KR101871871B1 - Data clustering apparatus using g-optics - Google Patents

Abstract

The present invention proposes a data clustering apparatus using G-OPTICS. The data clustering apparatus according to the present invention includes a storage unit for storing a G-OPTICS algorithm and a control unit for clustering data using the G-OPTICS algorithm. According to the present invention, data clustering is performed using the G-OPTICS algorithm using the GPU having superior performance compared to the existing clustering algorithm It is possible.

Description

DATA CLUSTERING APPARATUS USING G-OPTICS BACKGROUND OF THE INVENTION 1. Field of the Invention [0001]

The present invention relates to a data clustering apparatus using G-OPTICS, and more particularly, to a data clustering technique using a G-OPTICS algorithm that improves the performance of an OPTICS using a GPU.

Clustering is the process of creating a group of similar objects in a dataset, or clusters, and is one of the most important data mining problems that have been studied for a long time. Clustering algorithms are very diverse, among which density-based algorithms are widely used because they form a cluster of arbitrary types, filter out noises easily, and do not have to pre-define the number of clusters. DBSCAN is a representative density-based algorithm, and various studies have been conducted to improve performance to handle large-capacity datasets more efficiently.

)의 높은 복잡도로 가지며, 여기에서 N은 데이터셋의 크기이다. DB-SCAN, like many other data mining algorithms, has a weakness that clustering results are sensitive to parameter values. OPTICS is an algorithm for solving this weakness, and it calculates the ordering of equivalent objects and the clustering result corresponding to the smaller arbitrary 'given by the parameter. Like DBSCAN, OPTICS must calculate the distance to all other objects in the dataset, so O (

), Where N is the size of the dataset.

)에 대하여 F-OPTICS에 의하여 얻어진 클러스터링 결과는 OPTICS 또는 DBSCAN과 다를 수 있다. F-OPTICS에서는 파라미터의 개수를 줄이기 위하여 항상 = 1로 설정하였다. 하지만, 이 증가함에 따라 -neighbor를 구하기 위한 부담도 증가하므로, 과대한 을 설정하는 것은 알고리즘의 실행 부담을 늘리는 문제점이 있다.As shown in FIG. 1, OPTICS is an object q (n) having a minimum reachability distance from Seeds which is a union of neighbor objects N (p) of previous objects p, and F-OPTICS And append them to ordering O. The definition of the reachability distance of object q for object p is given by the following formulas (1) and (2). Here, MinPts-distance (p) is the distance to MinPts-nearest object of p, and D (p; q) is the distance between two objects p, q. The most time-consuming operation in OPTICS is to find -neighbor N (p) for all objects p in the dataset, and for this, the distance to all other objects must be calculated. In order to generate the core distance of the object p, the smallest MinPts of the distance values should be (partial) aligned. Existing algorithms for improving the performance of OPTICS have maximized this part. DeLi-Clu performed a MinPTS-nearest neighbor join using a multi-dimensional index structure such as R-tree. Experiments show that the performance is up to 20 times higher than OPTICS. The disadvantage of DeLi-Clu is that the dimensionality curse problem of multidimensional indices reduces the performance dramatically as the dimension of the data increases. F-OPTICS obtains a sampled neighbor instead of the actual -neighbor for object p as follows. A one-dimensional sequence S is obtained by projecting all objects in the dataset onto an arbitrary line L. Select any close objects within the MinPts count from object p in S. This operation is performed on several arbitrary lines, and the set of obtained close objects is called a sampled neighbor. By reducing the burden of acquiring -neighbor-spaces, F-OPTICS improved performance up to 38 times compared to DeLi-Clu. The object ordering obtained by F-OPTICS is similar to OPTICS but does not match. That is,

), The clustering result obtained by F-OPTICS may be different from OPTICS or DBSCAN. In F-OPTICS, it is always set to 1 in order to reduce the number of parameters. However, since the burden to obtain the -neighborness increases as the number increases, setting an excessive number increases the burden of the algorithm execution.

