Browsing by Subject "similarity search"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Algorithms for Constructing Exact Nearest Neighbor Graphs
    (2016-06) Anastasiu, David
    Nearest neighbor graphs (NNGs) contain the set of closest neighbors, and their similarities, for each of the objects in a set of objects. They are widely used in many real-world applications, such as clustering, online advertising, recommender systems, data cleaning, and query refinement. A brute-force method for constructing the graph requires O(n^2) similarity comparisons for a set of n objects. One way to reduce the number of comparisons is to ignore object pairs with low similarity, which are unimportant in many domains. Current methods for construction of the graph tackle the problem by either pruning the similarity search space, avoiding comparisons of objects that can be determined to not meet the similarity bounding conditions, or they solve the problem approximately, which can miss some of the neighbors. This thesis addresses the problem of efficiently constructing the exact nearest neighbor graph for a large set of objects, i.e., the graph that would be found by comparing each object against all other objects in the set. In this context, we address two specific problems. The epsilon-nearest neighbor graph (epsilon-NNG) construction problem, also known as all-pairs similarity search (APSS), seeks to find, for each object, all other objects with a similarity of at least some threshold epsilon. On the other hand, the k-nearest neighbor graph (k-NNG) construction problem seeks to find the k closest other objects to each object in the set. For both problems, we propose filtering techniques that are more effective than previous ones, and efficient serial and parallel algorithms to construct the graph. Our methods are ideally suited for sparse high dimensional data.
  • Thumbnail Image
    Item
    Efficiency of Shared-Memory Multiprocessors for a Genetic Sequence Similarity Search Algorithm
    (1997) Chi, Ed Huai-hsin; Shoop, Elizabeth; Carlis, John; Retzel, Ernest; Riedl, John
    Molecular biologists who conduct large-scale genetic sequencing projects are producing an ever-increasing amount of sequence data. GenBank, the primary repository for DNA sequence data, is doubling in size every 1.3 years. Keeping pace with the analysis of these data is a difficult task. One of the most successful technique, for analyzing genetic data is sequence similarity analysis-the comparison of unknown sequences against known sequences kept in databases. As biologists gather more sequence data, sequence similarity algorithms are more and more useful, but take longer and longer to run. BLAST is one of the most popular sequence similarity algorithms in me today, but its running time is approximately proportional to the size of the database. Sequence similarity analysis using BLAST is becoming a bottleneck in genetic sequence analysis. This paper analyzes the performance of BLAST on SMPs, to improve our theoretical and practical understanding of the scalability of the algorithm. Since the database sizes are growing faster than the improvements in processor speed we expect from Moore's law, multiprocessor architectures appear to be the only way to meet the need for performance.