CN110598733A - Multi-label distance measurement learning method based on interactive modeling - Google Patents

Multi-label distance measurement learning method based on interactive modeling Download PDF

Info

Publication number
CN110598733A
CN110598733A CN201910715600.9A CN201910715600A CN110598733A CN 110598733 A CN110598733 A CN 110598733A CN 201910715600 A CN201910715600 A CN 201910715600A CN 110598733 A CN110598733 A CN 110598733A
Authority
CN
China
Prior art keywords
label
distance measurement
learning
matrix
samples
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201910715600.9A
Other languages
Chinese (zh)
Inventor
不公告发明人
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing Zhigu Artificial Intelligence Research Institute Co Ltd
Original Assignee
Nanjing Zhigu Artificial Intelligence Research Institute Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing Zhigu Artificial Intelligence Research Institute Co Ltd filed Critical Nanjing Zhigu Artificial Intelligence Research Institute Co Ltd
Priority to CN201910715600.9A priority Critical patent/CN110598733A/en
Priority to PCT/CN2019/100001 priority patent/WO2021022571A1/en
Publication of CN110598733A publication Critical patent/CN110598733A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a multi-label distance measurement learning method based on interactive modeling, which comprises the steps of extracting training data from any multi-label application scene and carrying out multi-label labeling on the training data; the extracted training samples are preprocessed, so that the quality of the samples is improved; representing a distance metric matrix to be learned as a combined distance metric form; defining multi-marker semantic similarity based on feature and marker collaborative computation, and constructing a triple constraint set; combining the combined distance measurement and the triplet constraint set to construct a multi-mark distance measurement learning model, and carrying out optimization solution on the multi-mark distance measurement learning model; after learning distance measurement, mapping training data to a distance measurement space, and then learning by using the existing multi-label learning algorithm to obtain a multi-label classifier based on distance measurement learning; and inputting the sample to be predicted into the classifier to obtain a labeled sample. The invention can greatly reduce the time complexity of the multi-label learning system and increase the practicability of the multi-label learning framework.

