Algorithms for Similarity Search

Total Page:16

File Type:pdf, Size:1020Kb

Algorithms for Similarity Search Outline Similarity Search Algorithms for 1 High-level overview Similarity Search 2 Locality-sensitive hashing 3 Combinatorial approach Yury Lifshits Yahoo! Research Content Optimization ESSCASS 2010 1 High-level overview 2 Explore/exploit for Yahoo! frontpage 3 Ediscope project for social engagement analysis 1 / 27 2 / 27 Similarity Search: an Example More Formally Input: Set of objects Search space: object domain U, similarity function σ Task: Preprocess it Input: database S = fp1;:::; png ⊆ U Query: q 2 U Task: find argmaxpi σ(pi ; q) p2 Most similar p5 p4 p3 Query: New object q Task: Find the most p p similar one in the dataset 1 6 3 / 27 4 / 27 Applications (1/5) Information Retrieval Applications (2/5) Machine Learning Content-based retrieval (magnetic resonance images, tomography, CAD shapes, time series, texts) Spelling correction kNN classification rule: classify by majority of k nearest training examples. E.g. recognition of faces, Geographic databases (post-office problem) fingerprints, speaker identity, optical characters Searching for similar DNA sequences Nearest-neighbor interpolation Related pages web search Semantic search, concept matching 5 / 27 6 / 27 Applications (3/5) Data Mining Applications (4/5) Bipartite Problems Recommendation systems (most relevant movie to a Near-duplicate detection set of already watched ones) Plagiarism detection Personalized news aggregation (most relevant news articles to a given user's profile of interests) Computing co-occurrence similarity (for detecting synonyms, query extension, machine translation...) Behavioral targeting (most relevant ad for displaying to a given user) Key difference: Mostly, off-line problems Key differences: Query and database objects have different nature Objects are described by features and connections 7 / 27 8 / 27 Applications (5/5) As a Subroutine Variations of the Computation Task Additional requirements: Dynamic nearest neighbors: moving objects, deletes/inserts, changing similarity function Coding theory (maximum likelihood decoding) Related problems: Nearest neighbor: nearest museum to my hotel MPEG compression (searching for similar fragments Reverse nearest neighbor: all museums for which my hotel is the nearest in already compressed part) one Range queries: all museums up to 2km from my hotel Clustering Closest pair: closest pair of museum and hotel Spatial join: pairs of hotels and museums which are at most 1km apart Multiple nearest neighbors: nearest museums for each of these hotels Metric facility location: how to build hotels to minimize the sum of \museum | nearest hotel" distances 9 / 27 10 / 27 Brief History Ideal algorithm 1908 Voronoi diagram + Any data model 1967 kNN classification rule by Cover and Hart + Any dimension 1973 Post-office problem posed by Knuth + Any dataset 1997 The paper by Kleinberg, beginning of provable + Near-linear space upper/lower bounds + Logarithmic search time 2006 Similarity Search book by Zezula, Amato, Dohnal and Batko + Exact nearest neighbor 2008 First International Workshop on Similarity Search + Zero probability of error 2009 Amazon acquires SnapTell + Provable efficiency 11 / 27 12 / 27 Nearest Neighbors in Theory Theory: Four Techniques Branch and bound Greedy walks (p1; p2; p3; p4; p5) Sphere Rectangle Tree Orchard's Algorithm k-d-B tree Geometric near-neighbor access tree Excluded middle vantage point (p1; p2; p3) (p4; p5) forest mvp-tree Fixed-height fixed-queries tree AESA p2 ∗ p Vantage-point tree LAESA R -tree Burkhard-Keller (p1; p3) p2 p4 p5 3 p4 tree BBD tree Navigating Nets Voronoi tree Balanced aspect ratio q p1 p3 p1 tree Metric tree vps -tree M-tree Locality-Sensitive Epsilon nets Hashing SS-tree R-tree Spatial approximation tree Mappings: LSH, Works for small intrinsic dimension Multi-vantage point tree Bisector tree mb-tree Cover tree Hybrid tree random projections, minhashing Generalized hyperplane tree Slim tree Spill Tree Fixed queries tree X-tree k-d tree Balltree Quadtree Octree Post-office tree 13 / 27 14 / 27 Future of Similarity Search Cloud implementation Locality Sensitive Hashing Hadoop stack API (as in OpenCalais, Google Language API, WolframAlpha) Focus on memory 15 / 27 16 / 27 LSH vs. Ideal algorithm Approximate Algorithms c-Approximate r-range query: if there at least one { Any data model p 2 S : d(q; p) ≤ r return some p0 : d(q; p0) ≤ cr { Exact nearest neighbor { Zero probability