Abstract

     A Central Bank Digital Currency (CBDC) launched by the Bank of England could enable businesses to directly make electronic payments. It can be argued that digital payment is helpful in supply chain management applications. However, the adoption of CBDC in the supply chain could bring new turbulence since the CBDC value may fluctuate. Therefore, this paper intends to optimize the production plan of manufacturing supply chain based on a volatility clustering model by reducing CBDC value uncertainty. We apply both GARCH model and machine learning model to depict the CBDC volatility clustering. Empirically, we employed Baltic Dry Index, Bitcoin and exchange rate as main variables with sample period from 2015 to 2021 to evaluate the performance of the two models. On this basis, we reveal that our machine learning model overwhelmingly outperforms the GARCH model. Consequently, our result implies that manufacturing companies’ performance can be strengthened through CBDC uncertainty reduction.

Introduction

     Cryptocurrencies, represented by Bitcoin, have exhibited a significant impact on the current financial system (Easley et al., 2019, Zhang and Li, 2020, Huynh et al., 2021, Jin et al., 2021, Prat and Walter, 2021), especially after the outbreak of COVID-19 (Guo et al., 2021, Diniz-Maganini et al., 2021) and Fintech development (Le et al., 2021, Chaudhry et al., 2022). As a result, to confront and replace such an emerging private digital currency system, the Bank of England announced the launch of a Central Bank Digital Currency (CBDC), which served as a new form of digital currency issued by the Bank of England and can be used by households and businesses (BenDhaou and Rohman, 2018, Vessio, 2021).

     Since CBDC is still under development, copious of scholars advocate the adoption of blockchain techniques into CBDC systems, which can yield a number of benefits, such as increased payment safety and information transparency (Sun et al., 2017, Sethaput and Innet, 2021). Moreover, several countries have already signified the adoption of blockchain in their CBDC payment systems (Xu and Zou, 2021). On the other hand, bitcoin is an exceedingly representative digital coin based on the blockchain technique (Hughes et al., 2019, Jiang et al., 2022). Consequently, we use bitcoin to depict blockchain-based digital currency, which aligns with the future development trend of digital currency like CBDC.

Conclusion

     In summary, the main issue for this paper to accommodate is to deliver a sensible production plan based on the volatility clustering model regarding manufacturing companies with the application of CBDC in the supply chain. We unravel the fact that CBDC volatility can generate considerable effect on the supply chain management for manufacturing companies. Consequently, manufacturing companies formulate the production plan based on volatility clustering model can be helpful in the reduction of uncertainties from both CBDC and BDI. Our framework is based on the volatility clustering model, and we schedule company-related activities into low-volatility periods. Under such a framework, we can thereby reduce the overall uncertainty levels faced by manufacturing companies, as both CBDC and BDI uncertainties can be lessened through rescheduling.

     To capture the feature of CBDC volatility, we meld both BTC volatility and GBP volatility in the rescheduling model, where BTC represents the digital currency side of CBDC and GBP represents the sovereign currency side of CBDC. We first used the GARCH model as the volatility clustering model for production planning, where the model performance is inferior. We argue that if we can consider volatilities in both time series simultaneously and filter the noise from the time series, the model performance could be enhanced. Hence, we adopt the DBSCAN technique, which is a type of machine learning approach to achieve these two goals, and the performance actually has been improved. Thus, we can conclude that manufacturing companies’ production plans can be strengthened through uncertainty reduction. We combine BTC volatility with GBP volatility to reveal the CBDC volatility effect on SCM, and we schedule the production plan within those low volatility periods for both BTC and GBP accordingly. On this basis, our model verifies that the enhanced production plan can help companies to have better cost and revenue management, as it reduces costs and increases revenue.