Description

Multi-label distance measurement learning method based on interactive modeling
Technical Field
The invention relates to a multi-label distance measurement learning method based on interactive modeling, in particular to a multi-label distance measurement learning method based on interactive modeling of a feature space and a label space, which is suitable for any multi-label learning scene and belongs to the technical field of machine learning.
Background
In recent years, multi-label learning has received much attention from researchers and has resulted in a great deal of research effort. However, due to the combinability of the label space, the multi-label learning has high complexity and is difficult to be applied to the actual scene. The existing multi-label learning method mostly models the correlation among labels from a label space, and has less processing on a feature space. The multi-label data are analyzed, redundancy exists in the feature space, therefore, how to construct a proper multi-label distance measurement representation can greatly improve the learning performance of the multi-label system, and meanwhile, the combination of the distance measurement and the neighbor strategy reduces the complexity of the learning system, so that the practical application of multi-label learning is promoted. In view of the above problems, it is necessary to further study the learning method of the present invention.
Disclosure of Invention
The purpose of the invention is as follows: in order to solve the defects of the prior art, the invention aims to provide a multi-label distance measurement learning method based on interactive modeling, considering the correlation between labels and the structural interaction between the characteristic space and the label space from the aspect of characteristic space processing.
The invention adopts the following technical scheme:
the invention discloses a multi-mark distance measurement learning method based on interactive modeling, which comprises the following steps:
(1) extracting training data from any multi-label application scene, and performing multi-label labeling on the training data;
(2) preprocessing the extracted training samples, filtering and marking the samples with the occupancy rate smaller than a set threshold value, and improving the quality of the samples;
(3) based on a Mahalanobis distance measurement learning framework, considering the structural interaction between a feature space and a mark space in multi-mark data, and expressing a distance measurement matrix to be learned as a combined distance measurement form;
(4) defining multi-marker semantic similarity based on feature and marker collaborative computation, and constructing a triple constraint set;
(5) integrating the steps (3) and (4) to obtain a multi-mark distance measurement learning model, and carrying out optimization solution on the multi-mark distance measurement learning model;
(6) after learning distance measurement, mapping training data to a distance measurement space, and then learning by using an existing multi-label learning algorithm (MLKNN), thereby obtaining a multi-label classifier based on distance measurement learning;
(7) inputting a sample to be predicted into the classifier to obtain a labeled sample;
(8) performing sampling inspection on the labeling result, and if the labeling result is qualified, finishing the sampling inspection; otherwise, returning to the step (1), and continuing to extract the sample for model adjustment and updating.
The application scenes in step (1) include, but are not limited to, images, texts, video scenes, and the like.
Preferably, the step (3) represents the distance metric to be learned as a combined metric form as follows:
wherein, bl∈RdIs a combined basis vector, wl0 or more is a non-negative combination coefficient (weight) to be learned; l is the interval index of the sum symbol, varying from 1 to K. K is the upper bound of the summation symbol and is a self-defined value, and the method constructs a component basis vector for each token. T denotes the transpose of the matrix.
More preferably, b is constructed in combination with discriminative information of the markers in consideration of the structured interactivity between the markers and the features, and specifically, for each marker, the first K feature vectors are obtained as local metric vectors using Principal Component Analysis (PCA), so that K × q b are total, and the combination coefficient w is a vector of dimension 1 × Kq.
Further preferably, the step (4) comprises the steps of:
(41) obtaining a semantic similarity matrix S by performing collaborative calculation on the identification information of the feature space and the mark space:
sij＝yi TGyj
where G＝(αA+(1-α)C)
where j corresponds to sample xjSubscript of (i) corresponds to sample xiA subscript of (a); a is a matrix of q × q, q represents the number of marks, and the value of each row and column in the matrix is represented by alhM is the upper bound of the summation symbol and represents m samples; a islhCorresponding with a mark ylAnd yhA, since the correlation between the markers tends to be asymmetriclh≠ahl；clhCosine similarity, alpha and balance coefficient of corresponding combined base vector, and relative contribution of balance characteristic space and mark space to semantic similarity; c is a matrix of q x q, q represents the number of marks, and the value of each row and column in the matrix is represented by ClhThe calculation formula of (a), T represents the transposition of the matrix;
(42) setting a threshold parameter theta of semantic similarity, and aiming at a training sample x according to the definition of a semantic similarity matrixiCan be divided into the following similarity sets ZiAnd dissimilar set
Zi＝{xj|sij≥θ,j≠i,1≤j≤m}
