Research On Key Techniques Of Weakly-supervised Multi-label Learning

In the task of multi-label learning(MLL),each instance is annotated with multiple candidate labels.Generally,these annotated candidate labels are valid and complete.Thus,some traditional MLL methods can directly learn a multi-label classifier from such ideally annotated training data,and then make prediction for unseen instances.However,in real-world scenarios,such accurately labeled data is difficult to collect,especially for large-scale data sets,it is too time-consuming and labor-consuming to finish the task of annotating such a number of instances.Therefore,it is a good choice to collect large-scale training data by directly downloading data with semantic labels from the Internet.Unfortunately,the labels of these downloaded data are often noisy(no label,lack of fine-grained labels,missing labels,redundant labels or error labels).If we directly employ such ambiguous data to train the predictive model,the learned model will be biased.Under such background,weakly supervised multi-label learning has gradually become a research hotspot in the field of machine learning and computer vision.Weakly supervised multi-label learning is a general term for multi-label learning problems with imperfect or incomplete supervised information.Generally speaking,weak supervision can be roughly divided into three categories: the first is incomplete supervision,that is,only a few instances have label information,while a large number of instances are unlabeled;The second is not fine-grained labels,that is,the training examples only have coarse-grained annotation information,but lack fine-grained annotation information;The third is inaccurate supervision,that is,there are missing,redundant or wrong annotation information of training examples.In this paper,we focus on the case of redundant annotation information,and focus on two weakly supervised learning problems: partial label learning and partial multi-label learning.In partial label learning,only one label in the candidate label set is correct,while in partial multi-label learning,multiple candidate labels are vaild.Based on this,we conduct an in-depth research on the two above learning problems and propose six innovative algorithms as follows:1.Self-Paced Curriculum Learning based Partial Label Learning.This method simulates the human learning model “from easy to hard”,and introduces the curriculum learning and self-paced learning schemes to divide the difficulty degree of the training data,so as to add training data to learning model from easy to hard and guide the model from “naive” to “mature”.During the process of model learning,curriculum learning simulates human teachers to plan the course learning order in advance,and adjusts the difficulty order of training data through predefined knowledge;The self-paced learning simulates the autonomic learning of human students,and dynamically adjusts the followup learning tasks through the results of autonomic learning.The combination of such two strategies is analogous to “teacher-student collaborative” learning mode,which can not only flexibly utilize the prior curriculum knowledge to guide the model learning,but also avoid the inconsistency between the prior curriculum and the dynamically learned model,which leads the algorithm to achieve better learning performance.We have conducted enormous comparative experiments and ablation experiments,and the experimental results show that such “teacher-student collaborative” learning strategy can effectively improve the learning performance of partial label learning method.2.Graph Matching based Partial Label Learning.This method interprets the correspondences between instances and labels in partial label learning as “instance-label”matching,and for the first time transforms the task of partial label learning as “instancelabel” matching selection problem.By employing the graph matching model to explore instance similarity,instance dissimilarity and “instance-label” matching consistency,this method guides the learning model to finish accurate “instance-label” matching selection.During the process of model construction,since the traditional “one-to-one” graph matching algorithm can not meet the scenario that multiple instances may correspond to the same label in the partial label learning problem,we extend the traditional “oneto-one” probability matching algorithm to “many-to-one” constraint to adapt it to the natural characteristic of partial label learning problem.In addition,we also propose an“instance-label” matching prediction model,which assigns candidate labels to unseen instances through weighted k-nearest-neighbor reconstruction,and then employs graph matching strategy to obtain real “instance-label” matching.We have conducted enormous comparative experiments,and the experimental results show that our algorithm has stronger disambiguation ability against state-of-the-art methods.3.Deep Graph Matching based Partial Label Learning.This method overcomes two shortcomings of the traditional graph matching based partial labeling learning method:first,when model measures the instance and labeling relationship,the “instance-label”relationship is usually incorporated into the learning framework as a fixed prior knowledge,rather than adaptive learning;Second,the “instance-label” matching consistency relationship in the traditional graph matching architecture has high complexity,which decreases its efficiency in learning from large-scale data sets.During the process of model training,all instances and labels are constructed into instance graph and label graph respectively,and then each instance is connected to its candidate labels to integrate the two graphs into a unified matching graph.Then,the graph attention mechanism is employed to aggregate and update all node states on the instance graph,so as to excavate the structured representation of each instance.Finally,each candidate label is embedded into its corresponding instance,and the matching affinity of each “instance-label” is calculated by progressive cross-entropy loss.Enormous experimental results show that our algorithm achieves superior performance against state-of-the-art methods.4.Partial Label Learning by Combining Heterogeneous Loss with Sparse and LowRank Regularization.This method incorporates the advantages of pairwise ranking loss and pointwise reconstruction loss into a unified framework,where the constructed heterogeneous loss integrates the advantages of the two loss functions,so that the model can not only focus on the mapping relationship between feature space and label space,but also take the differential relationship between different labels,so as to provide richer label ranking and mapping reconstruction information for label disambiguation.During the process of model construction,we decompose the observed label matrix into a groundtruth label matrix and a noisy label matrix,then introduce the low-rank sparse mechanism to constrain the decomposed matrix to be sparse and low-rank,respectively.In our model,sparse ground-truth label matrix and low-rank noisy label matrix can accurately reflect the global label correlation of partial label data,and guide the model to achieve better ability of label disambiguation.We have conducted enormous comparative experiments and ablation experiments,the experimental results show that our proposed heterogeneous loss function and low-rank sparse constraint have achieved superior performance in solving partial label learning problems.5.Noisy Label Tolerance for Partial Multi-Label Learning.Considering that existing disambiguation-based partial multi-label methods may conduct unreliable disambiguation operation or the disambiguated labels may still incorrect,this method surpasses the traditional label disambiguation based model construction,and instead assumes that the label space of partial multi-label data is ideal and complete,while the feature information is missing.Then,the task of partial multi-label learning can be naturally transformed into a feature completion problem.During the process of model construction,we introduce a missing feature matrix and embed it into the observed feature matrix to obtain a ideal feature matrix with complete feature information.Then,we construct the mapping relationship between the completed feature matrix and the given label matrix to induce the desired multi-label classification model.Enormous experimental results show that our employed feature completion strategy can achieve desirable performance when dealing with partial multi-label learning problem.6.Prior Knowledge Regularized Self-Representation Model for Partial Multi-Label Learning.This method integrates the self-representation and prior label knowledge into a unified framework to solve the common problem that existing partial multi-label learning methods do not consider the noisy information reduction in feature space and valuable information exploration in label space,so as to learn a more discriminative feature representation for further label disambiguation.During the process of model construction,we first employ the low-rank self-representation model to learn the latent high-order relationship between different instances,then we introduce the prior label information as the feature complements to make the learned self-representation matrix be more discriminative.The core of this method is to take advantage of the data membership preference,which is derived from the prior label knowledge,to purify the discovered membership of the data and accordingly obtain more representative feature subspace for model induction.Enormous experiments on both synthetic and real-world data sets show that our proposed method can achieve superior performance than state-of-the-art methods.