Jinlong Huang

A non-parameter outlier detection algorithm based on Natural Neighbor

Outlier detection is an important task in data mining with numerous applications, including credit card fraud detection, video surveillance, etc. Although many Outlier detection algorithm have been proposed. However, for most of these algorithms faced a serious problem that it is very difficult to select an appropriate parameter when they run on a dataset. In this paper we use the method of Natural Neighbor to adaptively obtain the parameter, named Natural Value. We also propose a novel notion that Natural Outlier Factor (NOF) to measure the outliers and provide the algorithm based on Natural Neighbor (NaN) that does not require any parameters to compute the NOF of the objects in the database. The formal analysis and experiments show that this method can achieve good performance in outlier detection.

Natural neighbor-based clustering algorithm with local representatives

Clustering by identifying cluster centers is important for detecting patterns in a data set. However, many center-based clustering algorithms cannot process data sets containing non-spherical clusters. In this paper, we propose a novel clustering algorithm called NaNLORE based on natural neighbor and local representatives. Natural neighbor is a new neighbor concept and introduced to compute local density and find local representatives which are points with local maximum density. We first find local representatives and then select cluster centers from the local representatives. The density-adaptive distance is introduced to measure the distance between local representatives, which helps to solve the problem of clustering data sets with complex manifold structure. Cluster centers are characterized by higher density than their neighbors and a relatively large density-adaptive distance from any local representatives with higher density. In experiments, we compare the proposed algorithm NaNLORE with existing algorithms on synthetic and real data sets. Results show that NaNLORE performs better than existing algorithm, especially on clustering non-spherical data and manifold data.

Adaptive edited natural neighbor algorithm

Reduction techniques can reduce prohibitive computational costs and the storage requirements for classifying patterns while maintaining classification accuracy. The edited nearest neighbor rule is one of the most popular reduction technique, which removes noisy patterns that are not correctly classified by their k-nearest neighbors. However, selection of neighborhood parameters is an unsolved problem for the traditional neighborhood construction algorithms such as k-nearest neighbor and e-neighborhood. To solve the problem, we present a novel editing algorithm called adaptive Edited Natural Neighbor algorithm (ENaN). ENaN aims to eliminate the noisy patterns based on the concept of natural neighbor which are obtained adaptively by the search algorithm of natural neighbor. The main advantages are that ENaN does not need any parameters and can degrade the effect of noisy patterns. The adaptive ENaN algorithm can be easily applied into other reduction algorithms as a noisy filter. Experiments show that the proposed approach effectively removes the noisy patterns while keeping more reasonable class boundaries and improves the performance of two condensation methods in terms of both accuracy and reduction rate greatly.

QCC: a novel clustering algorithm based on Quasi-Cluster Centers

Cluster analysis aims at classifying objects into categories on the basis of their similarity and has been widely used in many areas such as pattern recognition and image processing. In this paper, we propose a novel clustering algorithm called QCC mainly based on the following ideas: the density of a cluster center is the highest in its K nearest neighborhood or reverse K nearest neighborhood, and clusters are divided by sparse regions. Besides, we define a novel concept of similarity between clusters to solve the complex-manifold problem. In experiments, we compare the proposed algorithm QCC with DBSCAN, DP and DAAP algorithms on synthetic and real-world datasets. Results show that QCC performs the best, and its superiority on clustering non-spherical data and complex-manifold data is especially large.

Weighted natural neighborhood graph: an adaptive structure for clustering and outlier detection with no neighborhood parameter

This paper aims at dealing with the practical shortages of nearest neighbor based data mining techniques, especially, clustering and outlier detection. In particular, when there are data sets with arbitrary shaped clusters and varying density, it is difficult to determine the proper parameters without a priori knowledge. To address this issue, we define a novel conception called natural neighbor, which can better reflect the relationship between the elements in a data set than k-nearest neighbor does, and we present a graph called weighted natural neighborhood graph for clustering and outlier detection. Furthermore, the whole process needs no parameter to deal with different data sets. Simulations on both synthetic data and real world data show the effectiveness of our proposed method.