(43) Respectively finding K samples with the nearest Euclidean distance in the similar set and the dissimilar set to form a set KiAndthe triplet constraint set R is then expressed as follows:
still more preferably, the learning model of step (5) is as follows:
ζi≥0,wi≥0(1≤i≤m)
whereinRepresenting the embedding spatial distance of the two samples,indicating the dissimilarity of the two token vectors.w is the non-negative combination coefficient (weight) to be learned in step (3), C is the balance coefficient, m represents the number of samples, ζ is the relaxation variable representing the loss, and R is the triplet constraint set.
The invention has the beneficial effects that: compared with the prior art, the multi-label distance measurement learning method based on interactive modeling provided by the invention can be used for processing any multi-label learning scene by fully utilizing the correlation among labels and the interactivity between the feature space and the label space, and reducing the complexity of a multi-label learning system and promoting the practical application of a multi-label learning framework by combining the distance measurement and the k nearest neighbor strategy.
Drawings
FIG. 1 is a flowchart of a multi-label distance metric learning method based on interactive modeling according to the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples.
As shown in FIG. 1, the invention is a multi-label distance metric learning method based on feature space and label space interactive modeling, comprising the following steps:
(1) for any multi-mark application scene as figureSampling images, videos, texts and the like to obtain training data, extracting corresponding features and carrying out manual labeling to obtain training data D { (x)i,yi) I is more than or equal to 1 and less than or equal to m, wherein xiE x is a d-dimensional feature vector,is an example xiThe set of labels of (1).
(2) And preprocessing the extracted training samples, and filtering out samples with the mark occupancy rate smaller than a set threshold value, so as to improve the sample quality.
(3) And based on the Mahalanobis distance metric learning framework, considering the structural interaction between the feature space and the mark space in the multi-mark data, and expressing the distance metric matrix to be learned as a combined distance metric form.
Based on the multi-labeled data characteristics, the distance metric is represented as a combined metric form:
wherein, bl∈RdIs a combined basis vector, wlNot less than 0 is the non-negative combination coefficient (weight) to be learned, and the importance of each basis vector is described. B is constructed in combination with the discriminative information of the markers, taking into account the structured interactivity between the markers and the features, and specifically, for each marker, the first K feature vectors are obtained as local metric vectors using Principal Component Analysis (PCA), so that K × q b are total, and the combination coefficient w is a vector of dimension 1 × Kq.
(4) Defining multi-marker semantic similarity based on feature and marker collaborative computation, and constructing a triple constraint set; the specific substeps are as follows:
(41) under the multi-label scene, due to the combinability of multiple types of labels, a semantic similarity matrix S is obtained by performing collaborative calculation on the identification information of the feature space and the label space:
sij＝yi TGyj
where G＝(αA+(1-α)C)
wherein a islhCorresponding with a mark yl,yhA, since the correlation between the markers tends to be asymmetriclh≠ahl。clhAnd (4) cosine similarity of the corresponding combined basis vectors, wherein alpha is a balance coefficient, and the relative contribution of the balance feature space and the mark space to semantic similarity is balanced.
(42) Based on the definition of the semantic similarity matrix, the training set can be divided into the following similarity set and
Zi＝{xj|sij≥θ,j≠i,1≤j≤m}
a dissimilar set:
wherein θ is a threshold parameter of semantic similarity.
(43) Finding out the ones with the shortest Euclidean distance in the similar set and the dissimilar set respectivelySamples, forming a set KiAndthe triplet constraint set R is then expressed as follows:
wherein R comprisesAnd (4) a triplet.
(5) And (4) integrating the steps (3) and (4) to obtain a multi-mark distance measurement learning model as follows:
ζi≥0,wi≥0(1≤i≤m)
whereinRepresenting the embedding spatial distance of the two samples,indicating the dissimilarity of the two token vectors.
This optimization problem can be solved for the combined weights w using a FISTA iterative optimization algorithm.
(6) After learning distance measurement, mapping training data to a distance measurement space, and then learning by using an existing multi-label learning algorithm to obtain a multi-label classifier based on distance measurement learning.
(7) And inputting the sample to be detected into the classifier to obtain a labeled sample. Firstly, the multi-label classifier is mapped to an learned distance metric space and then input into the multi-label classifier for label prediction.
(8) Performing sampling inspection on the labeling result, and if the labeling result is qualified, finishing the sampling inspection; otherwise, returning to the step (1), and continuing to extract the sample for model adjustment and updating.
In summary, the invention provides a multi-label distance measurement learning method based on feature space and label space interactive modeling, which makes full use of the correlation between labels and the interactivity between the feature space and the label space, can process any multi-label learning scene, and reduces the complexity of a multi-label learning system by combining distance measurement and a k nearest neighbor strategy, and promotes the practical application of a multi-label learning framework.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.

