International Core Journal of Engineering 2020-26 | Page 122

2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM) Semi-supervised Structured Sparse Graph Data Classification Shuai Shao Mingze Tang University of Electronic Science and Technology of China, Chengdu, China [email protected] Beijing University of Posts and Communications, Beijing, China [email protected] classifiers. The classifier design first needs to figure out the object and how it is represented. In general, multiple feature vectors are methods used to represent objects. In this way, the target can be regarded as a point in the multi-dimensional space, which is very beneficial for calculation and characterization. For example, when calculating the similarity or measuring the distance, the space midpoint can be used, or by using the main Methods such as composition analysis to reduce feature dimensions. The classification model based on machine learning has been developed for many years and there are many tools available [3]. Abstract—Face classification has been developed for many years with the aim of dividing the face image into the identity of the person to whom they belong. The existing classification methods still have some problems such as the classifier is difficult to conform to the multi-level graph model and the data in the real scene rarely has a valid label. In this paper, a semi-supervised structured sparse graph data classification method for face feature extraction is proposed. The method of transforming graph data into uniform length vector is used to construct a semi-supervised method using the manifold that preserves the global data structure and local data structure. Sparse graph data classifier was utilized in which sparse learning automatically obtains the connection relationship between the label data vector and the unlabeled data vector and its weight with the L2 norm to control the model. Finally it realizes the sparse representation of the semi-supervised data. Experiments based on standard face datasets show that this semi-supervised sparse graph data classification method has better classification performance. Kernel classifiers in machine learning have many advantages, such as the ability to process noisy data well. However, the mainstream supervised method in machine learning, whether the model is effective or not depends mainly on the number of data and whether the label is good, and the corresponding unsupervised method of machine learning has been paid attention to in the field of graph classification. The advantage of the unsupervised method is that the performance of the model is not completely dependent on the quantity of data and the quality of the annotation. Generally, the feature engineering pre-processes the data, and the data is first dimensioned and selected by prior knowledge, although the time consumption is relative to supervised methods are higher and more dependent on the scenario [4]. Keywords—Semi-supervised, structured sparse graph, face classification, manifold, L2 norm I. I NTRODUCTION In more and more practical applications, such as biological sequences, compounds, natural language texts, and semi-structured data, graphs are proposed and used as a modeling object. In this modeling approach, the target is represented as a graph, with a complex structure, rather than a traditional method of representing the target as a feature vector. Graph classification is a very important branch of the graph model and is widely used in the above applications. The challenge comes from the complex structure of the graph itself, and the lack of a model that is characterized by vectors. Machine learning algorithms are widely used as a primary means of graph classification. In the past machine learning algorithms, numerical vectors were used instead of graphs to represent specified targets. In the classification of graphs, more nuclear methods, such as SVM, are used. The assumption of using such a kernel method is that there may be some structural types in the feature vector extracted by the target. The purpose is to generate a classification hyperplane, and the extracted target features can be mapped into high- dimensional space, the problem is that such a kernel function is difficult to explicitly calculate the correlation and specific characteristics between structures [1, 2]. The semi-supervised classification problem is used to continue classification of nodes that are not labeled, and is used in directed graphs with positive and negative weighting, such as web page classification and document classification and biology. The graph nodes in the web page classification are used to represent the corresponding pages. The relationship between the pages is represented by the directed edges. At the same time, the graph connectivity is used when classifying the nodes without annotations. The prior knowledge required for this method is the graph. The mapping of each node's association and classification in the real scene, the typical a priori assumption is that the edge has a certain influence on the node category. It includes two structures: co-links with directed graphs and directed graphs with the same sub-nodes have been widely used in practical applications such as web page classification. The classifier design for the graph node classification problem can adopt the support vector machine algorithm. The problem is that this kind of kernel algorithm has no view global and only pays attention to the relationship between nodes. Further consideration is given to the global and localized techniques of directed graphs [5]. The machine learning classification models are derived from a large number of data training, using machine learning classification rules in training, Bayesian classification and support vector machines are typical machine learning 978-1-7281-4691-1/19/$31.00 ©2019 IEEE DOI 10.1109/AIAM48774.2019.00027 In the graph structure of the graph classifier method, the 100