Noisy multi-label classification method based on category feature construction

Abstract

The present invention proposes a noisy multi-label classification method based on category feature construction, which is suitable for multi-label data classification scenarios where the label space contains noise. The method includes the following steps: (1) the user collects multi-label data under the premise of ensuring that the true labels are completely obtained; (2) the noisy labels are disambiguated using the internal structure information of the feature space; (3) based on the disambiguated data set, a category feature map is constructed for each label; (4) a binary classification data set is constructed based on the category feature map of each label, and a binary classification model is trained; (5) the trained classification model is used to predict unseen examples to obtain the classification results for each category; (6) if the user is satisfied with the prediction results, the method ends; otherwise, the method goes to step (2) and the original labels are disambiguated again.

Description

Technical Field

[0001] The present invention relates to a multi-label data classification method, in particular to a noise label-oriented multi-label classification method, belonging to the technical field of weakly supervised classification. Background Art

[0002] In multi-label learning, an example is associated with multiple correct semantic labels. The goal of the learning system is to use the multi-label dataset to construct a mapping from the input space (feature space) to the output space (label space) to predict a set of correct labels for unseen examples. The most critical assumption in traditional multi-label learning tasks is that each example in the multi-label dataset is assigned its correct set of semantic labels. However, in real life, accurately assigning labels to each example is difficult. In many tasks, labeling errors, missing labels, or redundant labels may occur, resulting in noisy label information in the multi-label dataset. Frameworks that study such multi-label datasets are collectively referred to as noisy multi-label learning frameworks.

[0003] This paper studies noisy multi-label learning tasks in which labels are redundant. Each example is associated with a set of candidate labels, and the true label of the example is unknown but hidden in the candidate label set. The goal of the learning system is to learn the mapping relationship between the input space and the output space under known weak supervision information, which is used to predict the correct set of labels for unseen examples. To address the problem of noisy datasets, disambiguation methods are often used to first disambiguate the dataset before training the multi-label classification model. Such methods focus on how to disambiguate the dataset, which directly affects the classification performance of the learning model.

[0004] Existing disambiguation methods often use methods such as label correlation or meta-learning to disambiguate noisy datasets. These methods focus on the information in the output space (label space) rather than the input space (feature space). Summary of the Invention

[0005] To solve the above problems, the present invention discloses a noisy multi-label classification method based on category feature construction.

[0006] The problem of noisy labels in candidate labels is converted into the problem of noisy samples in the set of positive samples corresponding to the labels. The categorical features of the labels are then used to disambiguate the samples. The categorical features are then used to build a prediction model on the disambiguated dataset, improving the performance of the multi-label classification model.

[0007] The present invention focuses on the feature information of the input space, constructs the internal structure information of the corresponding feature for each category, and uses the internal structure information of the feature to perform category correction on samples with low confidence values, thereby completing the disambiguation operation of the noisy dataset. The method includes the following steps:

[0008] (1) The user collects multi-label data (which will inevitably contain noise) under the premise of ensuring that the true labels are fully obtained; (2) The noisy labels are disambiguated using the internal structural information of the feature space; (3) Based on the disambiguated dataset, a category feature map is constructed for each label; (4) A binary classification dataset is constructed based on the category feature map of each label, and a binary classification model is trained; (5) The trained classification model is used to predict unseen examples to obtain the classification results for each category; (6) If the user is satisfied with the prediction results, the process ends; otherwise, go to step (2) and disambiguate the original labels again.

