Residual cell optimization based super network representation learning method applied to breeding

Abstract

The application discloses a kind of based on residual unit optimization's super-network representation learning method applied to breeding, it is related to crop breeding technical field, method includes: constructing crop heterogeneous super-network, obtaining the representation vector of each node, and obtaining at least one to be modeled high-order tuple, the node representation sequence corresponding to each to be modeled high-order tuple is input into target model, the similarity of each to be modeled high-order tuple is obtained;Based on the similarity of each to be modeled high-order tuple, the representation vector of each node in crop heterogeneous super-network is optimized;Using the representation vector of all nodes after optimization, the prediction of crop variety is carried out.The application solves the problem that the existing method is insufficient in super-edge fusion, thereby improving the depth and fidelity of learning representation from high-order structure information, and while ensuring the ability to model complex relationships, the operability and pertinence of the method in real breeding scenarios are improved.

Description

Technical Field

[0001] This invention relates to the field of crop breeding technology, and in particular to a hypernetwork representation learning method based on residual unit optimization applied to breeding. Background Technology

[0002] In the field of crop breeding, a key technical challenge lies in how to accurately model the complex high-order relationships between varieties and traits, as well as between traits themselves, from massive, multi-dimensional phenotypic data, thereby predicting new varieties with ideal trait combinations. Traditional methods struggle to effectively handle these high-order interactions that go beyond pairwise relationships, hindering the progress of intelligent breeding. Hypernetwork representation learning technology offers a potential solution to this problem. Its goal is to embed nodes representing phenotypic traits into a low-dimensional vector space while preserving the complex high-order structural information defined by the variety, providing a computationally achievable representational basis for downstream variety prediction and design.

[0003] To address this challenge, existing technologies have developed various hypernetwork representation learning methods, which can be tested in breeding data analysis scenarios. Matrix factorization-based methods decompose the variety-trait association matrix to learn low-dimensional representations, such as hypergraph Laplacian mappings. Random walk-based methods design walk strategies on the variety hypernetwork, generate sequences, and then use language models to learn representations, such as Hyper2vec. Deep learning-based methods further utilize the powerful nonlinear fitting capabilities of neural networks. Graph neural network-based methods typically transform the hypernetwork into a regular graph and then apply standard GNNs, or design specialized hypergraph neural networks such as HyperGCN for message aggregation. Autoencoder-based methods learn representations by reconstructing node adjacency or hyperedge information. Furthermore, attention-based methods, such as HyperGAT, attempt to assign differentiated weights to nodes with different traits during information aggregation.

[0004] However, applying existing methods to crop breeding scenarios reveals several inherent drawbacks. First, there is insufficient modeling of high-order heterogeneity. Most existing methods treat all phenotypic traits constituting a variety as equally important or perform simple weighting, ignoring the fundamental differences in the contribution of different traits to a variety in real-world breeding. This results in the learned trait representation vectors losing their specific role information in different variety contexts. Second, there are issues of high computational complexity and poor scalability. Breeding data may involve a large number of varieties and traits. Especially when the number of certain related traits is large, deep learning methods based on complex aggregation operations face computational and memory bottlenecks, making them difficult to apply to large-scale breeding databases. Finally, there is insufficient hyperedge fusion. Many methods simply decompose the hyperedge representing a variety into multiple paired trait relationships. This approach fragments the variety characteristics defined by the trait combination as a whole, failing to fully capture the core high-order semantic of "multiple specific traits jointly determining a variety," thus limiting the accuracy of model predictions.

[0005] Therefore, existing hypernetwork representation learning methods have limitations in terms of accuracy, efficiency, and information fusion depth, addressing the urgent need for modeling high-order, heterogeneous relationships in crop breeding. There is a pressing need for a novel representation learning technique that can deeply integrate high-order tuple information, efficiently handle heterogeneous relationships, and is specifically adapted to the characteristics of breeding data, in order to more accurately serve the prediction and design of ideal crop varieties. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of the prior art, and the specific technical solution is as follows: 1) In a first aspect, the present invention provides a hypernetwork representation learning method based on residual unit optimization applied to breeding, the specific technical solution of which is as follows: Construct a heterogeneous supernetwork for crops, where nodes represent phenotypic traits of crop varieties and superedges represent crop varieties; Obtain the representation vector of each node in the heterogeneous supernetwork of crops, and obtain at least one higher-order tuple to be modeled, wherein each hyperedge corresponds to a higher-order tuple to be modeled. Input the node representation sequence corresponding to each higher-order tuple to be modeled into the target model to obtain the similarity of each higher-order tuple to be modeled; Based on the similarity of each higher-order tuple to be modeled, the representation vector of each node in the heterogeneous supernetwork of crops is optimized; Crop varieties are predicted using the optimized representation vectors of all nodes.

[0007] The beneficial effects of the supernetwork representation learning method based on residual unit optimization for breeding provided by this invention are as follows: By directly processing the higher-order tuples to be modeled and their node representation sequences corresponding to each hyperedge, and using the target model to calculate the similarity of the higher-order tuples to be modeled, the shortcomings of existing methods in modeling higher-order heterogeneity are effectively overcome. This allows the model to more precisely characterize the combination relationships and internal structure of nodes of different phenotypic traits within a specific crop variety. By iteratively optimizing the node representation vector based on the similarity of the higher-order tuples to be modeled, and using the optimized node representation vector for crop variety prediction, the technical solution achieves full fusion of the overall trait combination information represented by the hyperedge, solving the problem of insufficient hyperedge fusion in existing methods, thereby improving the depth and fidelity of learning representations from higher-order structural information. In addition, the entire method is closely designed around the data characteristics of the crop breeding field, constructing a heterogeneous supernetwork for crops and ultimately serving variety prediction. While ensuring the ability to model complex relationships, it improves the operability and relevance of the method in real breeding scenarios.

[0008] Based on the above scheme, the supernetwork representation learning method based on residual unit optimization applied to breeding, as proposed in this invention, can be further improved as follows.

[0009] Furthermore, the target model sequentially includes a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function. The one-dimensional convolutional network is used to extract basic features from the node representation sequence, the enhanced feature extraction residual unit is used to enhance the basic features into enhanced features, and the fully connected layer and activation function are used to calculate the similarity of the higher-order tuples to be modeled based on the enhanced features.

