Subtask-based image emotion mark distribution learning method

Abstract

The invention provides an image emotion mark distribution learning method based on subtasks, and aims to solve the problem that an existing mark distribution learning method based on auxiliary tasks depends on expert knowledge. The method comprises the steps of image emotion data collection, structured feature preprocessing, subtask mark distribution set construction, and a shallow / deep model-based learning method considering subtasks. Compared with the prior art, the method has the advantages that the sub-task mark space mask matrix is learned, the mark distribution is subdivided into the sub-task mark distribution, so that the pseudo supervision information containing new knowledge is generated, the sub-task mark distribution is constructed, the sub-task mark distribution is regarded as an auxiliary task for joint learning, and the prediction accuracy of the image emotion mark distribution is effectively improved. The method can be widely applied to the fields of mental health monitoring, emotion calculation, man-machine interaction and the like, and a reliable technical scheme is provided for complex emotion recognition.

Description

Technical Field

[0001] The present invention relates to the field of image emotion prediction, belongs to the category of computer vision, and particularly to the aspect of labeled distribution learning of image emotions. Background Art

[0002] With the rapid development of social media and human-computer interaction technology, image emotion recognition technology has been widely used in the fields of mental health monitoring, affective computing and human-computer interaction. Traditional image emotion recognition methods mainly rely on rough feature extraction, including edge detection, gray-level co-occurrence matrix or local binary pattern features, and rely on simple single-label annotation. With the rapid development of neural network methods, automated feature extraction has been widely used in emotion recognition tasks. However, single-label methods often face challenges in dealing with complex emotional states.

[0003] In recent years, label distribution learning, as an emerging machine learning paradigm, has been gradually applied in emotion recognition. Different from traditional single-label classification, label distribution learning assigns label correlations to sample examples and has become an effective solution to the problem of label ambiguity. Specifically for the image emotion prediction task, label distribution learning can handle the distribution of emotion labels with multiple complex emotions, providing more sufficient supervised information to improve the information richness and accuracy of image emotion prediction results.

[0004] Despite this, labeled distribution learning still faces the dilemma of insufficient data in real applications, because obtaining high-quality labeled distributions is time-consuming and laborious. To solve the problem of insufficient data in labeled distribution learning models in real application scenarios, auxiliary task learning provides a solution. Specifically, expert knowledge and prior knowledge in the field, additional data in similar fields, etc. are introduced as auxiliary tasks to guide the learning of the main task. Usually, when learning the main task and auxiliary tasks together, the model can learn more discriminative and robust representations, thereby improving performance. However, auxiliary tasks may constrain the generalization performance of the model. Due to its dependence on expert knowledge and prior knowledge in the field, the model can only solve specific problems in a single field. It is understood that some existing multi-label learning methods divide the label space to construct subtasks, which only have dependence on the main task itself. In addition, migrating this construction method to the labeled distribution learning task can generate pseudo-supervisory information containing new knowledge, which not only makes up for the problem of data scarcity in existing technologies, but also guides the learning of the main task.

[0005] Therefore, the present invention regards the subtasks obtained by dividing the label space as auxiliary tasks, explores the label distribution learning method based on the subtask, and applies it to the label distribution prediction of image emotion. Summary of the invention

[0006] Purpose of the invention: In order to alleviate the problem that the existing auxiliary task-based labeled distribution learning method is dependent on expert knowledge, the present invention provides a subtask-based image emotion labeled distribution learning method, which generates pseudo-supervisory information containing new knowledge by constructing an efficient subtask-based labeled distribution learning framework, thereby compensating for the problem of data scarcity in the prior art and guiding the model to learn more discriminative representations, thereby improving the accuracy and robustness of emotion prediction.

[0007] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution:

[0008] A subtask-based image emotion label distribution learning method comprises the following steps:

[0009] Collect a dataset of labeled distributions containing a variety of image emotions, use visual images as feature parts, and construct emotional label distributions;

[0010] Preprocessing the feature part of the labeled distribution data set, including image cropping and data normalization, to generate a structured feature representation;

[0011] Learning a subtask label space mask matrix to subdivide the label distribution into multiple subtask label distributions to generate pseudo-supervision information containing new knowledge;

[0012] The label distribution estimator is used to learn the subtask label distribution, and the structured features and all subtask label distribution sets are concatenated as features to train the model corresponding to the main task label space.

