Multi-Label Learning Based On Causal Inference

In recent years,with the development of computing hardware and the growth of labeled data,the AI(Artificial Intelligence)related research fields has ushered in another era of rapid development.With the rapid growth of data volume,how to effectively obtain valuable information from large volume data has become one of the hot research interests of current research.To deal with the high-dimensionality of multi-label learning,scholars respectively proposed label correlation and label-specific features.In the real-world scenario,labels are often correlated.For example,"smoking" and "lung cancer" are strongly correlated.Label correlation constrains the model based on the assumption that labels with higher correlation are likely to appear in the same instance.Label correlation improves the performance of multi-label learning to a certain extent.In addition,traditional multi-label learning algorithms usually predict different labels with the same set of features,while labels in the same instance can have different features,such as "flower" and "leaves".Label-specific features algorithms can extract the unique features of each label(label-specific features)which further improve the performance of multi-label learning algorithms.Most multi-label learning algorithms that consider label correlation and label-specific features assume that the correlation is symmetric.However,the correlation is usually asymmetric in the real-world data.To tackle this problem,this thesis proposes multi-label learning algorithms with causal inference.The main works of this thesis are as follows.(1)Most existing multi-label learning algorithms that consider label correlation assume that the correlation between the labels is symmetric,which ignores the asymmetric correlation between the labels.Only considering the symmetrical label correlation will introduce redundant information in the model.Results in performance drop in multi-label classifiers.Aiming at this problem,this thesis proposes a multi-label learning algorithm based on causal inference,which models the asymmetric correlation between labels by jointly learning label correlation and label causality to eliminate redundant information.The results of comparison experiments show that this method has a certain advantage.(2)Most present label-specific learning algorithms extract label-specific features by measuring Euclidean distance or constrain the coefficient matrix with the 7)-norm.However,with the increase of the data dimension,Euclidean distance becomes unreliable for measure the distance between instances in high-dimensional space.At the same time,extracting label-specific features through the 7)-norm relies on manual parameter selection.Either too high or too low sparsity will lead to poor classifier performance.In order to solve the above problems,this thesis proposes to obtain label-specific features by considering the causality between features and labels.The results of comparative experiments verify the effectiveness of the proposed method.(3)Causal inference algorithms are sensitive to the distribution of the data.Most causal inference algorithms are only applied to infer the causation between continuous data or causation between discrete data.However,in multi-label datasets,the features are usually continuous data,and the labels are usually discrete data.The distribution difference between the two sets is significant.Directly applies causal inference algorithms on such data results in poor performance.This thesis proposes to use label enhancement to process the labels to obtain continuous labels and perform causal inference based on continuous labels.The experimental results show the effectiveness of this method.