[0010] The beneficial effects of adopting the above-mentioned further approach are as follows: By clearly defining the target model as consisting of a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function in sequence, this method establishes a hierarchical feature processing pipeline. The one-dimensional convolutional network first extracts basic features from the node representation sequence. The enhanced feature extraction residual unit then performs deep enhancement on the basic features to capture complex patterns. Finally, the fully connected layer and activation function calculate the similarity of the higher-order tuples to be modeled based on the enhanced features. This structured design enables the model to systematically learn and fuse multi-level information from sequential input, thereby achieving a stable and accurate evaluation of the possibility of higher-order tuple relationships.

[0011] Furthermore, the process by which the enhanced feature extraction residual unit processes the basic features into enhanced features includes: performing a first convolution operation on the input basic features to obtain a first intermediate feature; performing a first batch normalization and a first activation function on the first intermediate feature to obtain a second intermediate feature; performing a second convolution operation on the second intermediate feature to obtain a third intermediate feature; performing a second batch normalization and a second activation function on the third intermediate feature to obtain a fourth intermediate feature; and performing a residual connection between the fourth intermediate feature and the basic features to obtain the enhanced feature.

[0012] The beneficial effects of adopting the above-mentioned further scheme are as follows: By specifically defining the process of transforming basic features into enhanced features through the enhanced feature extraction residual unit, this process includes two convolutions, batch normalization, activation function processing, and one residual connection. This design allows features to retain information of the original basic features through residual connections while undergoing nonlinear transformation and deep extraction. This operation not only prevents gradient decay or feature degradation problems that may occur in deep network training, but also achieves effective fusion of shallow basic features and deep abstract features, thereby enhancing the representational power and robustness of the final enhanced features.

[0013] Furthermore, based on the similarity of each higher-order tuple to be modeled, the representation vector of each node in the crop heterogeneous supernetwork is optimized, including: constructing a tuple relation loss function based on the similarity of all higher-order tuples to be modeled and their corresponding true labels; constructing a pairwise relation loss function based on the representation vectors of nodes in the crop heterogeneous supernetwork; combining the tuple relation loss function and the pairwise relation loss function with preset weights to obtain a joint loss function; and updating the representation vector of each node by minimizing the joint loss function.

[0014] The beneficial effects of adopting the above-mentioned further scheme are as follows: By constructing a tuple relation loss function and a pairwise relation loss function, and combining them into a joint loss function according to preset weights to optimize the node representation vector, this method achieves collaborative learning of higher-order tuple relations and pairwise relations. The tuple relation loss function drives the representation vector to learn the overall trait combination pattern of the variety, while the pairwise relation loss function strengthens the local association between traits. By minimizing the joint loss function of the two to update the node representation vector, the final representation vector can simultaneously encode the higher-order overall semantics and local pairwise similarity of the network, thereby obtaining a more comprehensive and higher-quality node embedding representation.

[0015] 2) In a second aspect, the present invention also provides a hypernetwork representation learning system based on residual unit optimization for use in breeding, the specific technical solution of which is as follows: It includes a construction module, an acquisition module, a similarity acquisition module, an optimization module, and a prediction module; The construction module is used to: construct a heterogeneous supernetwork of crops, where nodes represent phenotypic traits of crop varieties and superedges represent crop varieties; The acquisition module is used to: acquire the representation vector of each node in the heterogeneous supernetwork of crops, and acquire at least one high-order tuple to be modeled, wherein each hyperedge corresponds to one high-order tuple to be modeled; The similarity acquisition module is used to: input the node representation sequence corresponding to each higher-order tuple to be modeled into the target model, and obtain the similarity of each higher-order tuple to be modeled; The similarity acquisition module is used to optimize the representation vector of each node in the heterogeneous supernetwork of crops based on the similarity of each high-order tuple to be modeled. The prediction module is used to predict crop varieties using the optimized representation vectors of all nodes.

[0016] Based on the above scheme, the supernetwork representation learning system based on residual unit optimization applied to breeding, as proposed in this invention, can be further improved as follows.

[0017] Furthermore, the target model sequentially includes a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function. The one-dimensional convolutional network is used to extract basic features from the node representation sequence, the enhanced feature extraction residual unit is used to enhance the basic features into enhanced features, and the fully connected layer and activation function are used to calculate the similarity of the higher-order tuples to be modeled based on the enhanced features.

[0018] Furthermore, the process by which the enhanced feature extraction residual unit processes the basic features into enhanced features includes: performing a first convolution operation on the input basic features to obtain a first intermediate feature; performing a first batch normalization and a first activation function on the first intermediate feature to obtain a second intermediate feature; performing a second convolution operation on the second intermediate feature to obtain a third intermediate feature; performing a second batch normalization and a second activation function on the third intermediate feature to obtain a fourth intermediate feature; and performing a residual connection between the fourth intermediate feature and the basic features to obtain the enhanced feature.

[0019] Furthermore, the optimization module is specifically used to: construct a tuple relation loss function based on the similarity of all higher-order tuples to be modeled and their corresponding true labels; construct a pairwise relation loss function based on the representation vectors of nodes in the crop heterogeneous supernetwork; combine the tuple relation loss function and the pairwise relation loss function with preset weights to obtain a joint loss function; and update the representation vector of each node by minimizing the joint loss function.

[0020] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device implements any of the above-mentioned supernetwork representation learning methods based on residual unit optimization applied to breeding.

[0021] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for learning supernetwork representations based on residual unit optimization applied to breeding.

[0022] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a flowchart illustrating a supernetwork representation learning method based on residual unit optimization applied to breeding, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a heterogeneous supernetwork. Figure 3 A schematic diagram of the network structure of the enhanced feature extraction residual unit; Figure 4 This is a schematic diagram of the network structure of the HyReSim neural network model; Figure 5 The reconstruction results are for the GPS supernetwork reconstruction. Figure 6 The reconstruction results are for the drug hypernetic network. Figure 7 This is a schematic diagram of the structure of a supernetwork representation learning system based on residual unit optimization applied to breeding, according to an embodiment of the present invention. Detailed Implementation

[0024] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0025] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] like Figure 1 As shown in the figure, an embodiment of the present invention provides a supernetwork representation learning method based on residual unit optimization applied to breeding, comprising the following steps: S1. Construct a heterogeneous supernetwork for crops, where nodes represent phenotypic traits of crop varieties and hyperedges represent crop varieties. The specific implementation process is as follows: S10. Systematically collect a series of standardized phenotypic trait measurement data for crop varieties from historical breeding records, variety approval databases, or field trials. For example, collect... Each crop variety has complete records, and measurements were taken for each variety. These are key phenotypic traits. These data constitute the original input for building the network.

