A Multimodal Drug Interaction Event Prediction Method and System Based on Compact Representation Learning

Abstract

This invention provides a method and system for predicting multimodal drug interaction events based on compact representation learning. The method includes: acquiring multimodal features of each drug in a drug pair to be predicted, wherein the multimodal features include at least two of biological features, molecular structure features, and knowledge graph features; performing intramodal compact representation learning on each modal feature of each drug to eliminate redundant information within a single modality and generate compact sub-representations for each modality; using the compact sub-representations corresponding to the molecular structure features as anchors, performing cross-modal alignment of the compact sub-representations corresponding to the biological features and knowledge graph features through mutual information minimization constraints to eliminate intermodal redundant information; constructing a multimodal fusion representation for each drug based on the compact sub-representations of each modality; concatenating the multimodal fusion representations of the two drugs in the drug pair and inputting them into a classifier to output the prediction result of the interaction event type of the drug pair.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and artificial intelligence, specifically relating to a method and system for predicting multimodal drug interaction events based on compact representation learning. Background Technology

[0002] Predicting drug interaction events (DDIEs) plays a crucial role in ensuring medication safety and optimizing treatment regimens, especially given the increasing prevalence of multidrug therapy in the management of complex diseases. Unexpected DDIEs can lead to adverse drug reactions, decreased treatment efficacy, and even life-threatening situations. Therefore, effectively identifying potential DDIEs is of great significance for maintaining public health safety and advancing clinical research, particularly given the limitations of clinical trials during the new drug development phase.

[0003] Traditional DDIE prediction methods primarily rely on single-modal data, such as artificially designed features or molecular structures. These models often fail to adequately characterize the multi-factorial properties of pharmacological signals. Existing multimodal prediction methods typically extract features independently from each modality, neglecting intra- and intermodal redundancy. This leads to inconsistencies in drug embedding characterization and the omission of informative pharmacological details. (See attached instructions.) Figure 1 As shown, when all the multimodal features of ibuprofen are used directly without filtering (see red dashed box), redundant information in the biological, molecular, and knowledge graph modalities (such as the biological target BCL2 and the molecular substructure benzene ring) leads to incorrect predictions of toxicity events with labetalol. Conversely, after reducing redundancy (see green dashed box), the multimodal information of ibuprofen becomes more compact, clearly reflecting its core pharmacological properties, such as inhibition of the PTGS pathway and the carboxyl group associated with anti-inflammatory effects. This compact drug characterization naturally leads to the correct DDIE prediction that labetalol reduces the antihypertensive activity of ibuprofen. Summary of the Invention

[0004] To address the shortcomings and deficiencies of existing technologies, this invention provides a multimodal drug interaction event prediction method and system based on compact representation learning. It aims to solve the problems of insufficient prediction accuracy caused by excessive redundant information within and between modes, inconsistent drug representations, and omission of core pharmacological details in existing multimodal drug interaction event prediction methods.

[0005] This method first acquires the multimodal features of each drug in the drug pair to be predicted. The multimodal features include at least two of the following: biological features, molecular structure features, and knowledge graph features, comprehensively covering the intrinsic biomedical attributes and extrinsic contextual information of the drugs. Subsequently, drug representations are optimized through a two-stage compact representation learning process. Intramodal compact representation learning implements targeted processing for different modal features. Biological features and molecular structure features are reconstructed and optimized using an autoencoder to remove redundancy. Knowledge graph features are filtered using learnable pruning weights combined with random noise and temperature parameters, filtered by an activation function, and optimized for relation pruning through KL divergence constraints. Then, compact sub-representations of each modality are generated through a heterogeneous graph neural network. Intermodal compact representation learning uses the compact sub-representations corresponding to molecular structure features as anchors and achieves cross-modal alignment by minimizing a loss function constructed based on mutual information, further reducing intermodal redundancy and ensuring representation consistency. Based on this, a multimodal fusion representation of each drug is constructed by vector concatenation. The fusion representations of the two drugs in a drug pair are combined and input into the classifier. At the same time, an overall objective function including classification loss, intramodal compaction loss and intermodal compaction loss is constructed for end-to-end training. By balancing the weights of each loss through hyperparameters, accurate prediction of drug interaction event types can be achieved.

