Achieving intelligent task-based mobile widget organization and customization through machine learning techniques and user model generalization.

摘要

In a paper by Bostrom et al., an interface was presented providing rapid access to widget applications on mobile devices. Icons representing widgets were added and removed by users to a "canvas" that enabled them customize the interface to suit their primary objective. Though this implementation was sound and well received, we believed that it could be improved through the combination of two methods grounded in machine learning. These are the generalization of a data set for modeling a default user and a new algorithm named KAWS (K-Based Algorithm for Widget Selection). By accurately predicting potentially desirable widgets and automatically populating the widget canvas, there was potential to mitigate the amount of necessary interaction between the user and device resulting in a diminished physical and cognitive burden. To evaluate the ability of our collected data to form a generalized user model, we ran it against four machine learning algorithms IBK, KStar, Naive Bayes and J48 using 10-fold cross validation. We found that we were able to achieve an average of 56.9 percent correct class predictions while maintaining a relatively low variance and strong kappa statistic. When compared to a purchasing recommendation system and a personal assistant scheduling system that both use collaborative filtering and machine learning techniques to predict user preferences, our data generalization model was consistent with the two who maintained accuracies of around 50 percent. When this data was subsequently run against our new implementation, KAWS, we were able to reduce the average amount of requisite interaction by 11 percent when compared to the implementation by Bostrom et al.