1- Introduction

2- Background

3- Design and implementation of the parallel spectral clustering in Julia

4- Results

5- Discussion

6- Conclusion

References

Abstract

Spectral clustering is widely used in data mining, machine learning and other fields. It can identify the arbitrary shape of a sample space and converge to the global optimal solution. Compared with the traditional k-means algorithm, the spectral clustering algorithm has stronger adaptability to data and better clustering results. However, the computation of the algorithm is quite expensive. In this paper, an efficient parallel spectral clustering algorithm on multi-core processors in the Julia language is proposed, and we refer to it as juPSC. The Julia language is a high-performance, open-source programming language. The juPSC is composed of three procedures: (1) calculating the affinity matrix, (2) calculating the eigenvectors, and (3) conducting k-means clustering. Procedures (1) and (3) are computed by the efficient parallel algorithm, and the COO format is used to compress the affinity matrix. Two groups of experiments are conducted to verify the accuracy and efficiency of the juPSC. Experimental results indicate that (1) the juPSC achieves speedups of approximately 14×∼ 18× on a 24-core CPU and that (2) the serial version of the juPSC is faster than the Python version of scikitlearn. Moreover, the structure and functions of the juPSC are designed considering modularity, which is convenient for combination and further optimization with other parallel computing platforms.

Introduction

In recent years, machine learning has made great progress and become the preferred method for developing practical software, such as computer vision, speech recognition, and natural language processing [37,45,50,55]. Machine learning mainly includes supervised learning and unsupervised learning. Clustering is the main content of unsupervised learning. Among many clustering algorithms, the spectral clustering algorithm has become the popular one [32,53]. Spectral clustering is a technology originating from graph theory [17,29] that uses the edge connecting them to identify the nodes in the graph and allows us to cluster non-graphic data. Unsupervised clustering analysis algorithm can explore the internal group structure of data, which has been widely used in various data analysis occasions, including computer vision analysis, statistical analysis, image processing, medical information processing, biological science, social science, and psychology [19,44,51]. The basic principle of clustering analysis is to divide the data into different clusters. Members in the same cluster have similar characteristics, and members in different clusters have different characteristics. The main types of clustering algorithms include partitioning methods, hierarchical clustering, fuzzy clustering, density-based clustering, and model-based clustering [38]. The most widely used clustering algorithms are k-means [61], DBSCAN [39], ward hierarchical clustering [47], spectral clustering [53], birch algorithm [66], etc. It has been proven that the spectral clustering algorithm is more effective than other traditional clustering algorithms in references [46,56], but in the process of spectral clustering computation, the affinity matrix between nodes needs to be constructed, and storage of the affinity matrix requires much memory. It also takes a long time to achieve the first k eigenvectors of the Laplacian matrix. Thus, the spectral clustering algorithm is difficult to apply in the large-scale data processing.

1- Introduction

2- Related work

3- Preliminary

4- Clustering based domain adaptation

5- Experiments

6- Conclusion

References

Abstract

Despite great progress in face recognition tasks achieved by deep convolution neural networks (CNNs), these models often face challenges in real world tasks where training images gathered from Internet are different from test images because of different lighting condition, pose and image quality. These factors increase domain discrepancy between training (source domain) and testing (target domain) database and make the learnt models degenerate in application. Meanwhile, due to lack of labeled target data, directly fine-tuning the pre-learnt models becomes intractable and impractical. In this paper, we propose a new clustering-based domain adaptation method designed for face recognition task in which the source and target domain do not share any classes. Our method effectively learns the discriminative target feature by aligning the feature domain globally, and, at the meantime, distinguishing the target clusters locally. Specifically, it first learns a more reliable representation for clustering by minimizing global domain discrepancy to reduce domain gaps, and then applies simplified spectral clustering method to generate pseudo-labels in the domain-invariant feature space, and finally learns discriminative target representation. Comprehensive experiments on widely-used GBU, IJB-A/B/C and RFW databases clearly demonstrate the effectiveness of our newly proposed approach. State-of-the-art performance of GBU data set is achieved by only unsupervised adaptation from the target training data.