Claims (5)

1. A multi-label distance measurement learning method based on interactive modeling is characterized by comprising the following steps:
(1) extracting training data from any multi-label application scene, and performing multi-label labeling on the training data;
(2) preprocessing the extracted training samples, filtering and marking the samples with the occupancy rate smaller than a set threshold value, and improving the quality of the samples;
(3) based on a Mahalanobis distance measurement learning framework, considering the structural interaction between a feature space and a mark space in multi-mark data, and expressing a distance measurement matrix to be learned as a combined distance measurement form;
(4) defining multi-marker semantic similarity based on feature and marker collaborative computation, and constructing a triple constraint set;
(5) integrating the steps (3) and (4) to obtain a multi-mark distance measurement learning model, and carrying out optimization solution on the multi-mark distance measurement learning model;
(6) after learning distance measurement, mapping training data to a distance measurement space, and then learning by using the existing multi-label learning algorithm to obtain a multi-label classifier based on distance measurement learning;
(7) inputting a sample to be predicted into the multi-label classifier to obtain a labeled sample;
(8) performing sampling inspection on the labeling result, and if the labeling result is qualified, finishing the sampling inspection; otherwise, returning to the step (1), and continuing to extract the sample for model adjustment and updating.
2. The interaction modeling based multi-label distance metric learning method as claimed in claim 1, wherein the application scene in step (1) comprises image, text and video scenes.
3. The interaction modeling based multi-label distance metric learning method according to claim 1, wherein the distance metric to be learned is represented in step (3) as a combined metric form as follows:
wherein, bl∈RdIs a combined basis vector, wl≧ 0 is the non-negative combination coefficient (weight) to be learned, l is the interval index of the summation sign, varying from 1 to K. K is the upper bound of the summation symbol and is a self-defined value, and the method constructs a component basis vector for each token. T denotes the transpose of the matrix.
4. The interaction modeling based multi-label distance metric learning method according to claim 1, wherein the step (4) comprises the steps of:
(41) obtaining a semantic similarity matrix S by performing collaborative calculation on the identification information of the feature space and the mark space:
sij＝yi TGyj
where G＝(αA+(1-α)C)
wherein j corresponds toThis xjSubscript of (i) corresponds to sample xiA subscript of (a); a is a matrix of q × q, q represents the number of marks, and the value of each row and column in the matrix is represented by alhM is the upper bound of the summation symbol and represents m samples; a islhCorresponding with a mark ylAnd yhA, since the correlation between the markers tends to be asymmetriclh≠ahl(ii) a C is a matrix of q x q, q represents the number of marks, and the value of each row and column in the matrix is represented by ClhIs obtained by the calculation formula of (c)lhCorresponding to the cosine similarity of the combined base vector, T represents the transposition of the matrix; alpha is a balance coefficient, and the relative contribution of the balance feature space and the mark space to the semantic similarity is balanced;
(42) setting a threshold parameter theta of semantic similarity, and aiming at a training sample x according to the definition of a semantic similarity matrixiCan be divided into the following similarity sets ZiAnd dissimilar set
Zi＝{xj|sij≥θ,j≠i,1≤j≤m}
(43) Respectively finding K samples with the nearest Euclidean distance in the similar set and the dissimilar set to form a set KiAndthe triplet constraint set R is then expressed as follows:
5. the interactive modeling based multi-label distance metric learning method according to claim 1, wherein the learning model of the step (5) is as follows:
wherein w is a non-negative combination coefficient (weight) to be learned in the step (3), C is a balance coefficient, m represents the number of samples, ζ is a relaxation variable representing loss, and R is a triplet constraint set;representing the embedding spatial distance of the two samples,indicating the degree of dissimilarity of the two token vectors,
CN201910715600.9A 2019-08-05 2019-08-05 Multi-label distance measurement learning method based on interactive modeling Pending CN110598733A (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN201910715600.9A CN110598733A (en) 2019-08-05 2019-08-05 Multi-label distance measurement learning method based on interactive modeling
PCT/CN2019/100001 WO2021022571A1 (en) 2019-08-05 2019-08-09 Interactive modeling-based multi-label distance metric learning method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910715600.9A CN110598733A (en) 2019-08-05 2019-08-05 Multi-label distance measurement learning method based on interactive modeling

Publications (1)

Publication Number Publication Date
CN110598733A true CN110598733A (en) 2019-12-20

Family

ID=68853588

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910715600.9A Pending CN110598733A (en) 2019-08-05 2019-08-05 Multi-label distance measurement learning method based on interactive modeling

Country Status (2)

Country Link
CN (1) CN110598733A (en)
WO (1) WO2021022571A1 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111507293A (en) * 2020-04-22 2020-08-07 浙江工业大学 Signal classification method based on graph convolution model
CN111783788A (en) * 2020-06-24 2020-10-16 东南大学 Multi-label classification method facing label noise
CN111274377B (en) * 2020-01-23 2023-06-02 支付宝(杭州)信息技术有限公司 Method and system for training mark prediction model

