Improving k-nn search and subspace clustering based on local intrinsic dimensionality

In several novel applications such as multimedia and recommender systems, data is often represented as object feature vectors in high-dimensional spaces. The high-dimensional data is always a challenge for state-of-the-art algorithms, because of the so-called curse of dimensionality . As the dimensionality increases, the discriminative ability of similarity measures diminishes to the point where many data analysis algorithms, such as similarity search and clustering, that depend on them lose their effectiveness. One way to handle this challenge is by selecting the most important features, which is essential for providing compact object representations as well as improving the overall search and clustering performance. Having compact feature vectors can further reduce the storage space and the computational complexity of search and learning tasks. Support-Weighted Intrinsic Dimensionality (support-weighted ID) is a new promising feature selection criterion that estimates the contribution of each feature to the overall intrinsic dimensionality. Support-weighted ID identifies relevant features locally for each object, and penalizes those features that have locally lower discriminative power as well as higher density. In fact, support-weighted ID measures the ability of each feature to locally discriminate between objects in the dataset. Based on support-weighted ID, this dissertation introduces three main research contributions: First, this dissertation proposes NNWID-Descent, a similarity graph construction method that utilizes the support-weighted ID criterion to identify and retain relevant features locally for each object and enhance the overall graph quality. Second, with the aim to improve the accuracy and performance of cluster analysis, this dissertation introduces k-LIDoids, a subspace clustering algorithm that extends the utility of support-weighted ID within a clustering framework in order to gradually select the subset of informative and important features per cluster. k-LIDoids is able to construct clusters together with finding a low dimensional subspace for each cluster. Finally, using the compact object and cluster representations from NNWID-Descent and k-LIDoids, this dissertation defines LID-Fingerprint, a new binary fingerprinting and multi-level indexing framework for the high-dimensional data. LID-Fingerprint can be used for hiding the information as a way of preventing passive adversaries as well as providing an efficient and secure similarity search and retrieval for the data stored on the cloud. When compared to other state-of-the-art algorithms, the good practical performance provides an evidence for the effectiveness of the proposed algorithms for the data in high-dimensional spaces