Introduction

Benefiting from convolutional neural networks (CNNs) [1, 2, 3, 4, 5], deep face recognition (FR) has been the most efficient biometric technique for identity authentication and has been widely used in enormous areas such as military, finance, public security as well as our daily life. However, deep networks which perform perfectly on benchmark datasets may fail badly on real world applications. This is because the set of real world images is infinitely large and so it is hard for any dataset, no matter how big, to be representative of the complexity of the real world. One persuasive evidence is presented by P.J. Phillips’ study [6] which conducted a cross benchmark assessment of VGG model [7] for face recognition. The VGG model, trained on over 2.6 million face images of celebrities from the Web, is a typical FR systems and achieves 98.95% on LFW [8] and 97.30% on YTF [9]. However, It only obtains 26%, 52% and 85% on Ugly, Bad and Good partition of GBU database, even if all of images in GBU are nominally frontal. The main reason is a different distribution between training data (source domain) and testing data (target domain), referred to as domain or covariate shift. Visual examples of this domain shift are shown in Fig. 1. Each dataset in Fig. 1 displays a unique “signature” and thus one can easily distinguish them only by these signatures, which proves the existence of significant discrepancies. The images in CASIA-WebFace [10] are collected from Internet under unconstrained environment and most of the figures are celebrities and public taken in ambient lighting; The GBU [11] contains still frontal facial images and is taken outdoors or indoors in atriums and hallways with digital camera; IJB-A [12] covers large pose variations and contains many blurry video frames. Sometimes, the images of GBU and IJB-A datasets may be closer to the ones in real life which are taken with digital camera under different shooting environments and contain larger variations.

An u-shapelet is a sub-sequence of a time series used for the clustering of time series datasets. The purpose of this paper is to discover u-shapelets on uncertain time series. To achieve this goal, we propose a dissimilarity score called FOTS whose computation is based on the eigenvector decomposition and the comparison of the autocorrelation matrices of the time series. This score is robust to the presence of uncertainty; it is not very sensitive to transient changes; it allows capturing complex relationships between time series such as oscillations and trends, and it is also well adapted to the comparison of short time series. The FOTS score is used with the Scalable Unsupervised Shapelet Discovery algorithm for the clustering of 63 datasets, and it has shown a substantial improvement in the quality of the clustering with respect to the Rand Index. This work defines a novel framework for the clustering of uncertain time series.

Introduction

All measurements performed by a mechanical system contain uncertainty. Indeed, the uncertainty principle is partly a statement about the limitations of mechanical systems ability to perform measurements on a system without disturbing it [1]. Thus, time series from measurement instruments are uncertain. These time series produced by sensors constitute a vast proportion of the time series used in science, whether in medicine with ECGs, in physics with measurements recorded by telescopes, in computing with the Internet of Things and so on. Ignoring the uncertainty of the data during their analysis can lead to inaccurate conclusions [2], hence the need to implement uncertain data management techniques. Several recent studies have focused on the processing of uncertainty in data mining. Rizvandi et al.[3] studied CPU utilization time patterns of several MapReduce applications using Dynamic Time Warping and Euclidian distance for comparing times series, and they investigated the minimum distance/maximum similarity of these applications. Their results showed the effectiveness of their approach on a private cloud with up to 25 virtual nodes. Considering that time series data often contain uncertainty and that DUST is one of the latest methods that can deal with arbitrary probability distributions, but that its computational cost is high particularly when the dataset is large, Hwang et al. [4] demonstrated that the performance of DUST was much faster using GPU than the CPU-based implementation. Rehfeld and Kurths [5] investigated similarity estimators that could be suitable for the quantitative investigation of dependencies in irregular and age-uncertain time series like paleoclimate time series.

