Prototypes help to explain the predictions of deep classification models for time series. However, most models learn prototypes by randomly initializing an uncertain number of low-discriminative prototypes, which may lead to unstable models and unreliable results. To address these issues, we propose a new class D iscriminative P rototype L earning Net work (DPL-Net), which learns an appropriate number of class-discriminative prototypes, thus improving classification performance. Specifically, the proposed P rototype I nitialization M echanism (PIM) introduces a new proximity metric based on the silhouette coefficient and statistical metrics. It facilitates the automatic determination of the class-discriminative prototypes for each class. Then, the encoder layer encodes the prototypes derived from PIM and the input series using one-dimensional convolutional neural networks (1D-CNN). Finally, the prototype classification layer optimizes the prototypes according to the regularization terms, while simultaneously classifying the input sequence based on its similarity to the updated prototypes. The comparison experiments are conducted on 26 UCR datasets compared with 10 baselines. The results show that our proposed approach achieves the best accuracy on 11 datasets. Specifically, our method outperforms PIP, CSSL, and LSS by an average of 16.33%, 9.77% and 5.96% on 22, 14 and 16 datasets, respectively. The interpretability experimental results and the application analysis on spectral data indicate that the learned prototypes can provide reasonable explanations for the classification results of the model.

The extensive growth of data quantity has posed many challenges to data analysis and retrieval. Noise and redundancy are typical representatives of the above-mentioned challenges, which may reduce the reliability of analysis and retrieval results and increase storage and computing overhead. To solve the above problems, a two-stage data pre-processing framework for noise identification and data reduction, called ARIS, is proposed in this article. The first stage identifies and removes noises by the following steps: First, the influence space (IS) is introduced to elaborate data distribution. Second, a ranking factor (RF) is defined to describe the possibility that the points are regarded as noises, then, the definition of noise is given based on RF. Third, a clean dataset (CD) is obtained by removing noise from the original dataset. The second stage learns representative data and realizes data reduction. In this process, CD is divided into multiple small regions by IS. Then the reduced dataset is formed by collecting the representations of each region. The performance of ARIS is verified by experiments on artificial and real datasets. Experimental results show that ARIS effectively weakens the impact of noise and reduces the amount of data and significantly improves the accuracy of data analysis within a reasonable time cost range.

In this paper, a novel trajectory clustering algorithm - TAD - is proposed to extract trajectory Stays based on spatial-temporal density analysis of data. Two new metrics - NMAST (Neighbourhood Move Ability and Stay Time) density function and NT (Noise Tolerance) factor - are defined in this algorithm. Firstly, NMAST integrates the characteristics of Neighbourhood Move Ability (NMA, extended from the concept of Move Ability MA), Stay Time (ST), and evaluation factor Eμ to measure the spatial-temporal density of data. Secondly, NT utilizes the features of noise to dynamically evaluate and reduce the influence of noise. The experimental results on Geolife dataset shows that the distributions hidden in data are extracted more realistically, especially for various complex or special trajectories with long-duration gaps. Furthermore, our analytical method of trajectory data is particularly applied in the spectra of LAMOST survey to analyse the variation characteristics of sky-background. The results show a regular distribution on observational date which is relatively concentrated in the month of 1, 10, 11, 12 in each year. The laws discovered in this work would provide a reasonable support for the designation of observational plans, and the new trajectory analysis method would also provide the services for the astronomical data analysis and then for the further studies of formation and evolution of the universe.

In this study, we discover that the data skewness problem imposes adverse impacts on MapReduce-based parallel kNN-join operations running clusters. We propose a data partitioning approach-called kNN-DP-to alleviate load imbalance incurred by data skewness. The overarching goal of kNN-DP is to equally divide data objects into a large number of partitions, which are processed by mappers and reducers in parallel. At the heart of kNN-DP is a data partitioning module, which dynamically and judiciously partitions data to optimize kNN-join performance by suppressing data skewness on Hadoop clusters. Data partitioning decisions largely depends on data properties (e.g., distributions), the analysis of which is highly expensive for a massive amount of data. To speed up the data-property analysis, we incorporate a sampling technique to profile the data distribution of a small sample dataset representing big datasets. After building a data-partitioning cost model for parallel kNN-joins, we derive the time-complexity upper and lower bounds of parallel kNN-join algorithms. The cost model offers us a guidance to systematically investigate kNN-DP's performance. kNN-DP obtains global nearest neighbors using local nearest neighbors. To improve the accuracy of such an approximation solution, we augment each node's local data by a small amount of redundant data. We develop two kNN-DP-based schemes called LSH+ and z-value+, which seamlessly integrate kNN-DP with the existing LSH and z-value algorithms for kNN-join computing. We implement and evaluate LSH+ and z-value+ on a 24-node Hadoop cluster driven by both synthetic and real-world high-dimensional datasets. The experimental results show that kNN-DP significantly improves the performance of LSH and z-value while offering high extensibility and scalability on Hadoop clusters.

Traditional parallel algorithms for mining frequent itemsets aim to balance load by equally partitioning data among a group of computing nodes. We start this study by discovering a serious performance problem of the existing parallel Frequent Itemset Mining algorithms. Given a large dataset, data partitioning strategies in the existing solutions suffer high communication and mining overhead induced by redundant transactions transmitted among computing nodes. We address this problem by developing a data partitioning approach called FiDoop-DP using the MapReduce programming model. The overarching goal of FiDoop-DP is to boost the performance of parallel Frequent Itemset Mining on Hadoop clusters. At the heart of FiDoop-DP is the Voronoi diagram-based data partitioning technique, which exploits correlations among transactions. Incorporating the similarity metric and the Locality-Sensitive Hashing technique, FiDoop-DP places highly similar transactions into a data partition to improve locality without creating an excessive number of redundant transactions. We implement FiDoop-DP on a 24-node Hadoop cluster, driven by a wide range of datasets created by IBM Quest Market-Basket Synthetic Data Generator. Experimental results reveal that FiDoop-DP is conducive to reducing network and computing loads by the virtue of eliminating redundant transactions on Hadoop nodes. FiDoop-DP significantly improves the performance of the existing parallel frequent-pattern scheme by up to 31 percent with an average of 18 percent.