Research On Metric Learning Algorithms Based On Local Information

With the rapid development of the internet,high performance computing and other technologies,artificial intelligence has significant breakthroughs in theory and technology,which has a profound impact on the economy and society.Machine learning has been successfully applied to many areas of artificial intelligence,however,many traditional distance metric functions are difficult to capture the real data structure in complex and changing application scenarios.As an important research area in machine learning,metric learning has received a lot of attention for its ability to adaptively obtain discriminative metrics from data.At present,metric learning has made a series of important advances in theory,methods and applications,but it still faces many challenges,such as difficulty in information sample tuple mining,strong dependence on supervised information,easy overfitting of algorithms and limited scalability.(1)To address the difficulty of mining information sample tuples,an optimization model is proposed in which sample tuple construction and metric learning are mutually guided.In this model,large interval triple sample tuples are incorporated into a clustering model in which the target samples are replaced with cluster centers.As a result,the number of triple sample tuples formed with cluster centers is significantly reduced.Moreover,we use clustering to dynamically mine the sample tuples,while the clustering process is also influenced by metric learning,thus we integrate the information sample tuple mining with the metric learning process.Experimental results on synthetic and real datasets show that the proposed algorithm is able to dynamically mine the sample tuples,which improves the classification performance of the learned metric on the classifier.(2)To tackle the problem of strong dependence on supervised information,a metric learning model by perturbing hard-to-classify samples is proposed.The model first gives a perturbation method for the samples and defines the distance of the samples after perturbation.Secondly,the degree of difficulty in hard-to-classify samples is described by the loss of sample tuples constructed from the local nearest neighbor information of each sample.Subsequently,the loss of constructing a sample tuple with a hard sample is directly reduced by moving each hard sample in the similar sample direction while keeping the other samples unchanged,thus alleviating the performance degradation caused by the overfitting of metric learning to hard-to-classify samples.Experimental results show that the proposed metric learning model can effectively reduce the impact of hard-to-classify samples on metric learning by perturbing hardto-classify samples,and obtain better performance on real datasets,people re-identification and face verification datasets.(3)To overcome the problem that most multiple metric learning methods have too many parameters and are prone to overfitting,a multiple metric learning framework through local metric fusion is proposed.The framework unifies the fusion of similar local metrics into a single metric and is able to determine the number of local metrics adaptively.In addition,we apply the framework to a binary sample tuple to construct a specific optimization model and to obtain a closed-form solution of the model.Specifically,the model learns a global metric for all sample tuples,and under the global metric,a local metric is learned for some of the sample tuples with large losses to reduce their losses.Then,the number of metrics is gradually reduced by continuously fusing similar local metrics.The metrics from each fusion process are also applied to the 3-nearest neighbor classifier to continuously predict the performance of the classifier.The optimal number of metrics is determined by selecting the optimal classifier performance.Experimental results show that the number of local metrics can be effectively determined by fusing similar local metrics,and the construction of local metrics from the perspective of sample tuples can effectively enhance the fitting ability of multiple metric learning models,and obtain better results compared with other representative global metric learning and multiple metric learning methods.(4)To deal with the problem that metric learning is difficult to extend to high-dimensional data,a joint dimension reduction and multiple metric learning learning framework is proposed.The framework requires that the metrics learned in different similar localities should be consistent as much as possible under the reduced dimensional representation of the data,which can effectively reduce the negative impact of interfering features or noisy features on the metric learning.At the same time,the proposed model is solved simply and effectively.Experimental results show that the consistency of different local metrics under the reduced dimensional representation of data can effectively reduce the negative effect of interfering features or noise features on the metric learning,which makes the model better than the existing methods of single metric learning,multiple metric learning and joint dimension reduction and metric learning.In summary,this thesis proposes a series of metric learning methods based on local information such as the neighborhood of samples,the sample tuples and local clusters,combined with clustering,sample perturbation,dimension reduction and other techniques to address the problems of difficulty in information sample tuple mining,strong dependence on supervised information,easy overfitting of algorithms and limited scalability.These methods provide new ideas for model construction and optimal solution of metric learning,and have wide application prospects in image retrieval,people re-identification,face verification and other fields.