[0027] S11, Define the node set This instantiates the measured value or classification state of each specific phenotypic trait of each crop variety as a node in the hypernetwork. Since phenotypic traits have different categories, the node set consists of multiple types of nodes. Assume that a total of [number] nodes are defined. There are different types of phenotypic traits, such as morphological traits, yield traits, and quality traits. Then, the set of nodes... It is the union of all types of node instances, that is .here, This represents the set of nodes in the entire hypernetwork. Indicates the first The set of all node instances of a trait type. Is greater than or equal to The positive integers ensure the heterogeneity of the network. Each specific node uses... This indicates that it necessarily belongs to a specific... To more clearly represent trait nodes, we can denote them as follows: For the first The first type of trait A specific instance node, for example It can represent a specific node of the "plant height" trait. It can represent a specific "thousand-grain weight" trait node.

[0028] S12. Each hyperedge corresponds to a specific crop variety. For a given crop variety, the nodes corresponding to all its phenotypic traits are aggregated to form a hyperedge. Assume there are a total of... For each crop variety, the hyperedge set is... Include A super edge, denoted as Each hyperedge It is a set of nodes, that is ,in Indicates the first The number of trait nodes associated with each variety, and , These are trait nodes specifically belonging to this variety. These nodes originate from defined sets of nodes of different types. Therefore, a superedge typically contains multiple types of nodes.

[0029] S13, the heterogeneous node set defined in S21 and the set of superedges defined by S22 Combining to form a complete heterogeneous supernetwork of crops Its mathematical representation is This network It accurately depicts the complex relationship between crop varieties and multiple types of phenotypic traits, providing a structured data foundation for subsequent representation learning.

[0030] Among them, the heterogeneous hypernetwork of crops is a mathematical graph model used to characterize complex relationships in the field of crop breeding. In the heterogeneous hypernetwork of crops, nodes represent phenotypic traits of different categories of crop varieties, and hyperedges represent specific crop varieties. The heterogeneity of the network is reflected in the diversity of node types, i.e., the set of nodes. Depend on Different types of subsets Composition, in which Greater than or equal to The formal definition of this network is... It integrates the multidimensional trait information contained in a variety into a high-order connection structure, thereby enabling the more natural expression of breeding knowledge that transcends pairwise relationships.

[0031] In a hypernetwork, a hyperedge is a basic connection unit used to represent a higher-order coexistence relationship between a group of nodes. In a crop heterogeneous hypernetwork, each hyperedge corresponds to a specific crop variety. Defined as a subset of nodes, i.e. ,in It is the number of nodes contained in the superedge, and These nodes represent the various phenotypic traits exhibited by the variety. Unlike edges in a regular graph, a hyperedge can connect two or more nodes simultaneously, thus directly encoding the holistic information that "multiple traits jointly define the same variety."

[0032] The specific crop varieties can be: rice variety "Nanjing 46", corn variety "Zhengdan 958", wheat variety "Jimai 22", or Jinjing 818, etc., and can be set according to actual conditions. For example, a rice variety named "Jinjing 818" does not exist directly as a node in the hypernetwork, but is transformed into a hyperedge. This hyperedge will link to a series of nodes describing its phenotypic traits, such as "plant height 95 cm", "ear length 18 cm", and "resistance to rice blast", thus encapsulating the multivariate information of this variety in a high-order structure.

[0033] In this context, phenotypic traits of crop varieties refer to the observable and measurable morphological, physiological, and biochemical characteristics of the variety. In a heterogeneous supernetwork of crops, each specific instance of a phenotypic trait constitutes a node. These traits are categorized into different types. For example, morphological trait types might include specific nodes such as "plant height," "leaf color," and "ear type"; yield trait types might include specific nodes such as "number of ears per plant," "number of grains per ear," and "thousand-grain weight"; and quality trait types might include specific nodes such as "protein content" and "amylose content." Each such specific trait description, after acquiring a measurement value or state, becomes a node in the network. and belong to its corresponding type subset. .

[0034] S2. Obtain the representation vector of each node in the heterogeneous supernetwork of crops, and obtain at least one high-order tuple to be modeled, where each hyperedge corresponds to one high-order tuple to be modeled. The specific implementation process is as follows: S20. At the beginning of model training, it is necessary to prepare the crop heterogeneous supernetwork. Each node in the set is assigned an initial, low-dimensional real-valued vector representation. Let the node set be... The CCP There are n nodes, and the expected representation vector dimension is 1. The initialization process is for each node. One is generated randomly or assigned according to a specific strategy. A dimensional vector is used as its initial representation, denoted as . .here, Representing the first in the supernetwork 1 node It is a path from a node to its The dimension represents the mapping function of the vector. This is a preset hyperparameter. Because the network is heterogeneous, the nodes belong to... Different types, namely ( (This is a node type index). In practice, a common implementation technique is to index each node type... Maintain an independent trainable embedding matrix ,in It is the first The total number of nodes of each class. Then, for a specific node... If it belongs to the first Type and in The row index in is Then its initial representation vector is In this way, all nodes obtain an initial representation vector that can be used for subsequent optimization and adjustment of the model.

[0035] S21. The higher-order tuples to be modeled are directly derived from the set of superedges of the constructed crop heterogeneous supernetwork. Each hyperedge corresponds to a higher-order tuple to be modeled. Therefore, the hyperedge set is traversed. For each hyperedge in the array, the set of nodes it contains is transformed into an ordered sequence of nodes, which constitutes a higher-order tuple to be modeled. Let the hyperedge currently being processed be... ,in Indicates belonging to a superedge The 1 node This represents the number of nodes contained in the hyperedge. Arranging this set of nodes according to a certain rule yields a higher-order tuple to be modeled. , recorded as .here, Indicates the first A set of higher-order tuples to be modeled It is a permutation function used to determine the order of nodes in a tuple, which can be defined according to the order in which nodes were added to the network, the lexicographical order of node types, or other fixed rules. Ultimately, it can obtain at least one (actually) node from the entire hypernetwork. (Number) higher-order tuples to be modeled constitute the set of higher-order tuples to be modeled. .