[0013] In other embodiments, after the subtask label distribution set is constructed, a deeper model can also be used for learning: mapping the structured feature representation to obtain a subspace representation, connecting the subspace representation and the predicted subtask label distribution set, and training the label distribution learner based on the connection result.

[0014] Furthermore, the above method can be applied to various variants of labeled distribution learning tasks, including labeled distribution learning for classification, incomplete labeled distribution learning, and label enhancement.

[0015] Furthermore, the learning of the subtask label space mask matrix subdivides the label distribution into a plurality of subtask label distributions to generate pseudo-supervisory information containing new knowledge, including:

[0016] Considering the information diversity and local ignorance of the subtask label space, the mask matrix of the subtask label space is constructed and learned;

[0017] creating partitions according to the subtask label space mask matrix, and slicing the original label distribution matrix based on the partitions;

[0018] The subtask label distribution description is normalized to ensure that the subtask label distribution has a probabilistic simplex constraint.

[0019] Furthermore, the subtask label space mask matrix M∈{0,1} is obtained by solving the following problem: T×L , where T is the expected number of subtasks, L is the number of sentiment tags; the t-th row of the matrix M is m t , represents the mask corresponding to the possible tag space in the t-th subtask, and the matrix elements 1 and 0 represent whether the subtask tag space contains or does not contain the tag, respectively;

[0020]

[0021] Where N is the number of samples, λ is the balance hyperparameter, and d i represents the label distribution corresponding to the sample numbered i in the data set, T represents the transpose operation, and ||·|| represents the modulus operation of the vector.

[0022] In some embodiments, when predicting the sentiment tag distribution of a new image sample, the sample to be predicted and all predicted sub-task tag distributions are connected to obtain the connected data to be predicted, and the predicted main task tag distribution is obtained using an estimator.

[0023] In some embodiments, when predicting the sentiment label distribution of a new image sample, the sample to be predicted is input into the subspace representation operator to obtain a subspace representation, the subspace representation is input into the subtask label distribution prediction operator to obtain the subtask label distribution, the subspace representation and all predicted subtask label distributions are connected, and the predicted main task label distribution is obtained using the label distribution learner.

[0024] The types of learners involved include linear transformation matrices, nonlinear models, and neural networks based on encoder-encoder structures.

[0025] Furthermore, the predicted results of the marker distribution and the true value of the marker distribution can be reduced in dimension via the implicit Dirichlet allocation algorithm and mapped to the color space by the sigmoid mathematical operation. The structured features or subspace representations can be reduced in dimension via the t-distributed random neighbor embedding algorithm, so that different marker distributions can present different colors in the two-dimensional space obtained by the algorithm.

[0026] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned sub-task-based image emotion label distribution learning method.

[0027] Beneficial effects: The present invention uses subtasks as auxiliary tasks, which reduces the dependence of existing methods on expert knowledge, thereby enhancing the generalization of the model; the pseudo-supervisory information containing new knowledge generated by the present invention, namely the subtask label distribution, can expand the existing data, thereby enhancing the adaptability of the model under data scarcity conditions; the present invention uses subtasks to guide representation learning, so that the model is not prone to overfitting; in addition, the present invention also has "plug and play" capabilities, is more independent than the existing technology, and can be extended to derivative tasks of label distribution learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention.

[0029] Figure 2 The figure is a flowchart of a shallow model-based learning method that considers subtasks in an embodiment of the present invention.

[0030] Figure 3 Schematic diagram of the flow of a deep model-based learning method considering subtasks in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to more clearly understand the working principle and application of the present invention and further explain the present invention, the implementation details of the present invention will be described in detail below in conjunction with specific drawings and embodiments.

[0032] The definitions of the abbreviations and key terms involved in this embodiment are as follows:

[0033] (1) Label Distribution Learning (LDL)

[0034] (2) Local Binary Pattern (LBP)

[0035] (3) Stochastic Gradient Descent (SGD).

[0036] (4)Latent Dirichlet Allocation (LDA)

[0037] (5) t-Distributed Stochastic Neighbor Embedding (t-SNE; t-Distributed Stochastic Neighbor Embedding).