1. Introduction

2. Background: Metrics for detecting influential nodes in large biological networks

3. Proposed parallel algorithm for detecting influential nodes in large biological networks

4. Results

5. Discussion

6. Conclusion

Acknowledgments

References

Abstract

In biological networks, some nodes are more influential than others. The most influential nodes are those whose elimination induces a network collapse, and detecting these nodes is crucial in many circumstances. However, this is a difficult task when the size of the biological networks is large. In this paper, we have designed and implemented an efficient parallel algorithm for detecting influential nodes for large biological networks by exploiting a Graphics Processing Unit (GPU). The essential concept behind the proposed parallel algorithm is that several computationally expensive procedures in detecting influential nodes are redesigned and transformed into quite efficient GPUaccelerated primitives such as parallel sort, scan, and reduction. Four local metrics, including the Degree Centrality (DC), Companion Behavior (CB), Clustering Coefficient (CC), and H-Index, are used to measure the nodal influence. To evaluate the efficiency of the proposed parallel algorithm, five large real biological networks are employed in the experiments. The experimental results show that (1) the proposed parallel algorithm can achieve speedups of approximately 48∼94 over the corresponding serial algorithm; (2) compared to a baseline parallel algorithm developed on a multi-core CPU, the proposed parallel algorithm yields speedups of 5∼9 for DC and H-Index, while it is slightly slower for CB and CC due to the uneven degree distribution; and (3) when using DC and H-Index, the proposed parallel algorithm is capable of detecting the influential nodes in a large biological network consisting of 150 million edges in less than 3 s.

Introduction

In recent years, complex network analysis has received increasing attention. Many real complex systems can be abstractly regarded as complex networks for presenting the complexities of real systems [1], such as social networks, technological networks, information networks, and biological networks. Various methods have been proposed for mining information from complex networks. In particular, detecting the influential nodes in complex networks is a topic of interest drawing much attention in this research field [2–4].

Detecting influential nodes in complex networks can be exploited to mine the features and functions of these networks [2]. Much research has been conducted to rank and identify influential nodes in the aforementioned four network categories. In the detection of influential nodes in various networks, the first critical issue is to select or define the metrics for measuring the influence of each node; the second is to employ or develop specific algorithms to effectively and efficiently determine and rank the influential nodes.

Many metrics have been proposed for detecting the influential nodes in complex networks. These metrics can be roughly divided into two categories [4]: (1) local metrics that are calculated based on the local structures of networks and (2) global metrics that are calculated based on the global structures of networks. For the first category, the most commonly used metrics include the degree centrality, clustering coefficient, and H-Index [5]. For the second category, the most commonly used metrics are the betweenness centrality [6], closeness centrality, PageRank, k-core [7] computed using the k-shell decomposition [8], and the bidirectional k-core (B-core) [9].

Clustering is the most widely used unsupervised machine learning technique, having extensive applications in statistical analysis. We have multiple clustering algorithms available in theory and many more implementations available in practice. A bunch of literatures can be found focusing on the quality of clustering algorithms using various internal and external evaluation techniques. The motivation behind this work is the scarcity of literatures dealing with performance of clustering algorithms in terms of turnaround time. This paper summarizes the experimental analysis conducted on the performance of multiple clustering algorithms based on cardinality and dimensionality. The analysis is performed in R, which is a free and open source programming language mainly used for statistical computing. This work evaluates nine key algorithms coming under partitioning, hierarchical, density-based and model-based clustering approaches using different social media data sets. We captured performance trends of these algorithms in terms of turnaround time by varying the cardinality and dimensionality parameters of the data sets. Based on our experiments, CLARA, CLARANS, and k-means algorithms demonstrate best performances with varying cardinality. It is also observed that changes in dimensionality do not impact hierarchical clustering approaches whereas there is a positive influence on the execution time for partitioning, density-based and model-based clustering approaches.

Introduction

Data mining [1] is the process of extracting meaningful information from raw data through which underlying patterns and relationships are revealed. These revelations form useful knowledge that can be made use of various scientific, educational, and/or industrial scenarios. Based on the type of patterns to be processed, we can adopt appropriate data mining strategies which include, but not limited to classification, clustering, association, regression, etc. Clustering is the machine learning technique used for creating logical groups of similar entities from a data set. The aim of clustering process is to create distinct groups of elements in such a way that the entities from the same group will have similar properties whereas entities from different groups have dissimilar properties. It is an unsupervised learning technique which is widely used for per-forming statistical analysis of data. Since the volume of data being processed is increasing on a daily basis, clustering is extensively applied in almost all industrial segments. This work covers an empirical analysis of the performance of nine different clustering algorithms [2]. We captured the average processing time for each algorithm against varying number of records (cardinality) with constant number of attributes (dimensionality), and varying number of attributes with same number of records. The experiments were conducted using two distinct social media data sets.

