Techniques And Applications For Hierarchical Multi-label Classification

Hierarchical Multi-label Classification(HMC)aims to assign the most relevant multiple labels to entity data,which are organized and stored in a hierarchical structure(tree or DAG).Hierarchical Multi-label Classification provides a way to view the data from multiple perspectives and has a wide range of applications in real-world scenarios,such as text classification,image classification,functional genomics,etc.With the further integration of information technology and human production life,massive amounts of data in various forms are being generated.The constantly updated iterative data mining technologies and artificial intelligence methods bring new opportunities and challenges to HMC.Nowadays,despite the achievements of HMC,it still has a chance to further to further improve the the utility of HMC in practical scenarios when taking a deep insight into three research objects including hierarchy,entity data and classifier.However,there are three major challenges we need to address.First,for the hierarchy,the complex class dependencies of hierarchy leads to difficulties in understanding hierarchical structures.Second,for the entity data,the data modality is not single leading to difficulties in fusion between modal information.Finally,regarding the classifier,the prediction results produced by the models are incoherent,leading to a lack of model interpretability.To remedy these issues,this dissertation systematically conducts a series of exploratory research on methods and applications for the Hierarchical Multi-label Classification task.Specifically,for the hierarchy,we propose an attention-based recurrent network whose effects have been validated for relevant applications in the text domain.For the entity data,we propose a hierarchical multimodal method that takes into account the possible multimodal information of entity data.For the classifier,we propose a general hierarchical multi-label classification framework that combines path evaluation and selection,which improves the performance and interpretability of the classification.In particular,all the works rely on the leading online learning platform system "Zhixue"developed by iFLYTEK Co.,Ltd.,and the intelligent education platform "LUNA" developed by the USTC BDAA lab.Both the research issue and data are derived from real application scenarios,and all solutions are validated in real platforms that have practical application value.In addition,the relevant code has been open sourced on GitHub code site,accumulated more than one thousand stars,widely praised by the industrial open source community.The main work and contributions of this dissertation can be summarized as follows.First,for the hierarchy,we study the hierarchy association representation and modeling methods.Hierarchy representation and modeling hierarchy is the foundation of HMC tasks.Understanding and representing hierarchy information,and capturing class dependencies(e.g.,parent-child class relationships)in the hierarchy,can effectively help assign the most relevant label for each entity data.To this end,we propose a hierarchical attention-based mechanism for the HMC model,called Hierarchical Attentionbased Recurrent Neural Network(HARNN).Specifically,we first obtain a unified representation of text and hierarchy.Then,we design a Hierarchical Attention-based Memory unit to model the dependencies between different levels by capturing the associations between text and hierarchy in a top-down manner.Finally,we design a hybrid approach to combine the local predictions and global predictions of hierarchy.Extensive experiments show that HARNN is more accurate and stable than traditional methods for HMC in the text domain.Second,for the entity data,we study the data modality fusion and unification methods.In many applications in real scenarios,data collected by different means have multiple forms,for example,video data itself has multimodal information(image,text,audio,etc.),while traditional HMC methods mainly deal with unimodal data.Therefore,we propose an HMC model that incorporates multimodality,called Hierarchical Multi-modal Network(HMNet).In the educational video concept prediction task,we design a video keyframe extraction algorithm to divide the video into a series of video sections with corresponding subtitle text.Then,we use a multimodal encoder to obtain the unified representation for multimodality.Finally,we design a hierarchical predictor capable of fusing the multimodality representation,modeling the class dependencies,and predicting the hierarchical concepts of the video in a top-down manner.Extensive experiments verify that HMNet can effectively represent and fuse the different modalities of educational videos with good classification performance.Finally,for the classifier,we study the methods for evaluating and selecting prediction paths.Traditional HMC methods often produce incoherent prediction paths that do not satisfy the hierarchy constraints during prediction,and such prediction results lack interpretability and realistic meaning.To this end,we propose an HMC model that combines the evaluation and selection of prediction paths,called Hierarchical multilabel classification Network(HmcNet).Specifically,we first introduce the explicit and implicit hierarchy constraints,and obtain the unified representation of entity data and hierarchy.Then,we model the class dependencies between different levels by capturing the entity-hierarchy associations in a top-down manner.In this way,we ensure that the generated prediction results are as coherent as possible and satisfy the hierarchy constraints.Finally,we design an efficient Prune-based Coherent Prediction(PCP)strategy to select the optimal path.In the classification phase,HmcNet evaluates all possible prediction paths,selects the paths above the given threshold,and generates the final prediction results coherently.Extensive experiments fully validate that the classification performance of HmcNet on HMC tasks is higher than that of traditional methods,and has robust scalability and interpretability.