[0038] (6) Kullback-Leibler Divergence (KLD).

[0039] The embodiment of the present invention discloses a subtask-based image emotion label distribution learning method, which is abbreviated as S-LDL. The overall process of the technical solution is as follows: Figure 1 As shown, it includes image emotion data collection, structured feature preprocessing, sub-task label distribution set construction, shallow model-based learning method considering sub-tasks / deep model-based learning method considering sub-tasks, which will be introduced one by one below.

[0040] Step S1. Image emotion data collection: collect a labeled distribution dataset containing multiple image emotions, use visual images as feature parts, and construct the emotion label distribution.

[0041] Specifically, in this step, we collect a dataset containing a variety of labeled distributions that can induce human emotions, use visual images as feature parts, and define the original image feature matrix Where N is the number of samples. To facilitate data organization, it is assumed that the length and width of the image are both P, and C is the number of channels of the image. The emotional tag distribution is constructed based on crowdsourcing. The emotional tags corresponding to the emotional tag distribution include happiness, sadness, anger, surprise, disgust, fear and neutrality. The emotional tag distribution of all samples is defined as the matrix D, d i represents the label distribution corresponding to the sample numbered i in the data set. is the descriptive degree of example x with respect to label y. The label distribution satisfies the constraints of the probability simplex, including the constraint that the sum is one and the non-negativity constraint, that is, d i ∈Δ L-1 , where L is the number of sentiment tags, and the probability simplex of L-1 dimension Δ L-1 It is defined as follows mathematically:

[0042]

[0043] The source of the data set can be social media, online surveys or emotional databases, etc. Each sample of the data set is crowdsourced by a certain number of people and quantitatively evaluated based on the emotional markers. The conversion form of the crowdsourcing results to the facial emotion marker distribution matrix is ​​various, usually a normalization function.

[0044] Step S2. Structured feature processing: preprocessing the feature part of the labeled distribution dataset, including image cropping and data normalization, to generate a structured feature representation.

[0045] Specifically, in this step, the feature part of the data set is preprocessed, including the location and cropping of specific objects that induce emotions, and the LBP algorithm is used to extract the original image features into a Q-dimensional feature vector. The obtained feature vector is processed using a normalization function to obtain a structured feature matrix

[0046] Step S3. Construction of subtask tag distribution set,

[0047] The subtask label space mask matrix is ​​learned to subdivide the label distribution into multiple subtask label distributions to generate pseudo-supervisory information containing new knowledge, which ultimately promotes the method to learn more discriminative and robust representations. Specifically, it includes the following steps.

[0048] Step S301: Learn the subtask labeling space: define M∈{0,1} T×L is the subtask mask matrix, where T is the expected number of subtasks. The tth row of the matrix M, i.e., m t , represents the mask corresponding to the possible tag space in the t-th subtask, and the matrix elements 1 and 0 represent whether the tag is included or not in the subtask tag space, respectively.

[0049] To ensure that the knowledge generated in the subtask label space is rich and reliable and to avoid homogenized pseudo-supervisory information, it is necessary to reasonably control the ignorance of the sub-label distribution space. To address this challenge, this step uses the label distribution matrix to assign weights to the mask and considers the pairwise similarity between rows of the mask matrix. At the same time, the exponential technique is applied to convert the solution of M into a minimization problem, which is the following mathematical form:

[0050]

[0051] Among them, λ is a balancing hyperparameter. In the above problem objective, the first term avoids ignorance of the sub-label distribution space, and the second term encourages diversity of the sub-label distribution space. When optimizing the solution, it is important to make sure that each element M of the matrix M tj The following constraints are met:

[0052] M tj ∈{0,1}; t=1,…,T; j=1,…,L

[0053] The above problem is slightly more complicated than the common integer programming problem. For convenience, this step uses the SGD algorithm to solve it. In order to satisfy the matrix constraints, the result of this step will be finally processed by the sigmoid function and a threshold will be set. The default value is 0.95. If it is higher than this threshold, the matrix element is set to 1, otherwise, the matrix element is set to 0.