[0036] S22. Having obtained the initial representation vectors of all nodes and defined the set of higher-order tuples to be modeled. Next, it is necessary to model each higher-order tuple. Prepare to input data. Specifically, according to... In the order of the nodes, extract each node in turn. The current representation vector These vectors are concatenated along their feature dimensions to form a longer vector sequence, called the node representation sequence of the tuple. Mathematically, for a tuple... Its nodes represent sequences It can be represented as: Here, the semicolon indicates a concatenation operation of vectors along the feature dimension. The dimension is If the number of nodes on different superedges Inconsistencies usually require standardization processes such as padding or truncation to ensure that subsequent one-dimensional convolutional networks can handle fixed-length inputs, or to use convolutional strategies that can handle variable-length inputs.

[0037] In this context, a higher-order tuple to be modeled refers to a combination of higher-order relations that needs to be specifically learned and modeled by the neural network model during the hypernetwork representation learning process. In the context of a heterogeneous crop hypernetwork, a higher-order tuple to be modeled corresponds to a specific hyperedge, i.e., a specific crop variety. It is formally defined as an ordered sequence. Each element It is a phenotypic trait node belonging to this superedge. This tuple captures the higher-order structural information that "multiple specific trait nodes appear simultaneously in a single variety". The goal of the model is to learn so that the representation vector can effectively encode this higher-order coexistence pattern that transcends pairwise relationships, thereby enabling the computation of the probability or similarity of the relationship represented by the tuple.

[0038] S3. Input the node representation sequence corresponding to each higher-order tuple to be modeled into the target model to obtain the similarity of each higher-order tuple to be modeled. The target model consists of a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function. The one-dimensional convolutional network is used to extract basic features from the node representation sequence, the enhanced feature extraction residual unit is used to enhance the basic features into enhanced features, and the fully connected layer and activation function are used to calculate the similarity of the higher-order tuples to be modeled based on the enhanced features.

[0039] The process by which the enhanced feature extraction residual unit processes the basic features into enhanced features includes: performing a first convolution operation on the input basic features to obtain a first intermediate feature; performing a first batch normalization and a first activation function on the first intermediate feature to obtain a second intermediate feature; performing a second convolution operation on the second intermediate feature to obtain a third intermediate feature; performing a second batch normalization and a second activation function on the third intermediate feature to obtain a fourth intermediate feature; and performing a residual connection between the fourth intermediate feature and the basic features to obtain the enhanced features.

[0040] S3 is described as follows: S30. Train the HyReSim neural network model to obtain the target model. The HyReSim neural network model is a deep learning model specifically designed for processing heterogeneous hypernetwork data of crops. The model's layer structure, in order of information processing, is as follows: one-dimensional convolutional network, enhanced feature extraction residual unit, fully connected layer, and activation function. The purpose of the training process is to optimize the model parameters and node representation vectors so that the model can accurately calculate the similarity of the higher-order tuples to be modeled and the pairwise relationships between nodes. The entire training process is iterative. In each round of training, a batch of positive higher-order tuples to be modeled (i.e., real hyperedges) and negative higher-order tuples to be modeled (generated by destroying positive tuples) are sampled from the dataset. Their node representation sequences are input into the model to obtain the predicted similarity. At the same time, the pairwise relationship loss is calculated based on the node representation vector. Then, the tuple relationship loss function is calculated separately. Pairwise loss function According to the preset weight parameters The two loss functions are combined into a joint loss function. The joint loss function is calculated using the backpropagation algorithm. Regarding all model parameters (including convolutional kernel weights, fully connected layer weights, batch normalization parameters, etc.) and node representation vectors The gradients are then calculated. Finally, stochastic gradient descent or its variant optimizer is used to update these parameters and vectors. After multiple rounds of iterative training, when the loss function converges or reaches the preset number of training rounds, the model parameters are fixed, and the trained model at this point is the target model that can be used to calculate similarity.

[0041] S31. Obtaining the node representation sequence is a necessary data preparation process before model inference. For each higher-order tuple to be modeled... ,in, Represents the first tuple in the set. 1 node Given the order of the tuple (i.e., the number of nodes it contains), we need to extract its corresponding representation vector. For a tuple... Each node in Retrieve its corresponding dimensional representation vector Then, strictly follow the order of nodes in the tuple. The order in which the node representations are concatenated along the feature dimension is determined to form the node representation sequence of the tuple. This sequence can be formally represented as The semicolon indicates that the vectors are stacked in the column direction. For all higher-order tuples to be modeled in the network, repeating the above retrieval and concatenation operations yields the corresponding set of node representation sequences, which serve as the standard input to the target model.

[0042] S32. A one-dimensional convolutional network is the first component in the target model that processes serialized input. Its function is to represent the sequence from the concatenated nodes. This involves extracting preliminary feature representations with local perception capabilities, i.e., basic features. Specifically, this involves processing the sequence... Consider it as a length of The number of channels is The input is a one-dimensional signal. A one-dimensional convolutional network uses multiple one-dimensional convolutional kernels to perform sliding convolution calculations on this sequence. Let the weight vector of one convolutional kernel be... ,in It is the size of the convolution kernel (receptive field), which is in the sequence. Slide upwards, calculating the local value each time. The representation vector of each node is multiplied by the convolution kernel, and then passed through a non-linear activation function to generate a new feature value. After traversing the entire sequence, a new feature map is generated. By using multiple different convolution kernels, a one-dimensional convolutional network can extract basic features of various patterns from the original node representation sequence. Assuming that a... With a convolutional kernel and appropriate padding to maintain sequence length, the output of a one-dimensional convolutional network, i.e., the basic features, can be represented as a new feature matrix. .here, This represents the extracted basic feature matrix. It is the sequence length. It refers to the number of channels or dimensions of the basic features.

[0043] S33, Enhanced Feature Extraction Residual Unit receives basic features from a one-dimensional convolutional network. As input, and through a deep processing path containing residual connections, it outputs more expressive enhanced features. This process is executed strictly according to a defined sequence of sub-operations, specifically: 1) Basic features of the input Perform the first convolution operation. This convolution operation uses a set of convolution kernels to process the basic features. Further feature transformations are performed to obtain the first intermediate feature. .here, This indicates the first convolution operation. This represents the first intermediate feature obtained.

[0044] 2) For the first intermediate feature Perform the first batch normalization and the first activation function processing. Batch normalization... Each feature channel is standardized to stabilize the training process; then an activation function is applied to introduce nonlinearity, resulting in the second intermediate feature. .in, This indicates the first batch of normalization. This represents the activation function (such as ReLU). This represents the obtained second intermediate feature.