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113569967B (en) * 2021-07-30 2023-01-10 昆明学院 Tea leaf impurity data labeling method based on supervised machine learning
CN114168478B (en) * 2021-12-13 2024-05-28 东北大学 Software Bug detection method based on multi-graph multi-mark learning
CN114844677B (en) * 2022-03-31 2024-05-28 北京工业大学 Integrated SMOTE method
CN115205632B (en) * 2022-07-07 2023-07-18 山西大学 Semi-supervised multi-view metric learning method in Riemann space

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100570640C (en) * 2007-09-29 2009-12-16 清华大学 A kind of method for expressing of people's motion
CN104331442A (en) * 2014-10-24 2015-02-04 华为技术有限公司 Video classification method and device
CN108491865A (en) * 2018-03-05 2018-09-04 南京邮电大学 A kind of more example multiple labeling sorting techniques based on metric learning
CN109242039A (en) * 2018-09-27 2019-01-18 东南大学 It is a kind of based on candidates estimation Unlabeled data utilize method
CN110009017A (en) * 2019-03-25 2019-07-12 安徽工业大学 A kind of multi-angle of view multiple labeling classification method based on the study of visual angle generic character

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111274377B (en) * 2020-01-23 2023-06-02 支付宝(杭州)信息技术有限公司 Method and system for training mark prediction model
CN111507293A (en) * 2020-04-22 2020-08-07 浙江工业大学 Signal classification method based on graph convolution model
CN111783788A (en) * 2020-06-24 2020-10-16 东南大学 Multi-label classification method facing label noise

Also Published As

Publication number Publication date
WO2021022571A1 (en) 2021-02-11

Similar Documents

Publication Publication Date Title
CN110598733A (en) Multi-label distance measurement learning method based on interactive modeling
CN110414462B (en) Unsupervised cross-domain pedestrian re-identification method and system
CN109241894A (en) A kind of specific aim ticket contents identifying system and method based on form locating and deep learning
CN110321830A (en) A kind of Chinese character string picture OCR recognition methods neural network based
CN111414906A (en) Data synthesis and text recognition method for paper bill picture
CN107103326A (en) The collaboration conspicuousness detection method clustered based on super-pixel
Sun et al. Robust text detection in natural scene images by generalized color-enhanced contrasting extremal region and neural networks
CN109447123B (en) Pedestrian re-identification method based on label consistency constraint and stretching regularization dictionary learning
CN115205570B (en) Unsupervised cross-domain target re-identification method based on comparative learning
CN112733866A (en) Network construction method for improving text description correctness of controllable image
CN112541355A (en) Few-sample named entity identification method and system with entity boundary class decoupling
CN110347857B (en) Semantic annotation method of remote sensing image based on reinforcement learning
CN108681735A (en) Optical character recognition method based on convolutional neural networks deep learning model
CN111738355A (en) Image classification method and device with attention fused with mutual information and storage medium
CN111291705B (en) Pedestrian re-identification method crossing multiple target domains
CN112926379A (en) Method and device for constructing face recognition model
CN114332007A (en) Transformer-based industrial defect detection and identification method
CN110443273B (en) Zero-sample-confrontation learning method for cross-class identification of natural images
CN117690031B (en) SAM model-based small sample learning remote sensing image detection method
CN107452374A (en) Based on unidirectional from the various visual angles language identification method for marking auxiliary information
CN104331717B (en) The image classification method that a kind of integration characteristics dictionary structure is encoded with visual signature
CN111144469B (en) End-to-end multi-sequence text recognition method based on multi-dimensional associated time sequence classification neural network
CN116543269A (en) Cross-domain small sample fine granularity image recognition method based on self-supervision and model thereof
CN109145749B (en) Cross-data-set facial expression recognition model construction and recognition method
CN115797642A (en) Self-adaptive image semantic segmentation algorithm based on consistency regularization and semi-supervision field

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20191220