A novel outlier cluster detection algorithm without top-n parameter

Outlier detection is an important task in data mining with numerous applications, including credit card fraud detection, video surveillance, etc. Outlier detection has been widely focused and studied in recent years. The concept about outlier factor of object is extended to the case of cluster. Although many outlier detection algorithms have been proposed, most of them face the top-n problem, i.e., it is difficult to know how many points in a database are outliers. In this paper we propose a novel outlier cluster detection algorithm called ROCF based on the concept of mutual neighbor graph and on the idea that the size of outlier clusters is usually much smaller than the normal clusters. ROCF can automatically figure out the outlier rate of a database and effectively detect the outliers and outlier clusters without top-n parameter. The formal analysis and experiments show that this method can achieve good performance in outlier detection.

Natural Neighbor Reduction Algorithm for Instance-based Learning

Instance reduction is aimed at reducing prohibitive computational costs and the storage space for instance-based learning. The most frequently used methods include the condensation and edition approaches. Condensation method removes the patterns far from the decision boundary and do not contribute to better classification accuracy, while edition method removes noisy patterns to improve the classification accuracy. In this paper, a new hybrid algorithm called instance reduction algorithm based on natural neighbor and nearest enemy is presented. At first, an edition algorithm is proposed to filter noisy patterns and smooth the class boundaries by using natural neighbor. The main advantage of the algorithm is that it does not require any user-defined parameters. Then, using a new condensation method based on nearest enemy to reduce instances far from decision line. Through this algorithm, interior instances are discarded. Experiments show that the hybrid approach effectively reduces the number of instances while achieves higher classification accuracy along with competitive algorithms.

Natural neighbor-based clustering algorithm with density peeks

Clustering analysis has been widely used in many areas such as astronomy, bioinformatics, and pattern recognition. In 2014, Rodriguez proposed an algorithm based on the idea that cluster centers are characterized by a higher density than their neighbors and by a relatively large distance from points with higher density. But the density relies on cutoff distance, which might be affected by large statistical error, and the algorithm does not suit the clustering problem of multi-scale data. In this paper, a new neighbor concept Natural Neighbor is proposed. Natural neighbor-based density, is simple and well reflects the data distribution without any parameters. Then, we extend each cluster from its center by searching natural neighbors of points in this cluster, and we define extension rules to determine the cluster boundary. The experiment results show our algorithm is more effective on multi-scale data.

An efficient reduction algorithm based on natural neighbor and nearest enemy

Prototype reduction is aimed at reducing prohibitive computational costs and the storage space for pattern recognition. The most frequently used methods include the condensating and editing approaches. Condensating method removes the patterns far from the decision boundary and do not contribute to better classification accuracy, while editing method removes noise patterns to improve the classification accuracy. In this paper, a new hybrid algorithm called prototype reduction algorithm based on natural neighbor and nearest enemy is presented. At first, an editing algorithm is proposed to filter noisy patterns and smooth the class boundaries by using the concept of natural neighbor. The main advantage of the editing algorithm is that it does not require any user-defined parameters. Then, using a new condensing method based on nearest enemy to reduce prototypes far from decision line. Through this algorithm, interior prototypes are discarded. Experiments show that the hybrid approach effectively reduces the number of prototypes while achieves higher classification performance along with competitive prototype algorithms.

A Hierarchical Clustering Algorithm Based on Saturated Neighbor Graph

This paper proposes a hierarchical clustering algorithm based on Saturated Neighbor Graph -- hi-CLUBS and a new concept, natural nearest neighbor, which adopts a parameter-less algorithm of searching the natural neighbors for each point in a dataset. In the work, the Saturated Neighbor Graph is constructed by the natural nearest neighbor firstly. Then modularity is introduced into graph partitioning algorithm, with which the generated graph is partitioned into small sub-clusters without any parameters. Finally, these initial sub-clusters are repeatedly merged with another cluster according to similarity measurement based on connectivity and closeness, until the desired cluster number is reached. The results show that hi-CLUBS produces a set of final clusters achieves better quality than the traditional clustering algorithms.