[0009] The present invention is further preferred that the step (2) utilizes the internal structural information of the feature space to disambiguate the noisy labels, specifically comprising:

[0010] Assumptions represents the d-dimensional example space, Y={y1,y2,…,y q} represents a label space with q categories; given a noisy multi-label dataset D = {(x1, S1), (x2, S2), …, (x m ,S m )}, containing m examples, where is a d-dimensional attribute vector, and For example x i A corresponding set of candidate markers; in addition, Representation and example x i A corresponding set of non-candidate markers; Y i For example x i The set of true labels;

[0011] For each mark y k ∈Y, execute the following loop body:

[0012] 1) Divide the samples into positive sample sets and negative sample set

[0013]

[0014] 2) Use the k-means algorithm to and Clustered into m k clusters; the cluster centers of the positive and negative sample sets are defined as and

[0015]

[0016] 3) Train a binary classification model to obtain the confidence ranking of the samples, and select a certain number of samples with smaller confidence values to form the pseudo-positive sample subset U to be determined k ,

[0017] 4) Calculate sample x i (x i ∈U k ) to the distance between the positive and negative sample cluster centers to determine whether the sample is a noise sample, and if so, correct the sample category; Represents example x i With m k The average distance between the cluster centers of the positive samples, Represents example x i The average distance from the cluster center of each negative sample;

[0018]

[0019] The present invention is further preferred that step (3) constructs a category feature map for each tag based on the disambiguated dataset, specifically comprising:

[0020] (3.1) According to step (2)

[0021] Get the disambiguated dataset D′={(x1,S′1),(x2,S′2),…,(x m ,S′ m )};

[0022] (3.2) For each label y on the dataset D′ k ∈Y calculates its positive and negative sample sets;

[0023]

[0024] (3.3) Using the k-means algorithm, and Clustered into m k clusters. The cluster centers of the positive and negative sample sets are defined as and

[0025] (3.4) Calculate the sample and 2m k The Euclidean distance of the cluster centers maps the d-dimensional attributes of the original sample to 2m k Dimension, thus getting the label y k The class feature map φ k :

[0026]

[0027] The present invention is further preferred, wherein the step (4) constructs a binary classification data set based on the category feature map of each marker and trains a binary classification model, specifically comprising:

[0028] For each mark y k ∈Y, according to the mapping relationship φ in step (3) k Construct the dataset D′ into its corresponding binary classification training set D′ k , where when · is true, The value is 1, otherwise -1;

[0029]

[0030] In the binary classification dataset D′ k Above, marked y k Train a two-class classification model h k Similarly, for each tag in the tag space Y, a two-class classification model {h1,h2,...h q}.

[0031] Beneficial effects of the present invention:

[0032] 1. The method consists of two phases. In the first phase, a set of positive and negative samples is constructed for each label. The cluster centers of these sets are calculated to represent the internal structural information of the features corresponding to the label. This internal structural information is then used to perform category correction on samples with low confidence values. In the second phase, a category feature map is constructed for each label, the dataset is converted into a binary classification dataset corresponding to the label, and a binary classification model is learned. For unknown examples, category feature maps are first performed, and then the classification model is used to predict their label set. Learning and training based on this method can improve the effectiveness, stability, and robustness of multi-label classification models.

[0033] 2. In existing noisy multi-label classification methods, researchers often use label correlation information in the output space to disambiguate noisy labels, while ignoring the feature information in the input space. This paper disambiguates noisy datasets by obtaining the internal structure information of features through clustering, generating more representative feature information for each category, and to some extent reducing the impact of noisy labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a workflow diagram of a document classification device;

[0035] Figure 2 is a flow chart of the method of the present invention;

[0036] Figure 3 is a flowchart of the candidate token disambiguation phase;

[0037] Figure 4 It is a flowchart of the prediction model building phase. DETAILED DESCRIPTION

[0038] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inward" and "outward" refer to directions toward or away from the geometric center of a particular component, respectively.