1- Introduction

2- CB trip/close CC signals and CB operation

3- Theoretical basis of proposed method

4- Optimized AP clustering algorithm

5- Fault diagnosis method

6- Simulation results

7- Conclusion

References

Abstract

Online condition monitoring and fault diagnosis of circuit breakers (CBs) is a significant method to effectively improve the stability and reliability of the power system. However, the currently used fault diagnosis method still have certain defects including the inability to identify unknown faults for training samples. Therefore, this paper proposes an evolving method for fast and accurate online fault diagnosis of CBs. On the basis of collecting samples of CB trip/close coil current (CC) features, an optimized affinity propagation (AP) clustering algorithm to accurately extract the sample clustering exemplars is presented. Additionally, operating state identification and fault diagnosis of CBs is carried out by calculating the similarity coefficient between the new sample and exemplars online. Diagnosis of unknown faults is also achieved by introducing the threshold and comparing it with similarity coefficient results. Simulation results prove that the proposed method can precisely identify various known CBs faults and has the ability to recognizes unknown CBs fault samples even when the number of training samples is small, providing a foundation for CB fault location and condition-based maintenance.

Introduction

Circuit breakers (CBs) are recognized as one of the most crucial components to power equipment. They are the key to isolating faulty components driven by protection devices, and play a dual role in the protection and control of the power system [1]. Incorrect operation of CBs can cause a power grid accident or expand the scope of the accident. In severe cases of CB failure, the power system may collapse and cause major economic losses. Therefore, online monitoring and fault diagnosis of CBs have practical significance for enhancing the reliability and stability of the electric power system. According to the CIGRE surveys, more than 80 per cent of the fault of CBs is caused by operating mechanism and auxiliary control circuits failures. The trip and close coil current (CC) signal is an accessible and noninvasive parameter in CB online condition monitoring. Previous studies have determined that analysis of the CC characteristics can identify effectively many signs of various faults type occurring in control circuits and operating mechanism [2–4]. In recent years, increasingly advanced data-analytics algorithms have been applied to implement the fault diagnosis of various kinds of power equipment [5,6], and also provide a new approach in the fault diagnosis of CBs [7–10]. In Ref. [11], a fault diagnosis method was proposed which has high detection accuracy, and utilizes back-propagation neural network (BPNN) technique. BPNN requires a large number of training samples to ensure an accurate diagnosis. However, it is usually quite difficult to obtain general and sufficient CC fault samples in practical applications, which is because the operating frequency and average failure rate of CBs are relatively low, and different types of the CB may have different CC waveform. In addition, Ref. [12] proposed a fault diagnosis method combining CC characteristics and support vector machine algorithm, which can obtain better diagnostic results compared with the BPNN method in the small training sample case.