of error r cr + Provable efficiency q + Any dimension + Any dataset ∗ c-Approximate nearest neighbor query: return some + Near-linear space 0 0 p 2 S : d(p ; q) ≤ crNN , where rNN = minp2S d(p; q) +∗ Logarithmic search time Today we consider only range queries 17 / 27 18 / 27 Today's Focus Data model: d-dimensional Euclidean space: d R Locality-Sensitive Hashing: Our goal: provable performance bounds General Scheme Sublinear search time, near-linear preprocessing space Still an open problem: approximate nearest neighbor search with logarithmic search and linear preprocessing 19 / 27 20 / 27 Definition of LSH The Power of LSH Indyk&Motwani'98 Locality-sensitive hash family H with parameters Notation: ρ = log(1=P1) < 1 log(1=P2) (c; r; P1; P2): Theorem If kp − qk ≤ r then PrH[h(p) = h(q)] ≥ P1 Any (c; r; P1; P2)-locality-sensitive hashing leads to an If kp − qk ≥ cr then PrH[h(p) = h(q)] ≤ P2 algorithm forc-approximater-range search with (roughly)n ρ query time andn 1+ρ preprocessing space Proof in the next four slides 21 / 27 22 / 27 LSH: Preprocessing LSH: Search Composite hash function: g(p) =< h1(p);:::; hk (p) > 1 Compute g1(q);:::; gL(q) Preprocessing with parameters L; k: 2 Go to corresponding buckets and explicitly check 1 Choose at random L composite hash functions of k d(p; q) ≤?cr for every point there components each 3 Stopping conditions: (1) we found a satisfying 2 Hash every p 2 S into buckets g1(p);:::; gL(p) object or (2) we tried at least3 L objects Preprocessing space: O(Ln) Search time is O(L) 23 / 27 24 / 27 LSH: Analysis (1/2) LSH: Analysis (2/2) 1 − x ≤ e−x for x 2 [0; 1] Pk n = 1 L = P−k (−logδ) In order to have probability of error at most δ we set k; L such that: 2 1 k The expected number of cr-far objects to be tried is P2 Ln = L The chance to never be hashed to the same bucket as q for a Solution: k log n k L −P1 L true r-neighbor (1 − P1 ) ≤ e is at most δ k = log(1=P2) log n − log(1=P1) log(1=P2) log(1=P ) ρ Rewriting these constraints: L = P1 log(1/δ) = n 2 log(1/δ) = n log(1/δ) k −k P2 n = 1 L = P1 (−logδ) Preprocessing space O(Ln) ≈ n1+ρ+o(1) Search O(L) ≈ nρ+o(1) 25 / 27 26 / 27 LSH Based on p-Stable Distributions Datar, Immorlica, Indyk and Mirrokni'04 p · r h(p) = b + bc w r: random vector, every coordinate comes from N(0; 1) b: random shift from [0; 1] w: quantization width 1 ρ(c) < c 27 / 27.
Recommended publications
  • Lecture 04 Linear Structures Sort
    Algorithmics (6EAP) MTAT.03.238 Linear structures, sorting, searching, etc Jaak Vilo 2018 Fall Jaak Vilo 1 Big-Oh notation classes Class Informal Intuition Analogy f(n) ∈ ο ( g(n) ) f is dominated by g Strictly below < f(n) ∈ O( g(n) ) Bounded from above Upper bound ≤ f(n) ∈ Θ( g(n) ) Bounded from “equal to” = above and below f(n) ∈ Ω( g(n) ) Bounded from below Lower bound ≥ f(n) ∈ ω( g(n) ) f dominates g Strictly above > Conclusions • Algorithm complexity deals with the behavior in the long-term – worst case -- typical – average case -- quite hard – best case -- bogus, cheating • In practice, long-term sometimes not necessary – E.g. for sorting 20 elements, you dont need fancy algorithms… Linear, sequential, ordered, list … Memory, disk, tape etc – is an ordered sequentially addressed media. Physical ordered list ~ array • Memory /address/ – Garbage collection • Files (character/byte list/lines in text file,…) • Disk – Disk fragmentation Linear data structures: Arrays • Array • Hashed array tree • Bidirectional map • Heightmap • Bit array • Lookup table • Bit field • Matrix • Bitboard • Parallel array • Bitmap • Sorted array • Circular buffer • Sparse array • Control table • Sparse matrix • Image • Iliffe vector • Dynamic array • Variable-length array • Gap buffer Linear data structures: Lists • Doubly linked list • Array list • Xor linked list • Linked list • Zipper • Self-organizing list • Doubly connected edge • Skip list list • Unrolled linked list • Difference list • VList Lists: Array 0 1 size MAX_SIZE-1 3 6 7 5 2 L = int[MAX_SIZE]
    [Show full text]
  • Cover Trees for Nearest Neighbor