[0045] 3) Regarding the second intermediate feature Perform a second convolution operation. Use another set of convolution kernels... Convolution is performed on the top layer to obtain the third intermediate feature. .here, This indicates the second convolution operation. This represents the obtained third intermediate feature.

[0046] 4) Regarding the third intermediate feature A second batch normalization and a second activation function processing are performed. and After processing, the fourth intermediate feature is obtained. .in, This indicates the second batch of normalization. This represents the fourth intermediate feature obtained.

[0047] 5) The fourth intermediate feature With basic features Perform residual joins. This is done by... With the original input This is achieved by adding elements one by one to obtain the final enhanced feature. .here, This represents the enhanced features output by the enhanced feature extraction residual unit. This residual connection ensures that while feature extraction is deepened, the original basic feature information is not lost or degraded as the network depth increases.

[0048] S34. Calculating the similarity of the higher-order tuples to be modeled is the final output stage of the target model. First, the enhanced features are extracted from the output of the residual unit. Perform appropriate shaping, such as flattening it into a one-dimensional feature vector. ,in This is the flattened feature dimension. Then, this feature vector... The input is passed to a fully connected layer. The fully connected layer performs a linear transformation. ,in It is the weight matrix of the fully connected layer. It is a bias vector. This linear transformation will transform high-dimensional features. This is mapped to a scalar value. Then, an activation function is applied to this scalar value. Here, we use... Activation function . The function compresses the output of the linear transformation to Within the interval. This compressed scalar value. This refers to the higher-order tuples to be modeled as predicted by the model. The similarity is expressed by the formula: The similarity It can be interpreted as a tuple The probability or confidence score of the specific higher-order relationship (i.e., the corresponding combination of varieties and traits) is closer to 1, indicating that the model believes the relationship of the tuple is more likely to be true.

[0049] The Enhanced Feature Extraction Residual Unit (ERU) is a core component of the HyReSim neural network model, a specially designed residual network module. It consists of two consecutive convolutional operations, two batch normalization processes, two activation function applications, and a cross-layer identity residual connection path. This unit takes the basic features extracted by the previous layer as input and learns incremental transformations or supplementary information of the input features through its internal deep convolutional paths. The final operation within the unit is to directly add the learned incremental features to the original input basic features, thus obtaining the enhanced features. This design allows the unit to effectively fuse shallow basic features with deep abstract features, improving the model's ability to model complex high-order tuple relationships while mitigating gradient vanishing or information decay problems that may occur during deep network training.

[0050] Here, the similarity of the higher-order tuples to be modeled is the similarity of the HyReSim neural network model for a specific higher-order tuple to be modeled in the input. A quantized score is calculated. This score is a real number between 0 and 1, obtained through the fully connected layer at the end of the model. Activation function generation. Numerical similarity. This characterizes how the model determines the tuple based on the currently learned node representation vectors. This is a probability estimate of whether the multiple nodes (i.e., multiple phenotypic traits) constitute a true, reasonable higher-order combination (i.e., whether it belongs to a true species). In link prediction tasks, this similarity is directly used to assess the probability of the existence of unknown hyperedges; in model training, it is compared with the true labels of tuples. (1 for existence, 0 for non-existence) are compared to calculate the tuple relation loss. This drives the optimization of model parameters.

[0051] S4. Based on the similarity of each higher-order tuple to be modeled, the representation vector of each node in the heterogeneous supernetwork of crops is optimized. Specifically: S40. Based on the similarity of all the higher-order tuples to be modeled and their corresponding true labels, construct the tuple relation loss function. The specific implementation process is as follows: 1) In a training batch, assuming a total of [number] samples were taken. A batch set of higher-order tuples to be modeled. For each higher-order tuple to be modeled in the set. The model's forward propagation process outputs a predicted similarity value. As before, ,in This is the flattened vector of the enhanced feature corresponding to the tuple. Simultaneously, each higher-order tuple to be modeled... Each is accompanied by a pre-known, authentic label based on its source. This label is a binary scalar indicating whether the tuple represents a real higher-order relation. Therefore, for a batch of data, we obtain two tuples of length [missing information]. Vector: Model-predicted similarity vector and the corresponding real label vector .

[0052] 2) For a specific higher-order tuple to be modeled Its tuple relation loss Define the true label of the tuple. The model's predicted similarity value The product of. Here, That is, the similarity predicted by the model. Therefore, it can be equivalently understood as: in, It's a real label. It is the similarity predicted by the model. This definition means that for a real-world tuple ( The desired similarity predicted by the model. Make it as large as possible, so that As large as possible; for a non-existent negative tuple ( Regardless of the similarity predicted by the model The loss contribution is always zero regardless of the value. To transform it into a standard loss function that needs to be minimized, the formula is usually modified. A common and equivalent technique is to use the idea of ​​negative log-likelihood. The goal is to maximize the likelihood of the true tuple (i.e., ...). ), while minimizing the likelihood of negative example tuples. This can be achieved by defining a loss of the following form: for positive examples ( The loss is For negative examples ( The loss is This is exactly the form of the binary cross-entropy loss. Therefore, the loss of a single higher-order tuple to be modeled... This can be specifically implemented as follows: here, Represents a single tuple The resulting loss value, That is its true label. It is the similarity predicted by the model. It is the natural logarithm function. This formula ensures that when predicting... The loss is small when the label is close to the true label, and large when it deviates from the true label.

[0053] 3) To obtain an overall tuple relation loss function value that represents the performance of the current model across the entire batch of data, it is necessary to calculate the individual loss of each tuple. Summarize. The most direct and commonly used aggregation method is to calculate the arithmetic mean. Therefore, the batch tuple relation loss function... The structure is as follows: in, This represents the constructed tuple relation loss function value based on the current batch. It refers to the batch size. It is the index of the tuple within the batch. It is the first A set of higher-order tuples to be modeled It is its corresponding real label. It is the similarity between the model's predictions and the model's. This refers to the tuple relation loss function value that ultimately needs to participate in subsequent joint optimization, i.e. The specific implementation and calculation values ​​at the batch level.