[0006] This invention jointly models multi-source drug information, effectively highlighting key pharmacological signals such as critical targets, pathways, and functional groups. This not only makes the prediction results more interpretable, but also demonstrates excellent performance on datasets containing multiple drug-drug interaction events. In particular, it achieves significant performance improvement in new drug interaction prediction scenarios that did not appear during the training phase, and has strong generalization ability and application value.

[0007] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0008] A multimodal drug interaction event prediction method based on compact representation learning includes:

[0009] Obtain the multimodal features of each drug in the drug pair to be predicted, wherein the multimodal features include at least two of the following: biological features, molecular structure features, and knowledge graph features;

[0010] For each modal feature of each drug, intramodal compact representation learning is performed to eliminate redundant information within a single modality and generate compact sub-representations for each modality.

[0011] Using compact sub-representations corresponding to molecular structure features as anchors, cross-modal alignment of compact sub-representations corresponding to biological features and knowledge graph features is performed through mutual information minimization constraints to eliminate redundant information between modalities.

[0012] A multimodal fusion representation of each drug is constructed based on the compact sub-representation of each modality. The multimodal fusion representations of the two drugs in the drug pair are then concatenated and input into a classifier to output the prediction results of the interaction event types of the drug pair.

[0013] Furthermore, the intramodal compact representation learning includes:

[0014] Biological features and molecular structural features were reconstructed and optimized using an autoencoder, respectively.

[0015] After learningable pruning of the relation weights in the knowledge graph features, the knowledge compact sub-representation is generated by processing the heterogeneous graph neural network.

[0016] Furthermore, the learnable pruning includes:

[0017] Learnable pruned weights are generated based on the original relation weights. By combining random noise and temperature parameters, the pruned relation weight distribution is generated through differentiable sampling technology. The difference between the pruned relation weight distribution and the original relation weight distribution is constrained by KL divergence.

[0018] Furthermore, the multimodal fusion representation is generated by vector splicing of biological compact representation, molecular compact representation and knowledge compact representation.

[0019] Furthermore, the step of splicing the multimodal fusion representations of the two drugs in the drug pair specifically involves connecting the multimodal fusion representations of the two drugs in a preset dimensional order.

[0020] Furthermore, the method also includes constructing an overall loss function for end-to-end training, wherein the overall loss function includes classification loss, intra-modal compaction loss, and inter-modal compaction loss.

[0021] Furthermore, the intermodal compactness loss is the sum of the mutual information values ​​between biological compaction representation and molecular compaction representation, and between molecular compaction representation and knowledge compaction representation.

[0022] Furthermore, the multimodal features simultaneously include biological features, molecular structural features, and knowledge graph features.

[0023] Furthermore, at least one drug in the drug pair to be predicted did not appear during the model training phase.

[0024] Furthermore, a multimodal drug interaction event prediction system based on compact representation learning includes a processor, a memory, and a computer program stored in the memory, wherein when the processor executes the computer program, it implements the steps of the method described above.

[0025] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0026] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0027] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0028] The proposed two-stage compact learning framework systematically eliminates redundant information within and between biological, molecular structure, and knowledge graph modalities through intramodal feature reconstruction and relation pruning, as well as cross-modal mutual information minimization alignment anchored by molecular structure. This enables the generation of purer and more compact drug characterizations.

[0029] By constructing compact and information-rich drug characterizations, the model can more accurately capture key pharmacological signals of drug-drug interactions, reducing misjudgments caused by redundant noise, thereby improving overall prediction performance.

[0030] Since the model learns the core, compact pharmacological properties of drugs, rather than relying on noise or accidental associations that may exist in a specific dataset, the method of this invention can make more reliable interaction predictions based on the essential characteristics of new drugs that have not appeared in the training phase, thus demonstrating excellent generalization performance.

[0031] Through techniques such as relation weight pruning, the model can automatically identify and retain the core relations in the knowledge graph that are crucial to the prediction task. At the same time, compact cross-modal alignment makes features from different sources more semantically consistent, which helps to understand the underlying basis for the model to make specific predictions in subsequent analysis.