Korean Patent Laid-Open Publication No. 10-2015-0065433 (algorithm storage device and clustering device including the same)

SUMMARY OF THE INVENTION It is an object of the present invention to provide a data clustering apparatus using a G-OPTICS algorithm that improves the performance of OPTICS using a GPU.

The data clustering apparatus using G-OPTICS according to an aspect of the present invention includes a storage unit for storing a G-OPTICS algorithm and a control unit for clustering data using the G-OPTICS algorithm.

The apparatus may further include an output unit outputting a result of clustering data using the G-OPTICS algorithm of the control unit.

Preferably, the G-OPTICS algorithm may be an implementation of the OPTICS algorithm using a GPU.

According to the present invention, data clustering is performed using the G-OPTICS algorithm using the GPU having superior performance compared to the existing clustering algorithm It is possible.

FIG. 1 is a diagram illustrating an F-OPTICS algorithm, which is a conventional data clustering algorithm.
FIG. 2 is a block diagram of a data clustering apparatus using G-OPTICS according to an embodiment of the present invention. Referring to FIG.
3 is a diagram illustrating an example of a structure of a GPU used by a data clustering apparatus using G-OPTICS according to an embodiment of the present invention.
4 is a diagram illustrating a G-OPTICS algorithm used by a data clustering apparatus using G-OPTICS according to an embodiment of the present invention.
5 is a diagram illustrating a data structure of a G-OPTICS according to an embodiment of the present invention.
6 is a diagram illustrating a candidate object of the G-OPTIC algorithm according to an embodiment of the present invention.
FIG. 7 is a graph showing a comparison result between F-OPTICS and G-OPTICS performance.

In order to fully understand the present invention, operational advantages of the present invention, and objects achieved by the practice of the present invention, reference should be made to the accompanying drawings and the accompanying drawings which illustrate preferred embodiments of the present invention. It is also to be understood that the specific structure or functional description presented in the embodiments of the present invention is illustrated for the purpose of describing an embodiment according to the concept of the present invention only and embodiments according to the concept of the present invention may be embodied in various forms . Likewise, it should be understood that the present invention should not be construed as limited to the embodiments described herein, but includes all modifications, equivalents, and alternatives falling within the spirit and scope of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Like reference symbols in the drawings denote like elements.

FIG. 2 is a block diagram of a data clustering apparatus using G-OPTICS according to an embodiment of the present invention. The apparatus for clustering data using G-OPTICS according to an embodiment of the present invention includes a storage unit 10, (20), and may further include an output unit (30).

The storage unit 10 stores the G-OPTICS algorithm to be described below, and can be implemented in a DataBase format. The control unit 20 performs a function of clustering data using the G-OPTICS algorithm, and can be implemented as a CPU, an AP, or the like as a central processing unit. The output unit 30 may output the data clustering result of the controller 20 and may be implemented in the form of a display device such as a monitor.

FIG. 3 is a diagram illustrating an exemplary structure of a GPU used by a data clustering apparatus using G-OPTICS according to an embodiment of the present invention. FIG. OPTICS algorithm used by a data clustering apparatus using G-OPTICS according to an embodiment of the present invention. FIG. 5 is a diagram illustrating a data structure of a G-OPTICS according to an embodiment of the present invention. 6 illustrates a candidate object of the G-OPTIC algorithm according to an exemplary embodiment of the present invention. Referring to FIG. 6, a G-OPTICS algorithm according to an embodiment of the present invention and an apparatus for clustering data using the algorithm will be described.

The G-OPTICS proposed according to one aspect of the present invention improves the performance of the OPTICS using the GPU. FIG. 3 is a simplified diagram of the GPU's threads and memory structure. The GPU executes several millions of threads simultaneously, and all threads execute the same function (called a kernel function) for different data. The GPU thread is divided into blocks containing the same number of threads. Threads in one block share data using shared memory, and data sharing between threads in different blocks should use device memory. Shared memory is much smaller (> 1 TB / s) while it is smaller (eg, 48 KB). To increase the performance of the algorithm using the GPU, it is necessary to maximize the number of independent threads and maximize shared memory.

