CN-121705850-B - Dynamic graph disambiguation-based partial multi-label data classification method and system

CN121705850BCN 121705850 BCN121705850 BCN 121705850BCN-121705850-B

Abstract

The invention relates to the technical field of data processing, in particular to a dynamic graph disambiguation-based partial multi-label data classification method and a dynamic graph disambiguation-based partial multi-label data classification system, which are characterized in that a triple reliable label set is constructed according to output, confidence and consistency, and the weighted similarity between dependency construction examples of candidate partial multi-labels is based on the occurrence frequency of the candidate partial multi-labels in the triple reliable label set, the discrimination score of the candidate partial multi-labels, and (3) suppressing structural noise by using weighted similarity, constructing a multi-scale similarity graph and adaptively fusing, introducing a graph propagation mechanism for dynamically adjusting the label confidence, and realizing the co-evolution of the classifier and the label quality through progressive multi-stage training. The method effectively solves the problems of feature-semantic ambiguity, structural noise interference, unbalanced marking contribution, single-scale map limitation and the like.

Inventors

KONG XIANGJUN
XIAO YANSHAN
LI ZIANG
CHEN HUI
CHEN XIAODONG
CHEN YU

Assignees

广东工业大学

Dates

Publication Date: 20260512
Application Date: 20260212

Claims (10)

1. The method for classifying the partial multi-label data based on dynamic graph disambiguation is characterized by comprising the following steps of: S1, acquiring a partial multi-mark data set for training, and acquiring an example feature matrix and a candidate partial multi-mark matrix based on the partial multi-mark data set, wherein the partial multi-mark data set is derived from a natural scene image or a medical image; S2, pre-training an initial classifier based on the example feature matrix and the candidate partial multi-label matrix; S3, obtaining an initial prediction model based on the initial classifier and a first activation function; S4, carrying out repeated iterative training on the initial prediction model, purifying the candidate partial multi-label matrix in the training process to obtain a trained optimal prediction model, wherein the method at least comprises the following substeps: Performing repeated iterative training on the initial prediction model, constructing a triple reliable mark set based on a t-1 round prediction model according to output, confidence and consistency in a t round iteration, obtaining dynamic rarity weight of the candidate partial multiple marks based on the occurrence frequency of the candidate partial multiple marks in the triple reliable mark set and the discrimination score of the candidate partial multiple marks, constructing weighted similarity among examples based on the triple reliable mark set, the dynamic rarity weight and the dependence among examples, constructing a multi-scale fusion graph based on the weighted similarity, obtaining a purified soft mark matrix based on the multi-scale fusion graph, utilizing the purified soft mark matrix as the output of the t round iteration, and adjusting parameters of the t round prediction model to obtain the t round prediction model; S5, obtaining an instance to be detected, wherein the instance to be detected is derived from a natural scene image or a medical image, and classifying the instance to be detected based on the optimal prediction model.
2. The method according to claim 1, wherein in step S1, the feature vector of each instance in the instance feature matrix maps the candidate partial marker vector of the corresponding position in the candidate partial marker matrix, respectively, and the candidate partial marker matrix includes a true marker and an erroneous non-precise marker.
3. The method for classifying partial multi-label data based on dynamic graph disambiguation according to claim 1, wherein in step S4, said constructing a triple reliable label set based on the t-1 round prediction model according to output, confidence, consistency comprises at least the sub-steps of: Based on the t-1 round of prediction model, taking a plurality of candidate partial labels with highest output scores in the t-1 round of prediction model as a local relative reliable set; Based on the t-1 round of prediction model, taking a plurality of candidate partial labels with highest confidence scores in the t-1 round of prediction model as a global high confidence anchor set; based on the t-1 round of prediction model, taking a plurality of candidate partial labels with highest consistency scores in the t-1 round of prediction model as a label consistency set; And obtaining a triple reliable mark set from the union of the local relative reliable set, the global high-confidence anchor set and the mark consistency set.
4. The method of classifying partial multiple label data based on dynamic graph disambiguation according to claim 1, wherein in step S4, the discriminative score of the candidate partial multiple label is obtained based on a classifier instance feature importance analysis.
5. The method of classifying partial multi-label data based on dynamic graph disambiguation according to claim 1, wherein in step S4, the obtaining of dependencies between instances comprises at least the sub-steps of: Obtaining a mark co-occurrence matrix based on the triple reliable mark set and the occurrence frequency of the candidate partial marks in the triple reliable mark set, wherein the mark co-occurrence matrix is used for capturing the association strength between every two candidate partial marks; and when two instances share the candidate partial mark j, if the two instances also share other candidate partial marks l except the candidate partial mark j, the larger the dependence of the two instances.
6. The method for classifying partial multi-label data based on dynamic graph disambiguation according to claim 1, wherein in step S4, said constructing a multi-scale fusion graph based on said weighted similarity comprises at least the sub-steps of: constructing a local fine graph by using a proximity algorithm based on the weighted similarity; constructing a global manifold graph by utilizing an optimal transmission theory based on the weighted similarity; and constructing a multi-scale fusion graph based on the local fine graph and the global manifold graph.
7. The method for classifying partial multi-label data based on dynamic graph disambiguation of claim 6, wherein in step S4, said constructing a global manifold graph using optimal transport theory based on said weighted similarity comprises at least the sub-steps of: obtaining a cost matrix based on the weighted similarity; obtaining dynamic supply distribution based on the t-1 round prediction model, the candidate partial multiple marker matrix, the triple reliable marker set and the dynamic rarity weight; Obtaining dynamic demand distribution based on the t-1 round prediction model, the candidate partial multiple marker matrix, the triple reliable marker set and the dynamic rarity weight; And constructing a global manifold graph by utilizing entropy regularization optimal transmission based on the cost matrix, the dynamic supply distribution and the dynamic demand distribution.
8. The method of classifying partial multi-label data based on dynamic graph disambiguation according to claim 1, wherein in step S4, the obtaining of a clean soft label matrix based on the multi-scale fusion graph comprises at least the sub-steps of: Obtaining a confidence coefficient adjustment factor based on the t-1 round prediction model, a confidence coefficient nonlinear adjustment parameter, a reliable marker enhancement coefficient and the triple reliable marker set; And obtaining a purified soft marking matrix based on the multi-scale fusion map, the t-1 round prediction model, the confidence coefficient adjustment factor and the candidate partial multi-marking matrix.
9. The method for classifying partial multi-label data based on dynamic graph disambiguation according to claim 1, wherein in step S4, parameters of a t-th-round prediction model are adjusted by using the clean soft label matrix as an output of a t-th-round iteration, so as to obtain a t-round prediction model, and the method at least comprises the following sub-steps: the purification soft mark matrix is used as the output of the t-th round iteration, and the parameters of the t-th round prediction model are adjusted by adopting a gradient descent method to obtain a t-round fine adjustment prediction model; and training by utilizing binary cross entropy loss of the purification soft marking matrix and the purification hard marking matrix based on a t-round fine tuning prediction model to obtain a t-round prediction model.
10. A dynamic graph disambiguation-based meta-tagging data classification system for implementing the dynamic graph disambiguation-based meta-tagging data classification method of any one of claims 1 to 9, comprising: the data processing unit is used for acquiring an instance feature matrix based on the partial multi-label data set; an initial classifier construction unit for pre-training an initial classifier based on the example feature matrix and the candidate partial multiple marker matrix; The initial prediction model construction unit is used for obtaining an initial prediction model based on the initial classifier and a first activation function; The model iterative training unit is used for carrying out repeated iterative training on the initial prediction model, purifying the candidate partial multi-label matrix in the training process and obtaining a trained optimal prediction model; and the marking prediction unit is used for classifying the to-be-detected examples based on the optimal prediction model.