[0032] By unifying the classification task and the compact representation learning objective within an end-to-end training framework, and jointly optimizing the classification loss, intra-modal compact loss, and inter-modal compact loss, the model can simultaneously learn how to effectively eliminate redundancy and achieve accurate predictions, simplifying the training process and improving overall efficiency. Attached Figure Description

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0034] Figure 1 For the example given to illustrate the prediction of drug interaction events, the red dashed box represents redundant multimodal features that could lead to prediction errors, and the green dashed box represents compact and informative features that can guide accurate predictions.

[0035] Figure 2 This is a general framework diagram of the implementation of the solution in the embodiment of the present invention;

[0036] Figure 3 This is a visualization of the drug embeddings of three modalities learned on the DDIMDL dataset by the present invention and its variants. Detailed Implementation

[0037] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0038] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0039] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0040] This invention addresses the problem that existing multimodal DDIE prediction methods cannot fully model drug embedding representations. It proposes a multimodal drug interaction event prediction method and system based on compact representation learning. Through compact representation learning within and between modalities, and by constructing a fusion representation of the drug, more accurate drug interaction event predictions are achieved.

[0041] like Figure 2 As shown, the specific implementation process of the multimodal drug interaction event prediction method based on compact representation learning in this embodiment of the invention includes the following steps (stages):

[0042] S1, used for compact representation learning within modalities; modeling drugs from two perspectives—drug features and external knowledge—through a modal feature extractor. The former reflects the intrinsic biomedical properties of the drug, including biological characteristics and molecular structure; the latter encodes external contextual information, such as knowledge graphs.

[0043] S2, used for compact representation learning between modalities; aligning biological representations, knowledge representations, and molecular representations as anchors;

[0044] S3. After constructing the fusion characterization of the drug, predict drug interaction events.

[0045] In this embodiment, the specific process of compact representation learning within a modality includes the following steps:

[0046] S1.1 Calculate Jaccard similarity from three commonly used descriptors (i.e., chemical substructure, target, and enzyme) to obtain the initial biological characterization. .

[0047] S1.2 Obtain the SMILES string of the drug from DrugBank and parse it using RDKit to construct a molecular graph containing atomic and bond information; process the molecular graph using a graph neural network to generate a molecular characterization. .

[0048] S1.3, Integrating biological and molecular characterization , The input is fed into an autoencoder of a specific modality to generate an embedding. , where dim is the feature dimension of each modality compact sub-representation.

[0049] S1.4. The encoder projects each input into a low-dimensional representation, and then the decoder reconstructs the original feature space. The reconstruction loss is defined as:

[0050]

[0051] in and Indicates drug The original and reconstructed multimodal representations are given, where N is the number of drugs.

[0052] S1.5, Given a KG (Knowledge Graph) Introducing a relation pruning strategy to modify the original relation set Transformed into the following refined set (Relationship weight distribution after pruning):

[0053]

[0054] in Indicates embedding from the original relation Learnable pruning weights derived from [the text] It is noise sampled from the Gumbel distribution. It is a temperature hyperparameter used to control the smoothness of the Gumbel-Softmax distribution.

[0055] S1.6 Obtain refined relation embeddings (pruned relation embeddings) Adjusted via Kullback-Leibler (KL) divergence: .

[0056] S1.7, Obtain a compact KG Generate knowledge embeddings using heterogeneous graph neural networks .

[0057] S1.8, The overall intramodal compaction loss is calculated as follows:

[0058] .

[0059] In this embodiment, the specific process of compact representation learning between modalities includes the following steps:

[0060] S2.1, Introduce a mutual information minimization regularization term, two embeddings and Mutual information between them is defined as:

[0061]

[0062] in In this embodiment, the mutual information MI(・,・) is approximated by the InfoNCE loss function, and a neural network discriminator is used to distinguish between positive samples (different modal embeddings of the same drug) and negative samples (modal embeddings of different drugs), thereby achieving efficient estimation of mutual information.

[0063] S2.2 The overall intermodal compactness loss is calculated as follows:

[0064]

[0065]

[0066] Where E[·] represents the average of the calculation results of all drug samples in the training batch to achieve the expected estimate.

[0067] In this embodiment, drug interaction event prediction is performed after constructing the drug fusion characterization, specifically including the following steps:

[0068] S3.1 For each drug, a fusion characterization is constructed by concatenating its compact biological, molecular, and knowledge graph embeddings:

[0069]

[0070] S3.2, Given a drug pair The predicted drug interaction events are as follows:

[0071]

[0072] in It is a multilayer perceptron. This represents the number of categories of drug interaction events.

