Abstract:
"Collaborative filtering has seen considerable success in the areas regarding
information overload and e-commerce, while the current developed systems are
flawed in several respects. Two approaches, the distribution-based algorithm and
the blurring profile solution, are proposed to address several outstanding issues.
The main findings of this research include:
• The proposed distribution-based algorithm, like the traditional nearestneighbor
method, is easy to understand and easy to implement, while it
can provide a confidence level for each predicted rating.
• Compare to the nearest-neighbor method using the Pearson correlation
coefficient as a similarity metric, the distribution-based algorithm
performs significantly better, especially for the sparse dataset, where
less information about the active user and the neighbors are available.
• A technology-based solution, blurring each user’s profile by inserting
new ratings that did not actually exist, is proposed to protect users’
privacy. To quantify privacy provided by different approaches, the
degrees of obscurity and evenly-adding are introduced, where the
former is to protect the users’ main interests while the latter is to protect
the users’ strange interests.
• The proposed approaches can protect the users’ privacy, at the same
time, they can maintain the satisfied accuracy, and in some cases, they
can even increase the prediction accuracy. This mechanism may be used
to fill the missing data, which is required for the most model-based
algorithms..."--Conclusions
