent the user's desires.
Online machine learning (OML) techniques can be used to mitigate manual inputs, predicting personalised preferences from sensing the surrounding environment. User preferences and real-world environments are dynamic in nature, causing concept drifts that require learning algorithms to update or change their predictive models to remain effective. Within online environments, too few instances of data may be available to build an effective model. Transfer learning can be used to overcome this by sharing knowledge previously learnt from other users, referred to as a source domain, to improve the performance of predictions of new users, known as the target domain. This enables existing models to be used before sufficient data is collected in the target domain to learn the current concept.
Existing online transfer learning (OTL) techniques consider a source to be in an offline environment, however for many applications it may be preferable to consider the source to be in an online environment. The first novel contribution of this project is to develop a framework to transfer knowledge between an online source and target such that knowledge of newly encountered concepts in the source can be benefited from by the target. In addition to developing the framework to achieve OTL, a model weighting mechanism will be presented to allow source models to be used by the target predictor. The mechanism used will combine knowledge of concepts learnt across domains, increasing the influence a model has if it is more beneficial to the target learner.
The methodology developed will be extended to transfer knowledge of concept transitions. This information is used by OML algorithms such as RePro to enable pro-active predictions of future concepts. By transferring knowledge of concept transitions the online target learner is able to anticipate, rather than react to, concept drift.
Due to the dynamic nature of the source and target domains, knowledge initially transferred from source may impact the target learner in a different way as the environments evolve. Due to the changing environments, the knowledge transferred may become irrelevant to the target learner, the transfer of irrelevant or detrimental knowledge is known as negative transfer. To minimise negative transfer, a novel approach will be developed that re-evaluates the mapping between source and target domains, ensuring knowledge transfer is relevant and useful.
0-6 months: Related work research on OML and concept drift
6-12 months: Create a synthetic data generator using a representative application (smart home heating system). Implement OML algorithm (RePro)
12-18 months: Research existing OTL algorithms. Implement state-of-the-art OTL algorithm (Generalized Online Transfer Learning)
18-24 months: Design an OTL framework for concept drifting data streams where source and target in online environments. Create a synthetic data generator to simulate a drifting hyperplane
24-30 months: Research into distributed systems to consider when to transfer knowledge and how this relates to applications (discuss use case scenarios with industrial partner, JLR). Develop a weighting mechanism that evolves over time, allowing multiple models to be used by the target
30-36 months: Extend the novel OTL framework to transfer knowledge of concept transitions for proactive concept drift detection
36-42 months: Develop a novel transformation function to prevent negative transfer
42 months+: Thesis write up