[0054] To perform subsequent operations, define the sub-tag space set as And initialized to And, define the subtask tag distribution matrix set as And initialized to

[0055] Step S302: Slice the original label distribution matrix: Create partitions according to the subtask label space mask matrix, and slice the original label distribution matrix based on the partitions. Specifically, traverse each row of the matrix M, record the traversal index as t, and define is the possible t-th sub-tag space. If there is a sentiment tag index j such that M tj =1, then y j Add to collection

[0056] This step also requires checking for illegal subtask masks. (i.e., the subtask is identical to the main task) or (i.e., the subtask marker distribution cannot be constructed), skip the remaining operations on that row mask.

[0057] Step S303 normalizes the subtask tag distribution description to ensure that the subtask tag distribution has a probability simplex constraint. To avoid mathematical errors in subsequent operations, it is necessary to perform a clip operation on each description of the original tag distribution matrix, where the clip(x,a,b) function limits the input x to the range [a,b]. Definition D (t) Label the distribution matrix for the possible t-th subtask, specifically, all subtasks that satisfy clip(d ·j ,ε,1), where d ·j is the j-th column of matrix D.

[0058] After the clip operation, the subtask tag distribution description is normalized and the matrix D is traversed. (t) For each row of , let the traversal index be i, then the normalization steps are in the following mathematical form:

[0059]

[0060] Among them, the normalization operation The definition of can be in the following mathematical form:

[0061]

[0062] in,[·] j is the j-th element of the vector being operated on.

[0063] At the end of each row traversal of the matrix M, the subtask tag space and subtask tag distribution matrix that are successfully retained and not skipped due to the above illegal screening are aggregated and merged, that is, execution as well as

[0064] Finally, the module returns the subtask tag space collection And the subtask tag distribution matrix set The time complexity of the entire module is

[0065] The construction of the subtask label distribution set in this step does not rely on additional data in a field similar to the dataset, expert knowledge in any field, prior knowledge, etc.

[0066] Step S4. The joint learning of the main task label distribution and the subtask label distribution can adopt a learning method based on a shallow model or a learning method based on a deep model.

[0067] The following introduces a learning method based on a shallow model that considers subtasks. In short, it uses a shallow label distribution estimator and a simple connection operation to quickly build a model to jointly learn the main task label distribution and the subtask label distribution. The overall process is as follows Figure 2 Specifically, it can be divided into the following steps:

[0068] Step S401: Learn subtask label distribution estimator: select shallow label distribution estimator, total These estimators can be the same type of labeled distribution learning algorithms or different types, such as SA-IIS, LDSVR, AA-kNN, etc. Initialize the model parameters of each estimator.

[0069] The subtask label distribution matrix corresponding to different subtask label spaces is learned. All subtask markers in the distribution are traversed, and the traversal index is t, and a new dataset {X,D (t)}, and train the tth estimator on this data set, denoted as f (t) , for the new sample x′ to be predicted, the estimated subtask label distribution is obtained, that is:

[0070] d (t)′ ←f (t) (x′)

[0071] By the end of the traversal, the samples to be predicted have corresponding predicted subtask label distributions for all subtasks.

[0072] Step S402: label the distribution set D of the structured feature X and all subtasks (t) (in, ) to connect and obtain the connected data Z.

[0073] Step S403 learns the merged label distribution estimator: takes the connected data Z as features, constructs a new data set {Z, D}, and trains the model f corresponding to the main task label space.

[0074] Step S404 predicts the label distribution of unknown samples: predict the emotional label distribution of new image samples, link the sample to be predicted x′ and all predicted subtask label distributions d (t)′ (in, ), get the connected data to be predicted z′, and use the estimator to get the predicted main task label distribution, that is:

[0075] d′←f(z′)

[0076] Finally, the module returns the predicted main task label distribution d′ as the module output.