[0054] 4) In actual code implementation, due to yes The function's output is calculated directly. or exist When the value is very close to 0 or 1, it may cause underflow or result in an infinite value. Therefore, numerically stable implementations are usually adopted when constructing loss functions, such as directly using the binary cross-entropy loss function provided by the framework, which typically incorporates... Logarithmic operations, through equivalent mathematical transformations, prevent intermediate values ​​from falling into unstable regions. This ensures that the constructed loss function is technically robust and achievable.

[0055] Here, the true label corresponding to the high-order tuple to be modeled is a binary supervision signal used to indicate whether the high-order tuple to be modeled corresponds to a real hyperedge in the crop heterogeneous supernetwork. For a high-order tuple to be modeled Its real label Satisfy: If If it is a real variety-trait combination in the dataset (i.e., a real hyperedge), then ;like If it is a non-existent spurious combination generated through techniques such as negative sampling, then... During model training, these real labels serve as "standard answers" and are compared with the similarity of the model's predicted higher-order tuples to be modeled. The difference between the two is quantified into a tuple relation loss function, thereby driving the model to learn how to distinguish between real and fake higher-order relations and update the node representation vectors.

[0056] S41. Based on the representation vectors of nodes in the heterogeneous supernetwork of crops, construct a pairwise relation loss function. The specific implementation process is as follows: 1) The foundation for constructing the pairwise relation loss function is defining what constitutes a valid "pairwise relation." In the context of hypernetworks, this relation is typically defined through random walks or based on hyperedge co-occurrence. Specifically, for crop heterogeneous hypernetworks... A series of central nodes were obtained through sampling. Its context node Pairing. For example, a random walk can be performed along a superedge: starting from a node Start by randomly jumping to another node within the same superedge. ,So It is considered A context node, and vice versa. Suppose that a context node containing [a context node] is obtained through sampling. Training set of center-context node pairs ,in It is the central node. It is its context node. These pairings capture local, paired proximity between nodes in the network.

[0057] 2) For each center-context node pair A model based on node representation vectors is needed to evaluate the likelihood of their co-occurrence, given a center node. Observe the context node The conditional probability is defined as: In this formula, Indicates at a given central node Under the condition, node As the probability of its occurrence in its context. and These are the central nodes. and context nodes of Dimension represents a vector. Dot product. It measures the similarity between two vectors; the higher the similarity, the higher the numerator. The larger the value of , the better. It is a normalization term, or partition function, that applies to all possible nodes in the network. (Usually approximated by negative sampling in actual calculations) and the center node The similarity indices are summed to ensure that the output is an effective probability distribution. (Symbol) This represents the natural exponential function.

[0058] 3) It is computationally infeasible because the denominator involves operations on all nodes. The summation is computationally very expensive. The standard implementation uses a negative sampling strategy. For each positive example center-context pair... ,sampling One does not agree Nodes that constitute a contextual relationship are considered negative samples, denoted as... So, for this positive example pair and... One negative example, constructing the loss function (for a single center node). Using the form of binary cross-entropy: here, It is the loss generated by a single positive pair and its corresponding negative sample. It's the sigmoid function. The first term... This prompts the positive example context node With the central node The vector dot product should be as large as possible (i.e. (Approximately 1). Second item Promote negative sample nodes With the central node The vector dot product should be as small as possible (i.e. Approaching 1 means (negative and with a large absolute value). It represents the number of negative samples, and is a hyperparameter.

[0059] 4) For a training batch containing multiple center-context pairs, all individual losses need to be aggregated to form a single scalar loss value for gradient backpropagation. Assume the current training batch contains... Positive example center-context pair And sampled for each positive example. negative sample nodes Therefore, the pairwise loss function for this batch... Constructed as the average of all individual losses: Substituting the formula from the previous step, we get the specific expression: In this formula, This refers to the constructed pairwise loss function value based on the current batch. It is the number of positive pairs in the batch. It is the index of the positive pairs. and It is the first The central node and context node in a pair of positive examples It is the negative sampling quantity. It is a negative sample index. This is the first The positive example is the first sampled... One negative sample node. Function Used to obtain nodes Dimension represents a vector. that is A computable version implemented through negative sampling techniques in actual training will be used for subsequent joint optimization with the tuple relation loss function.

[0060] S42. Combine the tuple relation loss function and the pairwise relation loss function according to preset weights to obtain the joint loss function. The specific implementation process is as follows: 1) Before performing the combination operation, ensure that the tuple relation loss function and the pairwise relation loss function have been independently calculated based on the data of the current training batch. As shown in the previous process, the value of the tuple relation loss function is denoted as... It quantifies the model's error in predicting the similarity of higher-order tuples to be modeled. The value of the pairwise relation loss function is denoted as... It quantifies the model's error in characterizing pairwise co-occurrence relationships between nodes. These two loss values... and These are all scalars, and have usually been batch averaged, making them comparable and preparing them for subsequent weighted combinations.

[0061] 2) The combination process involves setting a parameter to control the proportion of the two loss functions. This parameter is the preset weight, denoted as... . It is a hyperparameter, and its value range is within arrive Between, that is This weight Used to control the loss function of tuple relationships The relative importance of these factors in the joint loss function, and Then, the pairwise loss function is controlled accordingly. The percentage. The value needs to be pre-set by the experimenter based on task requirements or through optimization on a validation set before model training begins. For example, if capturing the overall combination of traits of a variety (higher-order tuple relationships) is considered more important than capturing pairwise associations between individual traits (pairwise relationships), then the value can be set to... .

[0062] 3) Obtain independent loss values , And determine the weights Then, construct the joint loss function. The operation is a direct linear weighted summation. The specific formula is as follows: In this formula, This represents the final value of the joint loss function, which is a scalar. These are preset weight parameters. It is the value of the tuple relation loss function. This is the value of the pairwise loss function. Multiplication operation. and The two original losses are scaled separately, and the addition operation combines the scaled terms into a single total loss target. This calculation process is technically very simple and efficient, requiring only a few lines of code to complete.

[0063] 4) The constructed joint loss function It is a function of the model parameters and the node representation vectors. It represents a unified optimization objective: minimizing... This means simultaneously minimizing the prediction error of tuple relations and the prediction error of pairwise relations, but both are subject to weights. Adjustment. During backpropagation, the gradient will... Starting from, they flow to calculate respectively. and The subgraph, thus according to and The proportion of these parameters is used to synchronously update the model parameters (such as the parameters of one-dimensional convolutional networks, enhanced feature extraction residual units, and fully connected layers) for tuple relation learning and the representation vectors for pairwise relation learning. This includes the parameters that these two parts may share. In this way, the optimization of the node representation vector no longer depends solely on a single type of relation signal, but rather receives collaborative supervision from both higher-order tuples and pairwise relations.