I- INTRODUCTION

II- PROBLEM STATEMENT

III- RELATED WORK

IV- THE COLLEGE ADMISSION ALGORITHM

V- THE PROPOSED ALGORITHM

VI- DATASETS

VII- CLUSTERING RESULTS AND DISCUSSION

VIII- CONCLUSION

REFERENCES

Abstract

This paper proposes a novel supervised clustering algorithm to analyze large datasets. The proposed clustering algorithm models the problem as a matching problem between two disjoint sets of agents, namely, centroids and data points. This novel view of the clustering problem allows the proposed algorithm to be multi-objective, where each agent may have its own objective function. The proposed algorithm is used to maximize the purity and similarity in each cluster simultaneously. Our algorithm shows promising performance when tested using two different transportation datasets. The first dataset includes speed measurements along a section of Interstate 64 in the state of Virginia, while the second dataset includes the bike station status of a bike sharing system (BSS) in the San Francisco Bay Area. We clustered each dataset separately to examine how traffic and bike patterns change within clusters and then determined when and where the system would be congested or imbalanced, respectively. Using a spatial analysis of these congestion states or imbalance points, we propose potential solutions for decision makers and agencies to improve the operations of I-64 and the BSS. We demonstrate that the proposed algorithm produces better results than classical kmeans clustering algorithms when applied to our datasets with respect to a time event. The contributions of our paper are: 1) we developed a multi-objective clustering algorithm; 2) the algorithm is scalable (polynomial order), fast, and simple; and 3) the algorithm simultaneously identifies a stable number of clusters and clusters the data.

INTRODUCTION

WITH the growth of new technologies, smart cities and urban areas are adapting advanced devices to control and monitor transportation networks and thus provide better service to the public and private sectors. These devices collect data through many sensors in the city’s infrastructure. Agencies and researchers exploring the massive amounts of collected data often find it challenging to draw meaningful conclusions due the sheer size of the datasets. One way to deal with such data is to use clustering approaches. In the transportation field, operating agencies (such as departments of transportation) have been collecting data to improve the efficiency of the transportation network and provide a better service for all transportation modes. Clustering the travel times or speeds of transportation modes could help operating agencies to better manage the transportation network. In particular, the collected data could be reduced to find the cluster centroids (i.e., the means of the clusters) that represent the entire data with respect to a time event such as time of day, day of month, and month of the year. This could help operating agencies answer several questions related to traffic operations such as, “Can we discriminate between recurrent congestion and outliers?” and “Can we identify how many time periods we need to plan for in terms of resource and congestion management?”

abstract

The evaluation of subjective data is a very demanding task. The classification of the information gathered from human evaluators and the possible high noise levels introduced are ones of the most difficult issues to deal with. This situation leads to adopt individuals who can be considered as experts in the specific application domain. Thus, the development of Expert Systems (ES) that consider the opinion of these individuals have been appeared to mitigate the problem. In this work an original methodology for the selection of subjective sequential data for the training of ES is presented. The system is based on the arrangement of knowledge acquired from a group of human experts. An original similarity measure between the subjective evaluations is proposed. Homogeneous groups of experts are produced using this similarity through a clustering algorithm. The methodology was applied to a practical case of the Intelligent Transportation Systems (ITS) domain for the training of ES for driving risk prediction. The results confirm the relevance of selecting homogeneous information (grouping similar opinions) when generating a ground truth (a reliable signal) for the training of ES. Further, the results show the need of considering subjective sequential data when working with phenomena where a set of rules could not be easily learned from human experts, such as risk assessment.