[0039] like Figure 1 As shown, the process of document classification is taken as an example. The document storage device contains a large number of ambiguous subject documents, each of which is marked with multiple topics, of which only a small number of topics belong to the true subject of this document, and the number of true topics is unknown. When judging the true subject of a document, the user first selects N sample documents from all ambiguous documents and submits them to the document classifier (computer model) for processing. At this time, it is necessary to extract features from the documents to be classified. This step generally uses some feature dimension reduction techniques for Chinese text, such as term frequency (TF), term frequency-inverse document frequency (TF-IDF), etc. After extracting the features of the document, the document classifier proposed in the present invention disambiguates the ambiguous documents according to the internal structure information of the feature space corresponding to each topic in the first stage. In the second stage, based on the disambiguated subject document library, a category feature map is constructed for each topic, and a binary classification data set corresponding to the topic is constructed according to the category feature map of each topic, and a binary classification model is generated by training. After obtaining the final classification model, when there are documents with unknown topics that need to be predicted, the features are extracted in the same way, and then the classification model is used to predict and return the classification results. Figure 1 If the user is not satisfied with the result, he or she can select more sample documents from the document library and feed them back to the document classification device.

[0040] The method of the present invention is as follows Figure 2 As shown. Step 10 is the starting action. Assume that the sample document set D submitted by the user is D = {(x1, S1), (x2, S2), ..., (x m ,S m )}, where the sample is the d-dimensional eigenvector, and For the sample x i A corresponding set of candidate tags. i For sample x i The true label set, Y i Unknown and In step 11, the internal structure information of the feature space corresponding to each tag is used to disambiguate the ambiguous document set. Figure 3 In step 12, based on the disambiguated document set obtained in step 12, the category feature information of each tag is learned and a binary classification model is trained. Figure 4 In step 13, the prediction model obtained in step 12 is used to classify the unclassified documents in the document storage device and the classification results are returned. Finally, the process enters the end state of step 14.

[0041] Figure 3 A detailed description of step 11 is given, which uses the internal structure information of the feature space corresponding to each tag to disambiguate the ambiguous document set. Step 1100 is the starting action. Steps 1101 to 1113 form a loop, and in each round of the loop, the disambiguation operation is performed on the kth tag. Steps 1103 to 1112 form a loop, and in each round of the loop, a certain number of samples with low confidence are selected for category correction. Specifically, in step 1105, the tag y k The samples corresponding to ∈Y are divided into the positive sample set and negative sample set As shown below:

[0042] In step 1106 to step 1108, the k-means algorithm is used to respectively and Clustered into m k clusters, m k The calculation of is shown in formula (2). The cluster centers of the positive and negative sample sets are defined as and

[0043]

[0044] In step 1109, a binary classification model is trained to obtain the confidence ranking of the samples, and a certain number of samples with smaller confidence values are selected to form the pseudo-positive sample subset U to be determined. k , In step 1110, the samples x in the set of pseudo-positive samples to be determined are calculated respectively. i (x i ∈U k ) and the distance between the positive and negative sample cluster centers, as shown in formula (3). Represents example x i With m k The average distance between the cluster centers of the positive samples, Represents example x i The average distance from the cluster center of each negative sample.

[0045]

[0046] In step 1111, the distance calculated in step 1110 is used to determine whether the sample is a noise sample. If so, the sample category is corrected, as shown in formula (4).

[0047]

[0048] When all categories have completed the disambiguation operation, the process will enter the end state of step 1114.

[0049] Figure 4 A detailed description of step 12 is given. Based on the disambiguated document set D′={(x1, S′1),(x2, S′2),…,(x m ,S′ m )}, learn the category feature information for each tag and train the binary classification model. Steps 1201 to 1210 form a loop, and in each round of the loop, a binary classification model is constructed for the kth tag using its category feature. In step 1203, the tag y is k ∈Y calculates its positive and negative sample sets and As shown below

[0050]

[0051] In steps 1204 to 1206, the k-means algorithm is used to respectively and Clustered into m k clusters, m k The calculation of is shown in formula (2). The cluster centers of the positive and negative sample sets are defined as and In step 1207, the sample and 2m are calculated. k The Euclidean distance of the cluster centers maps the d-dimensional attributes of the original sample to 2m k Dimension, thus getting the label y k The class feature map φ k , as shown below:

[0052]

[0053] In step 1208, according to the mark y k The class feature map φ k Construct the document set D′ into its corresponding binary classification training set D′ k , as shown in formula (7), where when · is true, The value is 1, otherwise -1.