[0064] 5) In actual training, and The magnitude may vary depending on factors such as task definition, batch size, and negative sampling quantity. (Directly using a fixed...) Weighted summation can lead to one loss term dominating the optimization process. An enhanced implementation involves introducing dynamic adjustments or normalizing the loss values. For example, the moving averages of two loss functions over recent batches could be recorded and scaled before weighting to bring them to similar ranges. However, in the basic implementation, pre-set weights are used directly. Weighted summation is the most standard and direct method. Experimenters can manually adjust the loss terms by observing how they decrease during training. To achieve optimal performance. Ultimately, this joint loss function... The core optimization objective is to drive the entire HyReSim neural network model to learn and generate high-quality crop trait node representation vectors.

[0065] S43. Update the representation vector of each node by minimizing the joint loss function. The specific implementation process is as follows: 1) Before starting the optimization iteration, the joint loss function must be clearly defined. This depends on which trainable variables. These variables fall into two main categories: the first category is the set of parameters of the HyReSim neural network model itself. It includes all convolutional kernel weights and biases of the one-dimensional convolutional network, convolutional kernels and batch normalization layer parameters in the enhanced feature extraction residual units, and weight matrices of the fully connected layers. With bias The second category consists of the representation vectors of all nodes, which are essentially the parameters that the model needs to learn. Let's consider the set of nodes in a heterogeneous supernetwork for crops. Include Each node The representation vector is All these representation vectors can be viewed as a single trainable embedding matrix. Therefore, the complete set of parameters to be optimized is: . minimize The goal is to find an optimal set of... and .

[0066] 2) In each training iteration (usually for a batch of data), perform forward propagation and calculate the joint loss function value. Next, the backpropagation algorithm needs to be executed. The backpropagation algorithm uses a chain rule to deduce the results from the final output. Begin by calculating layer by layer backwards. The partial derivative for each trainable parameter, i.e., the gradient. Specifically, two types of gradients need to be calculated: ① Model parameter gradient: ,in .

[0067] ② Node representation of vector gradient: For each node Calculate the gradient of its representation vector. .

[0068] because This gradient also consists of two parts: one part comes from the influence of tuple relation loss on the representation vector, and the other part comes from the influence of pairwise relation loss on the representation vector. Modern deep learning frameworks such as PyTorch or TensorFlow can automatically and efficiently perform the calculation of all these gradients.

[0069] 3) After obtaining the gradient, use an optimization algorithm to update the parameters according to the gradient direction to reduce... The most basic algorithm is stochastic gradient descent. This algorithm requires a preset learning rate. (learning rate), which controls the step size for each update. The parameter update formula is as follows: For model parameters : For the node representation vector matrix (or each vector) ): here, It is the learning rate, which is a positive scalar hyperparameter. This indicates an assignment operation, where the current parameter value is subtracted from the product of the gradient and the learning rate to obtain the updated parameter value. It's important to note that in actual training, for computational efficiency, updates are typically only performed on the current training batch. The gradient of the representation vector of each node that has appeared in the algorithm is used, since the gradient of other nodes is zero. More advanced optimization algorithms, such as Adam or RMSprop, introduce mechanisms such as momentum and adaptive learning rate on top of SGD, but the core idea is still to utilize gradients. This guides the direction of parameter updates.

[0070] 4) The above two steps constitute a complete training iteration. This iteration will be repeated multiple times during model training. Each iteration uses a potentially different batch of data and calculates a new joint loss function. and its gradient, and update the model parameters. And node representation vector Through continuous iteration, the joint loss function... The values ​​generally show a downward trend, which means that the model's predictions of the similarity of the higher-order tuples to be modeled are becoming more accurate, and the pairwise relation information contained in the node representation vectors is also becoming richer. Iteration continues until a preset stopping condition is reached, such as: the total number of training rounds reaching a maximum value; the joint loss function... The model no longer shows a significant decrease on the validation set; or the early stopping strategy is triggered. When training stops, a set of optimized model parameters is obtained. and a final optimized and stable set of node representation vectors For all nodes .

[0071] 5) After updating through the process of minimizing the joint loss function described above, the representation vector of each node is... It is no longer a random initial value. It simultaneously integrates information from two supervisory signals: first, through tuple relation loss, the vector learns how to collaboratively combine with other vectors to form meaningful "variety" hyperedges; second, through pairwise relation loss, the vector learns to maintain appropriate similarity with other vectors in the local context. Therefore, the final node representation vector is a dense, low-dimensional mathematical representation that comprehensively reflects the high-order structure and pairwise associations of the corresponding phenotypic trait in the breeding network, laying a solid foundation for downstream crop variety prediction tasks.

[0072] S5. Using the optimized representation vectors of all nodes, predict crop varieties. The specific implementation process includes: S50. The goal of prediction is to recommend or evaluate the feasibility of a crop variety that does not yet exist or is being designed. This requires conceiving it as a specific combination of phenotypic traits. Suppose a target variety is conceived, consisting of a specific set of phenotypic trait nodes, denoted as the candidate node set. ,in It is the optimized set of nodes. A specific phenotypic trait node in the middle, This represents the number of traits involved in the candidate variety. According to the method definition, one variety corresponds to one higher-order tuple to be modeled. Therefore, this set of candidate nodes is arranged in a fixed order to construct a candidate higher-order tuple to be modeled and predicted. This tuple It refers to objects whose "ideality" or possibility of existence needs to be assessed.

[0073] S51. Load the final representation vectors of all nodes obtained after optimization by the joint loss function from the trained model or use them directly. For candidate higher-order tuples to be modeled Each node in Retrieve its optimized representation vector Then, strictly following the node order in the tuple, these optimized representation vectors are concatenated to form the node representation sequence of the candidate tuple. .

[0074] S52. Assemble the node representation sequence. The input is fed into a pre-trained target model with fixed parameters. The target model operates according to its inherent process: a one-dimensional convolutional network extracts basic features from the sequence; an augmented feature extraction residual unit processes the basic features into augmented features; and fully connected layers and activation functions ultimately calculate a similarity score based on the augmented features. This calculation process is a deterministic forward propagation, outputting a scalar value. : here, It is the similarity of the candidate higher-order tuples to be modeled, as predicted by the model. The parameter is The already trained target model, yes Flattened feature vectors after processing in the first few layers of the model and These are the weights and biases of the trained fully connected layer.

