In order to make full use of the historical knowledge and improve the accuracy of rating prediction,a recommendation model based on Lifelong Machine Learning(LML) is proposed to mine both user ratings and comments.The model accumulates knowledge from previous tasks and utilizes it in future tasks to help improve the rating prediction accuracy.Experimental results on real datasets show that compared with models without LML ability,the mean square error of the predicted ratings of this model is reduced by 5.4‰,and with the accumulation of knowledge,its error is continuely dropped.The accuracy of topic word classification results is improved.

In order to improve energy consumption,load balance,and Service Level Agreement(SLA) violation rate of cloud data centers,it is necessary to optimize virtual machine placement strategy.Therefore,based on the IaaS environment,a virtual machine migration adjustment method based on machine learning is proposed.The virtual machine is pre-placed according to the complementarity and imbalance of resource consumption,the deep neural network is used to predict the physical machine load level,and the Deep Q Network(DQN) is used to adjust the number of physical machines.Experimental results show that this method can effectively balance load distribution,reduce energy cost and SLA violation rate.

In the era of big data,traditional hidden hyperlink detection technology cannot quickly and accurately identify websites that encounter “hidden hyperlink attacks” on massive Web pages.To solve this problem,this paper introduces machine learning to the detection method for hidden hyperlink,which combines the characteristics of hidden hyperlink related texts,hidden hyperlink domains and the hidden structure of hidden hyperlink.The three models are constructed and compared using Classification and Regression Tree (CART),Gradient Boosted Decision Tree (GBDT) and Random Forest (RF).based on the proposed method.Experimental results show that the proposed method has high accuracy and reliability,and the classification accuracy of the detection model constructed by RF can reach 0.984.

Advanced Persistent Threat(APT) whose main goal is to steal information becomes a dangerous attack in recent years,which can bring high risk and persistent attack to the mobile devices.The detection features of mobile terminal intrusion detection are not readily available,so the accuracy of detection model is not high enough and there is the over-fitting problem,which lead to poor detection effect.For these problems,this paper proposes an optimized detection model using static detection technology to extract terminal devices’ static features which helps optimize the sensitivity of the model to the malicious application.It uses sliding window iterative algorithm to extract the delaying attack feature so as to optimize the model’s detection capability of delaying attack and uses Boost technology to fuse the classification algorithms including the decision tree,logistic regression and Bayesian classifier.Experimental results show that the model can increase the detection accuracy of APT effectively and avoid over-fitting problem.

A threat that Domain Name System(DNS) data is tampered by hackers endangers DNS applications. Due to the hidden characteristic of this threat,a quick and effective method to find dangerous changes in DNS data is needed urgently. Regarding to the problem,this paper proposes a method to monitor the DNS data based on machine learning,by which dangerous change in DNS data can be found quickly. Some domain names whose data are changed are chosen from a number of domain names,and their relevant information is individually analyzed in order to produce a tuple that is represented by a multi-dimensional attribute vector,which contains literal characteristics,forward-inverse match and so on. After that a class is labeled depending on whether the changes are bad or not so that an instance containing the tuple and their class label is built and consequently a training set is built. By analyzing the training set the two classification algorithms,decision tree and Support Vector Machine(SVM),build classifiers,which are used to detect whether changes in DNS data are dangerous or not. The 10-fold cross-validation is used to validate the two classifiers. It is found that the classifiers do well in finding dangerous changes in DNS data,in which the present results show that the classifier can reach a good precision,and their weighted average accuracies are 73. 8% and 82. 4% .

This paper presents a system for ellipsis identification in Chinese which is based on machine learning. The system can be used to select a number of features and feature combinations through preprocessing the corpus. And Chinese ellipsis identification can also be achieved by the ellipsis identification model built by Support Vector Machine(SVM). The performance of the system in different parser tree is studied as well. Experimental result shows that the system has F value of 84.01% on the standard parser tree, and 68.22% on automatic sentence parser tree.

To deal with the new time and space challenges of the machine learning problem algorithms from large scale data, this paper focuses on sparse-learning and categorizes the L1 regularized problem’s the-state-of-the-art solvers from the view of multi-stage and multi-step optimization schemes. It compares the algorithms’ convergence properties, time and space cost and the sparsity of these solvers. The analysis shows that those algorithms sufficiently exploiting the machine learning problem’s specific structure obtain better sparsity as well as faster convergence rate.

A natural feature recognition method based on machine learning is proposed for 3D registration in augmented reality application. This method increases the accuracy of key-points recognition and moves the computational burden from runtime matching to offline training by substituting specific classification for nearest-neighbor searching. Robust camera tracking and pose estimation can be obtained by the similarity of these matched key-points and the homography matrix. Experimental results demonstrate that this method is suitable for real-time application and is stable against illumination change, occlusion and perspective effect.

The machine learning algorithms play an important role in current spam filter, but a single machine learning algorithm has its own drawback which restrains it from further application in E-mail filter. This paper introduces some typical machine learning algorithms, and constructs a voting E-mail filter model based on multi-machine learning algorithms. The experiments show that this method makes use of every machine learning algorithm’s advantage, and offsets its disadvantage, and achieves better filter performance than a single algorithm.

Intrusion detection system has some defects, such as signatures being generated manually, updating difficulty and doing nothing in front of large data set. This paper discusses intrusion detection system with machine learning techniques. By making usage of Gene algorithm and Bayes classifiers, the defects mentioned above can be reduced to some extent and some tests have been done to show machine learning magic capability in intrusion detection system.