Dynamically updated parallel k-NN

Dynamically updated parallel k-NN

• 1.
  Dynamically updated parallelk-NN search algorithm using MPI
• 2.
  Keywords:  MPI (MessagePassing Interface)  Machine Learning  Classification  k-NN  Clustering  K-means
• 3.
  # MPI (Message Passing Interface)  MPI is an industrial standard that specifies library routines needed for writing message passing programs  MPI uses a library approach to support parallel programming
• 4.
  # Machine Learning Ability of a system to learn without being coded.  To get more proficient in performing a task from being familiar, with the time.
• 5.
  # Classification  Supervisedmachine Learning Approach.  Supervised in the sense that we know the class of the training instances to which it belongs.  To predict the class of a test instance on the basis of some similarity measure.
• 6.
  k-NN Classifier  Afamous classification approach  Based on the assumption that a instance has similar class to which it closest or close in the feature space.  To classify a test instance ,find its k-nearest neighbor according to some kind of similarity . Classify on the basis of majority .
• 7.
  # Clustering  UnsupervisedMachine Learning Approach.  Unsupervised as we don’t know the class of the training instances to which they belong.  Partition the data so as instances in one group are more similar but too different from the others.
• 8.
  K-means Clustering  Aclustering approach , to cluster the training instances into k-number of cluster. The value of k being provided by the user.  Clustering is done by a utilizing a SSE (sum of square error) function.
• 9.
  Problems with thek-NN  Time complexity  Sensitive to the local structure of the data.  Curse of dimensionality.
• 10.
  # Our ProposedApproach to solve k-NN parallel using MPI.  Pre-processing step  Perform clustering process on training set to divide it into p mutually exclusive partition {P1,P2,…,Pp}, where p is number of process .  Create the Representative Instance to represent each partition
• 11.
  # Step- IInd Fori =1 to p  Apply k-means approach  Evaluate nearest neighbor similarity of training instances with representative instance(centroid) of each partition.  Perform  Competence Enhancement – Repeated Wilson Editing Rule (noise removal)  Competence Preservation (removal of superfluous instance)  Store the outliers of each cluster separately.  Update the centroid of the cluster.  Repeat step I & II until number of instances in the selected one partition >=k.
• 12.
  # Step- IIIrd Take a test instance .  Select the partition whose R.I is closest to test instance.  Repeat until reach at the last sub-partition.  Apply the majority rule.  Select the class label who has majority for the test instance.
• 13.
  # Updation oftraining set  When the similarity value of the new test instance with the R.I. of the different partition exceeds the max radius value which we store during the pre-processing step.  Update the R.I. of that partition only, which is closed to the new test instance
• 14.
  # Research papersconsidered to design the dynamically updated parallel k-NN using MPI. # For Preprocessing (Clustering process ):  Efficient and Fast Initialization Algorithm for K-means Clustering By Mohammed El Agha, Wesam M. Ashour , Islamic University of Gaza, Gaza, Palestine  A new algorithm for initial cluster centers in k-means algorithm By Murat Erisoglu, Nazif Calis, Sadullah Sakallioglu Department of Statistics, Faculty of Science and Letters, Cukurova University, 01300 Adana, Turkey  An empirical comparison of four initialization methods for the K-Means algorithm By J.M. Pe~na *,1, J.A. Lozano, P. Larra~naga, Department of Computer Science and Artificial Intelligence, Intelligent Systems Group, University of the Basque Country, P.O. Box 649,E-20080 San Sebastian, Spain
• 15.
  # For findingk-NN and removal of Noise and Superfluous Instances.  Fast Condensed Nearest Neighbor Rule By Fabrizio Angiulli ICAR-CNR, Via Pietro Bucci 41C, 87036 Rende (CS), Italy  Advances in Instance Selection for Instance-Based Learning Algorithms By HENRY BRIGHTON, Language Evolution and Computation Research Unit, Department of Theoretical and Applied Linguistics, The University of Edinburgh, Edinburgh, EH8 9LL, UK  CHRIS MELLISH, Department of Artificial Intelligence, The University of Edinburgh, Edinburgh EH1 1HN, UK
• 16.
   Superlinear Parallelizationof k-Nearest Neighbor Retrieval By Antal van den Bosch Ko van der SlootILK Research Group Dept. of Communication and Information Sciences, Tilburg University, P.O. Box 90153, NL-5000 LE Tilburg, The Netherlands  Parallel Algorithms on Nearest Neighbor Search By BERKAY AYDIN, Georgia State University  K-Nearest-Neighbor Consistency in Data Clustering: Incorporating Local Information into Global Optimization By Chris Ding and Xiaofeng He  Instance-based classifiers applied to medical databases: Diagnosis and knowledge extraction By Francesco Gagliardi, Department of Philosophy, University of Rome
• 17.
  Thank You!...