[0073] Preferably, in this embodiment, both the encoder and decoder of the autoencoder consist of two fully connected layers with two hidden layer neurons of size dim × 2 and ReLU activation function; the graph neural network for processing molecular graphs is a graph convolutional network (GCN) with two hidden layers and output dimension dim for each layer; the heterogeneous graph neural network for processing knowledge graphs is a relational graph attention network (RGAT) with two attention layers; the MLP is a single fully connected layer with an input dimension of 6 × dim and an output dimension of C, and Softmax activation function.

[0074] S3.3 Treat drug interaction event prediction as a multi-class classification task and optimize it using cross-entropy loss:

[0075]

[0076] in It is the training set. It is the actual value.

[0077] S3.4 The overall objective function is calculated as follows:

[0078]

[0079] Here, α and β are hyperparameters that balance the compact loss within and between modes.

[0080] Preferably, in this embodiment, the model training uses the Adam optimizer, with an initial learning rate of 1e-4, a batch size of 64, and 200 training epochs. An early stopping strategy is adopted (training stops if the validation set AUC does not improve for 10 consecutive epochs). The initial values ​​of hyperparameters α and β are both set to 0.1, and can be tuned in the range of [0.01, 1.0] through grid search.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. By jointly modeling multi-source information such as biological features, molecular structure, and knowledge graphs, redundancy within and between modalities is significantly reduced, resulting in a consistent and compact drug representation, thereby more comprehensively and accurately characterizing the pharmacological properties of drugs.

[0083] 2. Effectively mining complementary information between different modalities highlights key pharmacological signals such as critical targets, pathways, and functional groups, making DDIE prediction results more interpretable and improving prediction performance and reliability.

[0084] Based on the above methods, this invention further proposes a multimodal drug interaction event prediction system based on compact representation learning, including a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the above-mentioned multimodal drug interaction event prediction method based on compact representation learning.

[0085] Furthermore, to evaluate the performance of the method of this invention on the drug interaction event prediction task, the performance was validated on the DDIMDL and MDF-SA-DDI datasets. The DDIMDL dataset contains 572 drugs, 74,528 drug interactions, and 65 drug interaction events; the MDF-SA-DDI dataset contains 1,258 drugs, 323,539 drug interactions, and 100 drug interaction events. The external biomedical knowledge graph contains 97,243 entities and 107 relationships. This invention uses ACC (precision), AUPR (area under the precision-recall curve), AUC (area under the receiver operating characteristic curve), F1 (macro-average F1 score), Pre (precision), and Rec (recall) to evaluate the performance of drug interaction event prediction.

[0086] This invention compares with seven representative state-of-the-art baselines, including traditional classification methods (RF and DNN), unimodal DDIE methods (DeepDDI), and multimodal DDIE methods (MDNN, HetDDI, MKG-FENN, and MolecBioNet). Experiments on predicting new drug-to-drug interactions were conducted on two datasets, where the new drugs only appeared in the test set. The results are shown in Table 1. This invention consistently outperforms all baseline methods, demonstrating the effectiveness of compact representation learning within and between modalities. In the more challenging "new drug" setting, this invention achieves a significant performance improvement of an average of 25.79%, highlighting its strong generalization ability on unseen drugs. Compared to traditional classification methods and unimodal methods, this invention integrates multimodal information sources, thus achieving a significant improvement of up to 200.91% on the Rec metric in the "new drug–new drug" setting of the DDIMDL dataset. Compared to multimodal methods, this invention explicitly learns more compact and informative drug embedding representations both intramodally and intermodally, resulting in a performance improvement of up to 60.51% compared to MolecBioNet.

[0087] Table 1. Experimental results (%) of the two datasets under the new drug setting.

[0088] (Best performance is indicated in bold, and second-best performance is indicated by underline).