G-OPTICS inherits the basic structure of OPTICS for algorithmic accuracy and can improve performance by parallelizing operations that do not have dependency between OPTICS and require a lot of execution time. Algorithms 3 and 4 in FIG. 4 are summarized for the G-OPTICS algorithm, and the parts that are performed in parallel in Algorithm 3 are the Algorithm lines (1) - (5), (14), and (16). In line (1) - (5) of Algorithm 3 of G-OPTICS, the method to obtain -neighbor, N (pi) for all objects pi in data set D is as follows.

In OPTICS, the task requiring the greatest computation time is to obtain the -neighbor (lines (7) and (14) of Algorithm 1 of the conventional OPTICS shown in FIG. 1) I can go crazy. G-OPTICS first constructs a data structure Lj (0 j < d 'd) as shown in FIG. 5, where d is a data dimension. Lj creates an array of pi (j; pi: id) pairs of IDs pi: id of pi with j-dimensional coordinate values pi from all objects pi (0i <N) j. &lt; / RTI &gt; Where N is the dataset size (N = jDj). To find the -neighborhood of an arbitrary object pi, first find all the candidate objects Cj within Lj within the distance difference from pi; j for all j. In FIG. 5, all objects from object pu to pv are applicable. Next, the actual distance between all candidate objects in Cj (pi) = \ 0j <d'Cj of all Cjs and pi is found, and only objects within the distance are selected. In FIG. 6, a region in which candidate objects are present is represented as a gray region, where d '= 2. Since this area is very small compared to the entire dataset area, the number of distance calculations between objects performed by G-OPTICS is greatly reduced and performance can be greatly improved. G-OPTICS finds the -neighbor for one object pi in each thread block and performs the distance calculation with all the objects in the candidate set C (pi) in parallel at the same time. In addition, G-OPTICS runs hundreds of threads simultaneously, which can greatly improve performance. d 'value is d

 Can be set to any value below.

n)의 복잡도를 갖는다. 여기에서 n은 C(pi) 내의 후보 객체들의 개수이다. 객체 p의 core distance를 구한 후에는 상기 공식 (2)를 이용하여 N(p) 내의 모든 객체들 q에 대하여 p에 대한 reachability distance를 산출한다. 이 값은 도 4의 Algorithms 4의 Update 함수 내의 라인 (2)에서 사용된다. 객체 p에 대한 core distance 및 객체 q(2N(p))에 대한 reachability distance의 계산과 p의 상태를 unprocessed로 설정하는 것은 N(p)를 구하는 동일한 쓰레드 블록에서 수행하므로 추가적인 부담이 적고 효율적으로 수행 가능하다(도 4의 Algorithm 3의 라인 (2)-(4)). 이러한 작업들을 동일한 쓰레드 블록에서 수행하는 또다른 장점은 shared memory를 활용함으로써 성능을 더욱 개선할 수 있다는 점이다. In order to obtain the core distance by the above formula (1) for the object p, the distance of the MinPts-th nearest object should be obtained by aligning the distances D (p; q) with all the objects q in N (p) in ascending order. Most efficient sorting algorithms, such as quick sort, heap sort, and merge sort, which are widely known, are sequential algorithms on a single thread and have the disadvantage that they are difficult to parallelize. G-OPTICS uses the bitonic sort algorithm [6] to perform alignment on the GPU. In a parallel environment, algorithms such as quick sort, heap sort, and merge sort have O (n) complexity, while bitonic sort algorithms have much lower O

n). Where n is the number of candidate objects in C (pi). After obtaining the core distance of object p, we use Equation (2) to calculate the reachability distance for p for all objects q in N (p). This value is used in line 2 in the Update function of Algorithms 4 of FIG. The computation of the reachability distance to the core distance and the object q (2N (p)) for object p and the state of p to unprocessed are performed in the same thread block for obtaining N (p) (Lines (2) - (4) of Algorithm 3 in Fig. 4). Another advantage of performing these tasks on the same thread block is that performance can be further improved by utilizing shared memory.