    Cover Trees for Nearest Neighbor Alina Beygelzimer [email protected] IBM Thomas J. Watson Research Center, Hawthorne, NY 10532 Sham Kakade [email protected] TTI-Chicago, 1427 E 60th Street, Chicago, IL 60637 John Langford [email protected] TTI-Chicago, 1427 E 60th Street, Chicago, IL 60637 Abstract The basic nearest neighbor problem is as follows: We present a tree data structure for fast Given a set S of n points in some metric space (X, d), nearest neighbor operations in general n- the problem is to preprocess S so that given a query point metric spaces (where the data set con- point p ∈ X, one can efficiently find a point q ∈ S sists of n points). The data structure re- which minimizes d(p, q). quires O(n) space regardless of the met- ric’s structure yet maintains all performance Context. For general metrics, finding (or even ap- properties of a navigating net [KL04a]. If proximating) the nearest neighbor of a point requires the point set has a bounded expansion con- Ω(n) time. The classical example is a uniform met- stant c, which is a measure of the intrinsic ric where every pair of points is near the same dis- dimensionality (as defined in [KR02]), the tance, so there is no structure to take advantage of. cover tree data structure can be constructed However, the metrics of practical interest typically do in O c6n log n time. Furthermore, nearest have some structure which can be exploited to yield neighbor queries require time only logarith- significant computational speedups.
    [Show full text]
  • Faster Cover Trees
    Faster Cover Trees Mike Izbicki [email protected] Christian Shelton [email protected] University of California Riverside, 900 University Ave, Riverside, CA 92521 Abstract est neighbor of p in X is defined as The cover tree data structure speeds up exact pnn = argmin d(p;q) nearest neighbor queries over arbitrary metric q2X−fpg spaces (Beygelzimer et al., 2006). This paper makes cover trees even faster. In particular, we The naive method for computing pnn involves a linear scan provide of all the data points and takes time q(n), but many data structures have been created to speed up this process. The 1. A simpler definition of the cover tree that kd-tree (Friedman et al., 1977) is probably the most fa- reduces the number of nodes from O(n) to mous. It is simple and effective in practice, but it can only exactly n, be used on Euclidean spaces. We must turn to other data 2. An additional invariant that makes queries structures when given an arbitrary metric space. The sim- faster in practice, plest and oldest of these structures is the ball tree (Omo- 3. Algorithms for constructing and querying hundro, 1989). Although attractive for its simplicity, it pro- the tree in parallel on multiprocessor sys- vides only the trivial runtime guarantee that queries will tems, and take time O(n). Subsequent research focused on provid- 4. A more cache efficient memory layout. ing stronger guarantees, producing more complicated data structures like the metric skip list (Karger & Ruhl, 2002) On standard benchmark datasets, we reduce the and the navigating net (Krauthgamer & Lee, 2004).
    [Show full text]
  • Canopy – Fast Sampling Using Cover Trees
    Canopy { Fast Sampling using Cover Trees Canopy { Fast Sampling using Cover Trees Manzil Zaheer Abstract Background. The amount of unorganized and unlabeled data, specifically images, is growing. To extract knowledge or insights from the data, exploratory data analysis tools such as searching and clustering have become the norm. The ability to search and cluster the data semantically facilitates tasks like user browsing, studying medical images, content recommendation, and com- putational advertisements among others. However, large image collections are plagued by lack of a good representation in a relevant metric space, thus making organization and navigation difficult. Moreover, the existing methods do not scale for basic operations like searching and clustering. Aim. To develop a fast and scalable search and clustering framework to enable discovery and exploration of extremely large scale data, in particular images. Data. To demonstrate the proposed method, we use two datasets of different nature. The first dataset comprises of satellite images of seven urban areas from around the world of size 200GB. The second dataset comprises of 100M images of size 40TB from Flickr, consisting of variety of photographs from day to day life without any labels or sometimes with noisy tags (∼1% of data). Methods. The excellent ability of deep networks in learning general representation is exploited to extract features from the images. An efficient hierarchical space partitioning data structure, called cover trees, is utilized so as to reduce search and reuse computations. Results. Using the visual search-by-example on satellite images runs in real-time on a single machine and the task of clustering into 64k classes converges in a single day using just 16 com- modity machines.