[0077] The following introduces a deep model-based learning method that considers subtasks, that is, using deep learning technology to jointly learn the main task label distribution and the subtask label distribution. The overall process is as follows: Figure 3 Specifically, the module includes the following steps:

[0078] Step S411: Mapping subspace representation: Mapping the structured feature representation to obtain a subspace representation, defining It is a subspace representation mapping operator whose parameters are guided by the joint learning of the subtask and the main task, making the representation more discriminative and robust. Specifically, this step can be expressed in the following mathematical form:

[0079]

[0080] Define ψ(·) as the subtask label distribution prediction operator, and the label distribution of each subtask can be obtained according to the input representation, that is, the following mathematical form:

[0081]

[0082] The learning of the subtask label distribution is performed in a way that minimizes the subtask learning loss, which is weighted by the sum of the descriptiveness grouped by labels, so that more reliable label spaces receive higher attention during the optimization process. Specifically, the subtask learning loss is defined as follows:

[0083]

[0084] in, They represent the true value of the main task description, the true value of the subtask description, and the predicted value of the subtask description, respectively.

[0085] Step S412: The subspace representation R and the predicted subtask label distribution set The connection is performed, specifically, Z = (R, ψ(R)).

[0086] Step S413 trains a labeled distribution learner according to the connection result, and defines the labeled distribution learner as ω(·), which can be any existing labeled distribution learning model (provided that the model can be expressed as a network structure in theory).

[0087] Step S414 predicts the label distribution of unknown samples: predict the emotional label distribution of new image samples, that is,

[0088] Finally, the module returns the predicted main task label distribution matrix As module output.

[0089] The final predicted value and the actual result of the label distribution can be reduced in dimension through the LDA algorithm and mapped to the color space by the sigmoid mathematical operation. The enhanced structured features obtained by the structured feature enhancement module can be reduced in dimension through the t-SNE algorithm, so that different label distributions can present different colors in the two-dimensional space obtained by the t-SNE algorithm.

[0090] The embodiment of the present invention can be implemented based on the deep learning open source tool TensorFlow 2.0. The LDL loss function is defined as follows:

[0091]

[0092] Among them, the specific implementation of l(·,·) is KLD, that is, the following mathematical form:

[0093]

[0094] Note that the final implementation of the present invention can be applied to various variant tasks of label distribution learning, such as classification-oriented label distribution learning, incomplete label distribution learning, and label enhancement, etc. The following specifically describes how to modify the content of the present invention to adapt to various variant tasks.

[0095] For classification-oriented labeled distribution learning tasks, the subtask construction still relies on the original labeled distribution matrix, and the main task loss is replaced by the task-specific loss Its specific definition is beyond the scope of this specification, so it will not be elaborated here. is the degenerate label distribution matrix, which is defined as follows:

[0096]

[0097] in,[·] ij The element with index j corresponding to the i-th row of the operated matrix.

[0098] For incomplete label distribution learning tasks, subtask construction relies on the label distribution matrix with limited observable range in The operations are defined as follows:

[0099]

[0100] in,[·] ij The definition of is the same as above. The main task loss is replaced by the task-specific loss of the incomplete labeled distribution learning task The specific definition is beyond the scope of this specification and will not be elaborated here.

[0101] For the label enhancement task, the subtask construction depends on the logical label matrix L. The theme of adaptation is similar to the above scheme, and the details are not expanded here.

[0102] The learners involved in all steps of the present invention can be represented as any learning model, including any linear transformation matrix, nonlinear model, and neural network based on encoder-decoder structure, etc. In this example, a neural network based on encoder-decoder structure is used to implement the learning method based on deep model in the present invention, and image emotion label distribution learning is performed based on this.

[0103] Use TensorFlow 2.0 to load the network-derived emotional label distribution dataset and perform label distribution learning tasks, using the LDL loss function as the main task loss. Define the parameters of the entire model as Θ, and train the model according to the following loss:

[0104]

[0105] Among them, α is the balancing hyperparameter of the subtask.

[0106] Through the above training method, the present invention can effectively improve the accuracy of sentiment tag distribution prediction in various data distribution environments, and provide a more accurate and reliable sentiment analysis tool for related applications.

[0107] The embodiment of the present invention also discloses a computer program product, including a computer program / instruction, which implements the steps of the aforementioned sub-task-based image emotion labeling distribution learning method when executed by a processor. The program / instruction code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program / instruction codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program / instruction code implements the steps of the method of the present invention when executed by the processor or controller. The program / instruction code can be executed completely on the machine, partially on the machine, partially on the machine as an independent software package and partially on a remote machine, or completely on a remote machine or server. What is not described in detail in the present invention is a well-known technology for those skilled in the art.