[0089]

[0090] To demonstrate the advantages of this invention in learning compact multimodal drug characterization, the three modal embeddings generated by "this invention w / o. Intra+Inter" (removing the intramodal and intermodal compact mechanisms) and "the complete invention" were visualized. Furthermore, in Figure 3 The drug atorvastatin (corresponding to the □ symbol) and its corresponding multimodal information are specifically highlighted. Without a compact mechanism, the embedding distributions of biological, molecular, and knowledge graph modalities are fragmented, resulting in inconsistent drug characterization. For example, some biological targets of atorvastatin (such as APOB) introduce "off-target" signals related to lipid metabolism unrelated to its primary function; simultaneously, recurring aromatic nuclei and pyrrolidine motifs in the molecular structure interfere with modeling the cholesterol-lowering effects of its hydroxy acid groups. This redundant information makes the embedding distribution scattered, hindering an effective characterization of its lipid-lowering mechanism. In contrast, this invention produces a more compact embedding distribution while preserving key pharmacological details rich in information. For example, by filtering redundant peripheral proteins and irrelevant knowledge graph relationships, the embeddings of atorvastatin in both biological and knowledge graph modalities consistently highlight its key targets. The resulting partial overlap between these two modalities aligns with atorvastatin's lipid regulation mechanism, providing a unified characterization for accurate prediction of drug interaction events.

[0091] This invention combines multi-source information such as biological features, molecular structure, and knowledge graphs to significantly reduce redundancy within and between modalities, obtaining a consistent and compact drug representation, thereby more comprehensively and accurately characterizing the pharmacological properties of drugs; it effectively mines complementary information between different modalities, highlighting core pharmacological signals such as key targets, pathways, and functional groups, making the prediction results of drug interaction events more interpretable and improving prediction performance and reliability.

[0092] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0093] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0094] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0096] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive other forms of multimodal drug interaction event prediction methods and systems based on compact representation learning. All equivalent changes and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A multimodal drug interaction event prediction method based on compact representation learning, characterized in that, include: Obtain the multimodal features of each drug in the drug pair to be predicted, wherein the multimodal features include at least two of the following: biological features, molecular structure features, and knowledge graph features; For each modal feature of each drug, intramodal compact representation learning is performed to eliminate redundant information within a single modality and generate compact sub-representations for each modality. Using compact sub-representations corresponding to molecular structure features as anchors, cross-modal alignment of compact sub-representations corresponding to biological features and knowledge graph features is performed through mutual information minimization constraints to eliminate redundant information between modalities. A multimodal fusion representation of each drug is constructed based on the compact sub-representation of each modality. The multimodal fusion representations of the two drugs in the drug pair are then concatenated and input into a classifier to output the prediction results of the interaction event types of the drug pair.

2. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 1, characterized in that: The intramodal compact representation learning includes: Biological features and molecular structural features were reconstructed and optimized using an autoencoder, respectively. After learningable pruning of the relation weights in the knowledge graph features, the knowledge compact sub-representation is generated by processing the heterogeneous graph neural network.

3. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 2, characterized in that: The learnable pruning includes: Learnable pruned weights are generated based on the original relation weights. By combining random noise and temperature parameters, the pruned relation weight distribution is generated through differentiable sampling technology. The difference between the pruned relation weight distribution and the original relation weight distribution is constrained by KL divergence.

4. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 1, characterized in that: The multimodal fusion representation is generated by vector splicing of biological compaction representation, molecular compaction representation and knowledge compaction representation.

5. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 4, characterized in that: The process of splicing the multimodal fusion representations of the two drugs in the drug pair specifically involves connecting the multimodal fusion representations of the two drugs in a predetermined dimensional order.

6. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 1, characterized in that: The method also includes constructing an overall loss function for end-to-end training, wherein the overall loss function includes classification loss, intra-modal compaction loss and inter-modal compaction loss.

7. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 6, characterized in that: The intermodal compactness loss is the sum of the mutual information values ​​between biological compaction representations and molecular compaction representations, and between molecular compaction representations and knowledge compaction representations.

8. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 1, characterized in that: The multimodal features simultaneously include biological features, molecular structural features, and knowledge graph features.

9. The method for predicting multimodal drug interaction events based on compact representation learning according to claim 1, characterized in that: At least one drug in the drug pair to be predicted did not appear during the model training phase.

10. A multimodal drug interaction event prediction system based on compact representation learning, characterized in that, It includes a processor, a memory, and a computer program stored in the memory, wherein when the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 9.

Images