    [Show full text]
  • Cover Trees for Nearest Neighbor
    Cover Trees for Nearest Neighbor Alina Beygelzimer [email protected] IBM Thomas J. Watson Research Center, Hawthorne, NY 10532 Sham Kakade [email protected] TTI-Chicago, 1427 E 60th Street, Chicago, IL 60637 John Langford [email protected] TTI-Chicago, 1427 E 60th Street, Chicago, IL 60637 Abstract The basic nearest neighbor problem is as follows: We present a tree data structure for fast Given a set S of n points in some metric space (X, d), nearest neighbor operations in general n- the problem is to preprocess S so that given a query point p X, one can efficiently find a point q S point metric spaces (where the data set con- ∈ ∈ sists of n points). The data structure re- which minimizes d(p, q). quires O(n) space regardless of the met- ric’s structure yet maintains all performance Context. For general metrics, finding (or even ap- properties of a navigating net [KL04a]. If proximating) the nearest neighbor of a point requires the point set has a bounded expansion con- Ω(n) time. The classical example is a uniform met- stant c, which is a measure of the intrinsic ric where every pair of points is near the same dis- dimensionality (as defined in [KR02]), the tance, so there is no structure to take advantage of. cover tree data structure can be constructed However, the metrics of practical interest typically do in O c6n log n time. Furthermore, nearest have some structure which can be exploited to yield neighbor queries require time only logarith- significant computational speedups.
    [Show full text]
  • SURVEY on VARIOUS-WIDTH CLUSTERING for EFFICIENT K-NEAREST NEIGHBOR SEARCH
    Vol-3 Issue-2 2017 IJARIIE-ISSN(O)-2395-4396 SURVEY ON VARIOUS-WIDTH CLUSTERING FOR EFFICIENT k-NEAREST NEIGHBOR SEARCH Sneha N. Aware1, Prof. Dr. Varsha H. Patil2 1 PG Student, Department of Computer Engineering,Matoshri College of Engineering & Research Centre,Nashik ,Maharashtra, India. 2 Head of Department of Computer Engineering, Matoshri College of Engineering & Research Centre, Nashik, Maharashtra, India. ABSTRACT Over recent decades, database sizes have grown large. Due to the large sizes it create new challenges, because many machine learning algorithms are not able to process such a large volume of information. The k-nearest neighbor (k- NN) is widely used in many machines learning problem. k-NN approach incurs a large computational cost. In this research a k-NN approach based on various-width clustering is presented. The k-NN search technique is based on VWC is used to efficiently find k-NNs for a query object from a given data set and create clusters with various widths. This reduces clustering time in addition to balancing the number of produced clusters and their respective sizes. Keyword: - clustering,k-Nearest Neighbor, various-width clustering, high dimensional, etc. 1. INTRODUCTION:- The k-nearest neighbor approach (k-NN) has been extensively used as a powerful non-parametric technique in many scientific and engineering applications. To find the closest result a k-NN technique is widely used in much application area. This approach incurs a large computational area. General k-NN problem states that, suppose we have n number of objects and a Q query object, then distance from the query object to all training object need to be calculated in order to find k closest objects to Q.
    [Show full text]
  • An Empirical Comparison of Exact Nearest Neighbour Algorithms
    An Empirical Comparison of Exact Nearest Neighbour Algorithms Ashraf M. Kibriya and Eibe Frank Department of Computer Science University of Waikato Hamilton, New Zealand {amk14, eibe}@cs.waikato.ac.nz Abstract. Nearest neighbour search (NNS) is an old problem that is of practical importance in a number of fields. It involves finding, for a given point q, called the query, one or more points from a given set of points that are nearest to the query q. Since the initial inception of the problem a great number of algorithms and techniques have been proposed for its solution. However, it remains the case that many of the proposed algorithms have not been compared against each other on a wide variety of datasets. This research attempts to fill this gap to some extent by presenting a detailed empirical comparison of three prominent data structures for exact NNS: KD-Trees, Metric Trees, and Cover Trees. Our results suggest that there is generally little gain in using Metric Trees or Cover Trees instead of KD-Trees for the standard NNS problem. 1 Introduction The problem of nearest neighbour search (NNS) is old [1] and comes up in many fields of practical interest. It has been extensively studied and a large number of data structures, algorithms and techniques have been proposed for its solution. Although nearest neighbour search is the most dominant term used, it is also known as the best-match, closest-match, closest-point and the post office problem. The term similarity search is also often used in the information retrieval field and the database community.