[0108] The above are embodiments of the present invention. Those skilled in the art should know that the present invention is not limited to the above embodiments. Without departing from the spirit and scope of the present invention, the present invention will also have various variations and improvements, and such variations and improvements all fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A subtask-based image emotion label distribution learning method, characterized in that: The steps include: Collect a dataset of labeled distributions containing a variety of image emotions, use visual images as feature parts, and construct emotional label distributions; Preprocessing the feature part of the labeled distribution data set, including image cropping and data normalization, to generate a structured feature representation; Learning a subtask label space mask matrix to subdivide the label distribution into multiple subtask label distributions to generate pseudo-supervision information containing new knowledge; The label distribution estimator is used to learn the subtask label distribution, and the structured features and all subtask label distribution sets are concatenated as features to train the model corresponding to the main task label space.

2. A subtask-based image emotion label distribution learning method, characterized in that: The steps include: Collect a dataset of labeled distributions containing a variety of image emotions, use visual images as feature parts, and construct emotional label distributions; Preprocessing the feature part of the labeled distribution data set, including image cropping and data normalization, to generate a structured feature representation; Learning a subtask label space mask matrix to subdivide the label distribution into multiple subtask label distributions to generate pseudo-supervision information containing new knowledge; The structured feature representation is mapped to obtain a subspace representation, the subspace representation and the predicted subtask label distribution set are connected, and the label distribution learner is trained based on the connection result.

3. The subtask-based image emotion label distribution learning method according to claim 1 or 2, characterized in that: The method is applied to multiple variants of labeled distribution learning, including labeled distribution learning for classification, incomplete labeled distribution learning, and label augmentation.

4. The subtask-based image emotion label distribution learning method according to claim 1 or 2, characterized in that: The learning of the subtask label space mask matrix subdivides the label distribution into multiple subtask label distributions to generate pseudo-supervisory information containing new knowledge, including: Considering the information diversity and local ignorance of the subtask label space, the mask matrix of the subtask label space is constructed and learned; creating partitions according to the subtask label space mask matrix, and slicing the original label distribution matrix based on the partitions; The subtask label distribution description is normalized to ensure that the subtask label distribution has a probabilistic simplex constraint.

5. The subtask-based image emotion label distribution learning method according to claim 1 or 2, characterized in that: The subtask label space mask matrix M∈{0,1} is obtained by solving the following problem T×L , where T is the expected number of subtasks, L is the number of sentiment tags; the t-th row of the matrix M is m t , represents the mask corresponding to the possible tag space in the t-th subtask, and the matrix elements 1 and 0 represent whether the subtask tag space contains or does not contain the tag, respectively; Where N is the number of samples, λ is the balance hyperparameter, and d i represents the label distribution corresponding to the sample numbered i in the data set, represents the transpose operation, and ||·|| represents the modulus operation of the vector.

6. The subtask-based image emotion label distribution learning method according to claim 1, characterized in that: When predicting the sentiment tag distribution of a new image sample, the sample to be predicted and all predicted sub-task tag distributions are connected to obtain the connected data to be predicted, and the predicted main task tag distribution is obtained using the estimator.

7. The subtask-based image emotion label distribution learning method according to claim 2, characterized in that: When predicting the sentiment label distribution of a new image sample, the sample to be predicted is input into the subspace representation operator to obtain the subspace representation, the subspace representation is input into the subtask label distribution prediction operator to obtain the subtask label distribution, the subspace representation and all the predicted subtask label distributions are connected, and the label distribution learner is used to obtain the predicted main task label distribution.

8. The subtask-based image emotion label distribution learning method according to claim 2, characterized in that: The types of learners involved include linear transformation matrices, nonlinear models, and neural networks based on encoder-encoder structures.

9. The subtask-based image emotion label distribution learning method according to claim 1 or 2, characterized in that: The predicted results of the label distribution and the true value of the label distribution can be reduced in dimension through the implicit Dirichlet allocation algorithm and mapped to the color space by the sigmoid mathematical operation. The structured features or subspace representation can be reduced in dimension through the t-distributed random neighbor embedding algorithm, so that different label distributions can present different colors in the two-dimensional space obtained by the algorithm.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images