Online machine learning

Stream mining

• Conventional machine learning assumes that the training data is available as a whole set.
• In stream mining, core assumption is that training examples can be briefly inspected for a single time only.
• Arrive in a high speed stream and then must be discarded to make room for subsequent examples.
• Algorithm must update its model incrementally as each element is inspected.
• It should ideally be able to apply the model any time between training examples.
• the goal of classification is to produce a model that can predict the class of unlabelled examples by training on examples whose label/class is supplied.
• Data is assumed to have small and fixed no. of columns or attributes/features. This can be thought of as tuples.
• The number of examples (rows) are very large and they are assumed to keep arriving in a stream infinitely.
• The labels that can be applied to the data are limited, typically less than ten.

Approaches

• Wrapper approach : collect a stream into a batch and then use batch ML. Not very good.
• Adaptation approach: custom built for stream learning.
  • Decision Tree: Good choice.
  • Neural Networks: Also good.