Description

Dynamic graph disambiguation-based partial multi-label data classification method and system Technical Field The invention relates to the technical field of data processing, in particular to a method and a system for classifying partial multi-label data based on dynamic graph disambiguation. Background In data processing tasks, multi-tag learning aims at identifying multiple semantic tags from a single sample. However, training data in real scenes often contain a lot of noise due to subjectivity, omission or errors of manual labeling, resulting in "biased-labeled" data, i.e., a set of candidate labels associated with each sample are intermixed with a partial error label. The efficient cleaning, correction and information extraction of incomplete and inaccurate data are key preprocessing links for improving the performance of a downstream classification model, and belong to the technical field of typical and challenging data processing. The existing partial multi-mark learning method usually adopts a data processing framework based on a graph, and filtering of noise marks and recovery of real marks are expected to be realized by constructing a similarity graph among samples and transmitting mark confidence along edges. However, this type of method has a fundamental disadvantage in processing the data, in that the construction of the graph structure and subsequent similarity calculations are often dominated by marks that occur at high frequencies (but may be of low discrimination and even structured noise). During data processing, rare marks that occur less frequently, but may have higher category discrimination, are severely impaired in their contribution. This results in a substantial bias of the constructed graph structure towards the correlation patterns reflecting common markers, rather than the true semantic structure of the data, which in turn makes systematic deviations of the marker purification from the results of confidence estimation, limiting the upper precision limit of the final classification model. Therefore, how to effectively suppress the interference of high-frequency noise and enhance the influence of rare discrimination marks when data is marked more often becomes a key technical problem to be solved in the field. Disclosure of Invention The invention aims to overcome the defects that high-frequency noise interference is difficult to effectively inhibit and rare discrimination marking influence cannot be fully enhanced when partial multi-marking is carried out on data in the prior art, and provides a partial multi-marking data classification method and system based on dynamic graph disambiguation, which can effectively inhibit high-frequency noise interference and enhance rare discrimination marking influence when partial multi-marking is carried out on data. According to an aspect of the embodiment of the present invention, there is provided a method for classifying partial multi-label data based on dynamic graph disambiguation, the method comprising the steps of: S1, acquiring a partial multi-mark data set for training, and acquiring an example feature matrix and a candidate partial multi-mark matrix based on the partial multi-mark data set, wherein the partial multi-mark data set is derived from a natural scene image or a medical image; S2, pre-training an initial classifier based on the example feature matrix and the candidate partial multi-label matrix; S3, obtaining an initial prediction model based on the initial classifier and a first activation function; S4, carrying out repeated iterative training on the initial prediction model, purifying the candidate partial multi-label matrix in the training process to obtain a trained optimal prediction model, wherein the method at least comprises the following substeps: Performing repeated iterative training on the initial prediction model, constructing a triple reliable mark set based on a t-1 round prediction model according to output, confidence and consistency in a t round iteration, obtaining dynamic rarity weight of the candidate partial multiple marks based on the occurrence frequency of the candidate partial multiple marks in the triple reliable mark set and the discrimination score of the candidate partial multiple marks, constructing weighted similarity among examples based on the triple reliable mark set, the dynamic rarity weight and the dependence among examples, constructing a multi-scale fusion graph based on the weighted similarity, obtaining a purified soft mark matrix based on the multi-scale fusion graph, utilizing the purified soft mark matrix as the output of the t round iteration, and adjusting parameters of the t round prediction model to obtain the t round prediction model; S5, obtaining an instance to be detected, wherein the instance to be detected is derived from a natural scene image or a medical image, and classifying the instance to be detected based on the optimal prediction model. In an optional manner, in