Machine Learning

As part of my doctoral school course on web retrieval and data mining, Hugo Zaragoz (now at Yahoo! Research) gave a crash course on machine learning in the context of data mining. A few notes:

Machine learning could be inspiring in the context of ubicomp systems. For example in improving context detection. However one current challenge is “unsupervised learning”, that is when the features of the target result are not known and only assumed (that’s often the case in uncontrolled ubicomp system). There are 4 major way to handle uncertain information:

• Baysian statistics (Gaussian processes)
• Probalistic models (Graphical model)
• Statistical learning (SVM (Support Vector Machine) perception)
• Symbolic learning

For model validation, it is important not to test the model with the data used for training the model.

Relation to my thesis: Machine learning is certainly a way for systems to become more context-aware and make less errors. As I might end-up modeling uncertainty in the context of collaboration in a ubicomp system, I got a few clues on how to properly validate a model (not mixing training set of data with test sets).