Jian-Pei Zhang

A parallel SVM training algorithm on large-scale classification problems

Support vector machine (SVM) has become a popular classification tool but the main disadvantages of SVM algorithms are their large memory requirement and computation time to deal with very large datasets. To speed up the process of training SVM, parallel methods have been proposed by splitting the problem into smaller subsets and training a network to assign samples of different subsets. A parallel training algorithm on large-scale classification problems is proposed, in which multiple SVM classifiers are applied and may be trained in a distributed computer system. As an improvement algorithm of cascade SVM, the support vectors are obtained according to the data samples distance mean and the feedback is not the whole final output but alternating to avoid the problem that the learning results are subject to the distribution state of the data samples in different subsets. The experiment results on real-world text dataset show that this parallel SVM training algorithm is efficient and has more satisfying accuracy compared with standard cascade SVM algorithm in classification precision.

A divisional incremental training algorithm of support vector machine

Support vector machine (SVM) has become a popular classification tool but the main disadvantages of SVM are their large memory requirement and computation time to deal with very large datasets. Therefore we prefer to incremental learning algorithms especially when the data available are obtained at different intervals. The key of SVM to incremental training is to assure the final results consists of almost all support vectors. This paper proposes a divisional incremental training algorithm of SVM, considering the possible impact of new training data to history learning results. Training data are divided into smaller sets to decrease the computation complexity and the support vectors are obtained in a crossed way. The experiment results on the real-world test dataset show that the classification accuracy is satisfying, and the efficiency of proposed incremental algorithm is superior to that of batch SVM model.

Dynamic Incremental SVM learning Algorithm for Mining Data Streams

Incremental SVM framework is often designed to deal with large-scale learning and classification problems. The paper presents a new dynamic incremental learning algorithm for mining data streams. The multiple classifiers are constructed according to the statistic characters of batched training data in data streams. The feature space of all data is partitioned according to the performance of each classifier and the statistical characters on each region are counted. The classifier that has the best performance on the region near the test data is selected as the final output. The experimental results confirm the feasibility and validity of the proposed algorithm.

A New Kind of Hybrid Cache Coherence Protocol for Multiprocessor with D-Cache

This paper introduces a new kind of hybrid Cache coherence protocol-MECSIF, which applicants for multiprocessor environment, based on hybrid cache line write strategy. Through the introduction of a small dictionary-D-Cache in system architecture, protocol overcomes the shortcoming of snoopy coherence protocol that data request was undifferentiated broadcasted. Protocol extends data block state so that eliminates ping-pang phenomenon, uses hybrid cache line write strategy to reduce L1 cache miss ratio. Simulation results show that the MECSIF protocol extent improves the efficiency of processor data access comparing with MESI protocol.

Research on Porting Technology of Kaffe Virtual Machine Based on ARM7/uClinux

Kaffe was ported to ARM7/uClinux which was popular embedded platform in this paper, and Java runtime environment was built to support executing Java applications. According to characteristics of ARM7 such as the length of instructions and data types, the interpreter core and the data structure of Kaffe were modified, which made ARM7 with the revised Kaffe can access Memory and solve the problem which came from the MMU-Less in the system, and provided technical method for porting Kaffe to other system platforms which was MMU-Less. Finally, make used of test tool set to test the completed system, and the test result did not only verify the systems availability, but also provided the reference for the designer of embedded system to choose Java virtual machine. KeywordsJava virtual machine; Kaffe, ARM7 processor; uClinux operating syste;.

A New Approach to Creating Spatial Index with R-Tree

Aimed at the traditional method expends time and overlapping area is big, this paper proposes a new creation method with R-tree. Through ranking the coordinate of spatial objects center point on 2D, a number of objects, which have minimal span, are picked up to establish the R-tree. The static batch load method optimizes spatial index structure, which can improve the spatial utilization and reduce the overlapping area. Experimental results show that this method can achieve a higher spatial utilization and reduce the time consumed, then increase the index capability. Therefore, the proposed method is correct and effective.

A category on the cache invalidation for wireless mobile environments

Wireless computing becomes most popular with the development of mobile computers. Cache technique in wireless computing is crucial because it facilitates the data access at clients for reducing servers’ loading, hence improve the performance. However, conventional cache technique requires coherence between servers and clients because of frequent disconnection. The cache invalidation strategy becomes an ideal method. In this paper, a category on the cache invalidation is proposed. To evaluate system performance, a mathematical model is proposed. It will develop high performance cache technique for practical wireless mobile computing.

An algorithm for mining strongly correlated pairs in relational table

Given a user-specified minimum correlation threshold and a relational table, the problem of mining all-strong correlated pairs is to find all attribute value pairs with Pearsons correlation coefficients above the minimum correlation threshold. However, algorithms developed for transaction database will generate invalid candidate pairs due to fundamental property of the itemsets in relational table (i.e. 1NF, they cannot contain more that one item per table column) and hence encounter additional and unnecessary computation cost. In this paper, using this property, the join step in the candidate generation phase is adapted to reflect this and to prune candidate set by not taking into itemsets which are not in 1NF. Furthermore, we propose other techniques to reduce the number of candidate pairs that are to be examined in the refinement step, even when the upper bound based pruning technique is useless in case of very low correlation threshold. Experimental results from real data sets exhibit that our algorithm can produce smaller candidate set and be faster than previous algorithms.

A heuristic algorithm to incremental support vector machine learning

Incremental learning techniques are possible solutions to handle vast data as information from Internet updating gets faster. Support vector machine works well for incremental learning model with impressive performance for its outstanding power to summarize the data space in a concise way. This paper proposes a heuristic algorithm to incremental learning with SVM taking the possible impact of new training data to history data into account. The idea of this heuristic algorithm is that the partition difference set has less elements, and existing hyperplane is much closer to the optimal one. New support vectors in this algorithm consist of existing support vectors and partition difference set of new training data and history data by separating hyperplane. The algorithm improves classification precision by adding partition difference set, and decreases the computation complexity by constructing new classification hyperplane on support vector set. The experimental results show that this heuristic algorithm is efficient and effective to improve the classification precision.

Transfer learning over big data

Transfer learning has emerged as a new learning technique facilitating an improved learning result of one task by integrating the well learnt knowledge from another related task. While much research has been devoted to develop the transfer learning algorithms in the field of long text analysis, the development of the transfer learning techniques over the short texts still remains challenging. The challenge of short text data analysis arises due to its sparse nature, noise words, syntactical structure and colloquial terminologies used. In this paper, we propose AutoTL(Automatic Transfer Learning), a transfer learning framework in short text analysis with automatic training data selection and no requirement of data priori probability distribution. In addition, AutoTL enables an accurate and effective learning by transferring the knowledge automatically learnt from the online information. Our experimental results confirm the effectiveness and efficiency of our proposed technique.