    [Show full text]
  • COVER TREES for NEAREST NEIGHBOR Nearest Neighbor
    COVER TREES FOR NEAREST NEIGHBOR ALINA BEYGELZIMER, SHAM KAKADE, AND JOHN LANGFORD ABSTRACT. We present a tree data structure for fast nearest neighbor operations in general Ò-point metric ´Òµ spaces. The data structure requires Ç space regardless of the metric’s structure. If the point set has an ¾ expansion constant in the sense of Karger and Ruhl [KR02], the data structure can be constructed ¡ ¡ ½¾ Ò ÐÓ Ò Ç ÐÓ Ò in Ç time. Nearest neighbor queries obeying the expansion bound require time. ½ ´Òµ Ç ´ Òµ In addition, the nearest neighbor of Ç points can be queried in time. We experimentally test the algorithm showing speedups over the brute force search varying between 1 and 2000 on natural machine learning datasets. 1. INTRODUCTION Nearest neighbor search is a fundamental problem with a number of applications in peer-to-peer networks, lossy data compression, vision, dimensionality reduction, computational biology, machine learning, and Ò physical simulation. The standard setting is as follows: Given a set Ë of points in some metric space µ Ë Ô ¾ ´ , the problem is to preprocess so that given a query point , one can efficiently find a point ¾ Ë ´Ô Õ µ Õ which minimizes . For general metrics, finding (or even approximating) the nearest neighbor Òµ of a point requires ª´ time. However, the metrics and datasets of practical interest typically have some structure which may be exploited to yield significant computational speedups. Most research has focused on the special case when the metric is Euclidean. Although many strong theoretical guarantees have been made here, this case is quite limiting in a variety of settings when the data do not naturally lie in a Euclidean space or when the data is embedded in a high dimensional space, but has a low dimensional intrinsic structure [TSL00, RS00] (as is common in many machine learning problems).
    [Show full text]
  • All-Near-Neighbor Search Among High-Dimensional Data Via Hierarchical Sparse Graph Code Filtering
    All-Near-Neighbor Search Among High-Dimensional Data via Hierarchical Sparse Graph Code Filtering by Alexandros-Stavros Iliopoulos Department of Computer Science Duke University Date: Approved: Xiaobai Sun, Advisor Pankaj K. Agarwal Nikos Pitsianis Carlo Tomasi Dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Computer Science in the Graduate School of Duke University 2020 Abstract All-Near-Neighbor Search Among High-Dimensional Data via Hierarchical Sparse Graph Code Filtering by Alexandros-Stavros Iliopoulos Department of Computer Science Duke University Date: Approved: Xiaobai Sun, Advisor Pankaj K. Agarwal Nikos Pitsianis Carlo Tomasi An abstract of a dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Computer Science in the Graduate School of Duke University 2020 Copyright c 2020 by Alexandros-Stavros Iliopoulos All rights reserved except○ the rights granted by the Creative Commons Attribution-Noncommercial Licence Abstract This thesis addresses the problem of all-to-all near-neighbor (all-NN) search among a large dataset of discrete points in a high-dimensional feature space endowed with a distance metric. Near-neighbor search is fundamental to numerous computational tasks arising in modern data analysis, modeling, and knowledge discovery. The ac- curacy and efficiency of near-neighbor search depends on the underlying structure of the dataset. In emerging data-driven analysis applications, the dataset is not necessarily stationary, the structure is not necessarily built once for all queries, and search is not necessarily limited to a few query points. To facilitate accurate and efficient near-neighbor search in a stationary or dynamic dataset, we makeasys- tematic investigation of all-NN search at multiple resolution levels, with attention to several standing or previously overlooked issues associated with high-dimensional feature spaces.