Introduction

The practice of Knowledge Engineering (Van Do, Le Thi, & Nguyen, 2018) has become a very useful approach to solve complex problems that require a high level of human expertise. This discipline involves integrating knowledge into computer systems which emulates the decision-making ability of a human expert in a specific domain. The systems in charge of achieving these tasks are the Expert Systems (ES) (Agarwal & Goel, 2014). The building, maintaining and development of ES (Djamal et al., 2017) are mainly based on the interaction between the knowledge engineer and the domain expert (Yau & Sattar, 1994). The devel- opment of a reliable ES requires a deep understanding and a good representation of the knowledge of the domain expert. In most of the cases, the knowledge representation is based on a set of rules (a production system) that ease the explanation of the decision-making made by the inference engine (Wick & Slagle, 1989). These rules are build from the knowledge acquired from human experts with the application of Machine Learning techniques (such as Neural Networks (Lin & Zhang, 2012), Deep Learning (Wei, He, Chen, Zhou, & Tang, 2017), Decision Trees (Sriram & Yuan, 2012), Fuzzy Logic (Wang, Lee, & Ho, 2007), Bayesian methods (WenBin, XiaoLing, YiJun, & Yu, 2010), Genetic Algorithms (Daza et al., 2011), among others). Knowledge acquisition is a process which aims to extract knowledge, experience and problem-solving procedures from one or more domain experts. Several techniques have been proposed for a correct knowledge acquisition (see Hua, 2008 for a complete review). Nevertheless, there are several problems that must be considered when acquiring knowledge from human experts (Gaines, 1987): • Experts may not be able to express their knowledge in a structured way • Experts may not be aware of the significance of the knowledge they have used. • The expressed knowledge may be irrelevant, incomplete or not understandable. In some cases, depending on the field of application, it may be easier to extract the knowledge from human experts through a continuous scale. This is the case of the risk assessment, where the knowledge could be acquired in a predefined scale (e.g. from 0, no risk, to 100, maximum risk). Here, the knowledge of the experts is gathered in form of subjective sequential data (Prelec, 2004) and could be treated as time series for its study and integration (see, for instance, de Diego, Crespo, Siordia, Conde, & Cabello, 2011; de Diego, Siordia, Conde, & Cabello, 2011; Siordia, de Diego, Conde, & Cabello, 2011a). However, the integration of several opinions into a unique ground truth (i.e. a reliable signal) is a hard-to-achieve task (Liou & Nunamaker, 1990). Two different scenarios appear. The consideration of knowledge from too few experts could provide a ground truth with insufficient information. In contrast, the consideration of knowledge from too many experts could generate a noisy ground truth due to the appearance of possible contradictions between their evaluations (Turban, 1991). Different statistical approaches have been proposed in the past (see, for instance, meta-analysis methods in Brockwell & Gordon (2001)).

A novel clustering algorithm by fast search and find of density peaks (DP) was proposed in Science, 2014. It has attracted much attention from researchers. It can easily select clusters centers with decision graph. However, it cannot be used to cluster manifold data sets as the existing distance measurement is not suitable to evaluate the dissimilarity between objects on manifold structure. Some researchers use graph-based distance to measure the dissimilarity between objects on manifold clusters, but computing the graph-based distance on the original data set is time consuming. An improved density peaks clustering algorithm based on shared-neighbors between local cores, SLORE-DP, is proposed in this paper. First, it finds local cores to represent the data set and redefines the graph-based distance between local cores with sharedneighbors-based distance. Then natural neighbor-based density and the new defined graph-based distance are used to construct decision graph on local cores and DP algorithm is employed to cluster local cores. Finally, the remaining points are assigned to the same cluster as their local cores belong to. Since we use the new defined graph-based distance to estimate the dissimilarity between local cores, SLORE-DP can be used to cluster manifold data sets and at the same time it only calculates the shortest path between local cores, which greatly reduces the running time of the algorithm. We do experiments on several synthetic data sets containing manifold clusters and several real data sets from UCI. The results show that SLORE-DP is more effective and efficient than other algorithms when clustering manifold data sets.

Introduction

As an unsupervised learning, clustering is an important method for data analysis. It has been widely used in the field of pattern recognition, image processing, and information retrieval. It is designed to divide objects into multiple clusters, so that similar objects are in the same cluster while different objects are in different clusters. Many clustering algorithms have been proposed over the past few decades. According to different strategies, these algorithms can be roughly grouped into partitioning methods, density-based methods, hierarchical methods, model-based methods and grid-based methods. Among them, partitioning, density-based and hierarchical algorithms, due to their simple principle, are the most popular. K-means [1] and K-medoids [2] are typical partitioning algorithms. However, their performance depends on the selection of initial cluster centers. To avoid selecting cluster centers, AP algorithm [3] treats all objects as potential centers. K-AP [4] is an improved AP algorithm. It uses the immediate result of K clusters by introducing a constraint in the process of message passing. However, since each point is always allocated to the nearest center, these algorithms cannot discover arbitrary-shaped clusters. DBSCAN [5] is a typical density-based clustering algorithm. It defines clusters as dense regions separated by sparse regions. Dcore [6] is a hybrid decentralized approach which is based on finding density cores instead of centroids.