[0054]

[0055] In step 1209, based on the above two-classification dataset D′ k , is the marker y k Train a two-class classification model h k After q rounds of iteration, each tag in the tag space Y is trained to obtain a two-class classification model to form the final multi-tag classification model {h1,h2,...h q}, then enter the end state of step 1211.

[0056] This paper proposes a noisy multi-label classification method based on generic feature construction. The method is divided into two stages. In the first stage, a set of positive and negative samples is constructed for each label, and the cluster centers of the positive and negative sample sets are calculated to represent the internal structural information of the features corresponding to the label. The internal structural information of the features is then used to perform category correction on samples with low confidence values. In the second stage, a generic feature map is constructed for each label, the dataset is converted into a binary classification dataset corresponding to the label, and a binary classification model is learned. For unknown examples, generic feature mapping is first performed, and then the classification model is used to predict the label set. Learning and training based on this method can improve the effectiveness, stability, and robustness of multi-label classification models.

[0057] The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above-mentioned embodiment, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A noisy multi-label classification method based on categorical feature construction, characterized by: The method comprises the following steps: Step (1) The user collects multi-label data under the premise of ensuring that the true labels are fully obtained; Step (2) uses the internal structure information of the feature space to disambiguate the noisy labels; Step (3) constructs a category feature map for each token based on the disambiguated dataset; Step (4) constructing a binary classification dataset based on the class feature map of each tag and training a binary classification model; Step (5) predicts the unseen examples based on the learned classification model to obtain the classification results for each category; Step (6) If the user is satisfied with the prediction result, the process ends; otherwise, go to step (2) and perform disambiguation on the original tag again. The step (2) uses the internal structural information of the feature space to disambiguate the noisy labels, specifically including: Assumptions represents the d-dimensional example space, Y={y1,y2,…,y q } represents a label space with q categories; given a noisy multi-label dataset D = {(x1, S1), (x2, S2), …, (x m ,S m )}, containing m examples, where is a d-dimensional attribute vector, and For example x i A corresponding set of candidate markers; in addition, Representation and example x i A corresponding set of non-candidate markers; Y i For example x i The set of true labels; For each mark y k ∈Y, execute the following loop body: 1) Divide the samples into positive sample sets and negative sample set 2) Use the k-means algorithm to and Clustered into m k clusters; the cluster centers of the positive and negative sample sets are defined as and 3) Train a binary classification model to obtain the confidence ranking of the samples, and select a certain number of samples with smaller confidence values to form the pseudo-positive sample subset U to be determined k , 4) Calculate sample x i The distance from the positive and negative sample cluster centers is used to determine whether the sample is a noise sample, x i ∈U k ; If yes, correct the sample category; Represents example x i With m k The average distance between the cluster centers of the positive samples, Represents example x i The average distance from the cluster center of each negative sample; Step (3) constructs a category feature map for each tag based on the disambiguated dataset, specifically including: (3.1) According to step (2) Get the disambiguated dataset D′={(x1,S′1),(x2,S′2),…,(x m ,S′ m )}; (3.2) For each label y on the dataset D′ k ∈Y calculates its positive and negative sample sets; (3.3) Using the k-means algorithm, and Clustered into m k clusters; the cluster centers of the positive and negative sample sets are defined as and (3.4) Calculate the sample and 2m k The Euclidean distance of the cluster centers maps the d-dimensional attributes of the original sample to 2m k Dimension, thus getting the label y k The class feature map φ k : The step (4) constructs a binary classification dataset based on the category feature map of each tag and trains a binary classification model, specifically including: For each mark y k ∈Y, according to the mapping relationship φ in step (3) k Construct the dataset D′ into its corresponding binary classification training set D′ k , where when · is true, The value is 1, otherwise -1; In the binary classification dataset D′ k Above, marked y k Train a two-class classification model h k Similarly, for each tag in the tag space Y, a two-class classification model {h1,h2,...h q }.

Images