    [Show full text]
  • Improving Dual-Tree Algorithms
    IMPROVING DUAL-TREE ALGORITHMS A Dissertation Presented to The Academic Faculty By Ryan R. Curtin In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in Electrical and Computer Engineering School of Electrical and Computer Engineering Georgia Institute of Technology December 2015 Copyright © 2015 by Ryan R. Curtin IMPROVING DUAL-TREE ALGORITHMS Approved by: Dr. David V. Anderson, Committee Chair Dr. Mark A. Clements Associate Professor, School of Electrical and Associate Professor, School of Electrical and Computer Engineering Computer Engineering Georgia Institute of Technology Georgia Institute of Technology Dr. Charles L. Isbell, Jr., Advisor Dr. Polo Chau Professor, School of Interactive Computing Assistant Professor, School of Computational Georgia Institute of Technology Science and Engineering Georgia Institute of Technology Dr. Richard W. Vuduc Associate Professor, School of Computational Science and Engineering Georgia Institute of Technology Date Approved: August 18, 2015 This document is dedicated solely to my cats, who do not and will not ever have the capacity to understand even the title of this manuscript, and who, thanks to domestication, are actually entirely incapable of leading any sort of autonomous lifestyle and thus are mortally dependent on my completion of simple maintenance tasks. TABLE OF CONTENTS LIST OF TABLES ................................... vii LIST OF FIGURES .................................. ix CHAPTER 1 THE POINT ............................. 1 CHAPTER 2 INTRODUCTION .......................... 3 2.1 An abridged history of statistical computing . .3 2.2 A less abridged history of the development of trees . .4 2.3 The explosion of single-tree algorithms . 10 2.4 A smorgasboard of trees . 12 2.5 The fast multipole method and query amortization .
    [Show full text]
  • Non-Rigid Surface Registration Using Cover Tree Based Clustering and Nearest Neighbor Search
    Non-rigid Surface Registration using Cover Tree based Clustering and Nearest Neighbor Search Manal H. Alassaf 1, 2, Yeny Yim3 and James K. Hahn1 1Department of Computer Science, George Washington University, Washington DC, U.S.A. 2Department of computer science, Taif University, Taif, Saudi Arabia 3Samsung Electronics, Suwon, Geonggi-Do, Republic of Korea Keywords: Non-rigid Registration, Iterative Closest Point Algorithm, ICP, Cover Tree, Clustering, Nearest Neighbor Search. Abstract: We propose a novel non-rigid registration method that computes the correspondences of two deformable surfaces using the cover tree. The aim is to find the correct correspondences without landmark selection and to reduce the computational complexity. The source surface S is initially aligned to the target surface T to generate a cover tree from the densely distributed surface points. The cover tree is constructed by taking into account the positions and normal vectors of the points and used for hierarchical clustering and nearest neighbor search. The cover tree based clustering divides the two surfaces into several clusters based on the geometric features, and each cluster on the source surface is transformed to its corresponding cluster on the target. The nearest neighbor search from the cover tree reduces the search space for correspondence computation, and the source surface is deformed to the target by optimizing the point pairs. The correct correspondence of a given source point is determined by choosing one target point with the best correspondence measure from the k nearest neighbors. The proposed energy function with Jacobian penalty allows deforming the surface accurately and with less deformation folding. 1 INTRODUCTION it is quite efficient especially for low dimensional data.
    [Show full text]
  • Introduction to Cover Tree
    Introduction to Cover Tree Javen Qinfeng Shi SML of NICTA RSISE of ANU 2006 Oct. 15 Outline Goal Related works Cover tree definition Optimized Implement in C++ Complexity Further work Outline Goal Related works Cover tree definition Efficiently Implement in C++ Advantages and disadvantages Further work Goal Speed up Nearest-neighbour search[John Langford etc. 2006] ZPreprocess a dataset S of n points in some metric space X so that given a query point p ∈ X , a point q ∈ S which minimises d(p,q) can be efficiently found Also Kmeans Outline Goal Related works Cover tree definition Example Optimized Implement in C++ Advantages and disadvantages Further work Relevant works Kd tree [Friedman, Bentley & Finkel 1977] Ball tree family: ZBall tree [Omohundro,1991] ZMetric tree [Uhlmann, 1991] ZAnchors metric tree [Moore 2000] Z(Improved) Anchors metric tree [Charles2003] Navigating net [R.Krauthgamer, J.Lee 2004] ? Lots of discussions and proofs on its computational bound Kd tree Univariate axis-aligned splits Split on widest dimension O(N log N) to build, O(N) space Kd tree Kd tree y x (axis) Ball tree/metric tree A set of points in R2 [Uhlmann 1991], [Omohundro 1991] A ball-tree: level 1 [Uhlmann 1991], [Omohundro 1991] A ball-tree: level 2 A ball-tree: level 3 A ball-tree: level 4 A ball-tree: level 5 Outline Goal Related works Cover tree definition ZBasic definition ZHow to build a cover tree ZQuery a point/points Optimized Implement in C Advantages and disadvantages Further work Basic definition A cover tree
    [Show full text]