The method of performing the Update function in the line 14 of the Algorithm 3 of the G-OPTICS in FIG. 4 is as follows. The Update function of OPTICS inserts or relocates all objects q in N (p) to the appropriate positions in Seeds so that the reachability distance values are sorted in ascending order. However, this task can have the complexity of O (N2) in the worst case because it needs to compare reachability distances with all objects in Seeds for each object q. In G-OPTICS, each object q is processed in different concurrent threads and has the complexity of O (1) because it inserts the object in Seeds irrespective of reachability distance order. The search for objects with the minimum reachability distance in Seeds is performed on line (16) in Algorithm 3 of the Diagram. To accomplish this, we can find objects with the minimum reachability distance in each thread by distributing the objects in Seeds to multiple concurrent threads, and then find objects with the minimum reachability distance among all the threads from all threads.

FIG. 7 is a graph showing a comparison result between F-OPTICS and G-OPTICS performance. Hereinafter, the description will be made and the d '= 4 is always set in the experiments described below. Since F-OPTICS improves performance up to 520 times and 38 times, respectively, compared to OPTICS and DeLi-Clu, G-OPTICS compares with F-OPTICS. The hardware platform used in this experiment is a workstation equipped with Intel Core i7-3960X CPU, 32GB main memory, 256GB SSD, and Nvidia GeForce GTX 1080 GPU, and implemented using CUDA Toolkit 7.5 on Microsoft Windows 7 64-bit Edition. F-OPTICS utilized open source. It is assumed that the data used in this experiment is synthetic data, and that the coordinate values of all objects are included in the [0.0, 1.0] area. The default parameter values were set to N = 64K, = 0.1, MinPts = 4, d = 8.

The graphs (a) to (d) shown in FIG. 7 show experimental results of changing data set size (N), threshold (), MinPts, and data dimension (d) parameters. Actual execution time in graph is bar chart, G-OPTICS performance improvement rate is line chart, and actual execution time is log scale. The default value is set except for the parameter values that are changed in each experiment. As shown in FIG. 6, the performance of G-OPTICS has been greatly improved up to 90.2 times compared with F-OPTICS. In FIG. 7 (a), the performance improvement ratio decreases as the size of the data set increases. This is because the burden of performing-alignment on the object-p is increased rapidly in G-OPTICS. As shown in FIG. 7 (b), the size of the neighbors of the neighbors C (p) increases in the G-OPTICS, thereby increasing the execution time and decreasing the performance improvement ratio. In FIG. 7 (c), the execution time of both G-OPTICS and F-OPTICS hardly changed according to the MinPts value, and the performance improvement ratio remained almost the same. In FIG. 7 (d), as the data dimension increases, the execution time of the F-OPTICS increases and the performance improvement ratio of the G-OPTICS increases.

The present invention has been described in detail with reference to preferred embodiments. It will be apparent to those skilled in the art that the present invention is not limited to the embodiments described above and that various modifications and changes may be made by one of ordinary skill in the art without departing from the scope of the present invention, It is to be understood that the technical idea of the present invention extends to the extent possible.

Claims (3)

A storage unit for storing a G-OPTICS algorithm that implements an OPTICS algorithm using a GPU; And
And a controller for clustering data using the G-OPTICS algorithm,
The GPU runs millions of threads simultaneously, all threads execute the same function on different data,
GPU threads are divided into blocks containing the same number of threads so that threads in one block share data using shared memory and data sharing between threads in different blocks uses device memory. Data clustering device using OPTICS.
The method according to claim 1,
Further comprising an output unit outputting a result of clustering data using the G-OPTICS algorithm of the control unit. delete

Images