[0075] S53. Calculated similarity That is, the model corresponds to the envisioned crop variety. The prediction result. This value In the interval Internally, its numerical value can be directly interpreted as: under the pattern learned by the model based on historical data, the value derived from the set of phenotypic traits. The candidate variety constituted is scored as a reasonable, feasible, or "ideal" variety. The closer the score is to the correct value, the higher the probability score. This indicates that the more a trait combination conforms to the higher-order association patterns learned by the model from successful varieties, the higher the confidence level in predicting it as a superior variety. Breeding experts can set a threshold. ,For example .when If the candidate variety is deemed worthy of further field trials or focused attention, then adjustments to the trait combination can be considered; otherwise, adjustments can be made. Another application scenario is comparing multiple different candidate variety designs: constructing their respective higher-order tuples to be modeled and calculating their similarity. Then, they are sorted from high to low similarity to provide data-driven prioritization suggestions for breeding selection.

[0076] S54. Besides evaluating a given combination of traits, inverse or exploratory predictions can also be performed. For example, some key trait nodes of a variety can be fixed, and then the model can be used to search in the representation space for other trait nodes that maximize the similarity of the final tuples, thereby assisting in the design of an "ideal" variety that meets specific constraints. This requires searching in the node representation space or discrete candidate nodes using optimization algorithms. Regardless of the form, the technical basis is the use of optimized node representation vectors. The similarity between the trained model and the new high-order tuples to be modeled, which are composed of these vectors, is calculated, and the result of this calculation is used as the core basis for prediction.

[0077] Specifically, the prediction result refers to the similarity score output by the model after the optimized node representation vectors, which form the candidate high-order tuples to be modeled, are input into the trained HyReSim target model. This score is a quantified, continuous numerical value that integrates all the knowledge the model has learned from historical breeding data regarding higher-order tuple relationships and pairwise relationships. Prediction Results It is not a simple classification label, but a probabilistic estimate or confidence score that measures the degree of matching between a candidate variety and a known successful variety pattern. Breeders use the magnitude of this prediction result and its comparison with other candidate options to guide their breeding decisions, such as selecting high-scoring trait combinations as key experimental materials or rejecting combinations with too low scores.

[0078] This invention relates to the following technical content: 1) Hypernetworks: In crop breeding, complex relationships between traits and varieties can be mathematically modeled using hypernetworks. Hypernetworks are often abstracted as hypergraphs. .in, represent A collection of different types of nodes. Representing the A set of nodes of various types. In this invention, nodes correspond to phenotypic traits of different categories of crop varieties, such as morphological traits, yield traits, etc., therefore... This constitutes a heterogeneous supernetwork. It is a set of superedges, each superedge Include Each node represents a specific crop variety, and its nodes represent the characteristics of that variety. A phenotypic trait. If for any All Then the hypernet is - Uniform supernetwork. Figure 2 An example of a heterogeneous supernetwork is shown, in which, , , There are three different node types, which simulates a scenario in breeding where varieties consist of multiple types of traits.

[0079] 2) Residual Networks: Residual Networks (ResNet) are a deep neural network architecture that effectively alleviates the vanishing gradient problem during deep network training by introducing residual connections, making it possible to build and train very deep networks. The core idea of ​​residual networks is to learn the residual mapping between inputs and outputs, rather than directly learning the unreferenced mapping.

[0080] Residual connections are a fundamental component of residual networks, allowing direct shortcut connections to be established between network layers. Specifically, for a residual block, let its input be... The nonlinear transformation performed by the residual block is denoted as Then the final output of this block is This design allows the original input information to be... The ability to directly pass gradients to subsequent layers helps maintain efficient gradient flow as network depth increases and preserves crucial low-level feature information. Residual blocks are the basic building blocks of residual networks. A typical residual block contains a set of convolutional layers, activation functions, and a residual connection spanning these layers. Convolutional layers learn spatial mappings of features; activation functions (such as ReLU) introduce non-linearity; and the residual connection adds the block's input to the output of the convolutional transformation. Residual networks employ a modular design, building deep models by stacking multiple residual blocks; this structure is both clear and easily scalable.

[0081] 3) Based on the concept of residual networks, this invention designs a specialized enhanced feature extraction residual unit to process feature sequences extracted from hypernetwork tuples. For example... Figure 3 As shown. Figure 3The paper details the specific data processing flow and structure within the enhanced feature extraction residual unit. The input to this unit is the basic features extracted by the previous one-dimensional convolutional network. The data first undergoes a preliminary feature transformation through a one-dimensional convolution (Conv1D) operation, followed by the introduction of non-linearity through the ReLU activation function. Next, a max-pooling (MaxPooling1D) layer reduces the dimensionality of the feature sequence while preserving key information. Afterward, the data flows into the core of the unit, where it is further processed by a second one-dimensional convolution (Conv1D) layer, followed immediately by batch normalization and ReLU activation. Then… To extract fine-grained features with low computational cost, the process employs depthwise separable one-dimensional convolution (DV1D), followed by a pointwise one-dimensional convolution (Conv1D) to fuse the feature channels obtained from the previous operations. This result is then subjected to batch normalization. Finally, a crucial step involves an add operation, where the output of this batch normalization process is residually concatenated with the basic features initially input to this unit. The result of this addition is then input into a ReLU activation function to generate the final enhanced feature. This enhanced feature, as the output of this unit, is passed to subsequent fully connected layers of the target model to calculate the similarity of the higher-order tuples to be modeled. This residual unit uses a combination of depthwise separable convolution and pointwise convolution. Depthwise separable convolution first extracts fine-grained spatial features from the input with low computational cost; subsequently, pointwise convolution (i.e., 1x1 convolution) fuses and combines these features between channels. This design enables the enhanced feature extraction residual unit to efficiently learn the complex high-order tuple relationships contained in the node representation sequence, and enhances the model's ability to capture deep abstract features through multi-level feature fusion, while avoiding information degradation or noise accumulation problems caused by network deepening.

[0082] 4) The HyReSim neural network model is the core of this invention. Its goal is to fully encode the structural information in the hypernetwork into the low-dimensional representation vectors of the nodes by collaboratively optimizing pairwise and high-order tuple relationships. For example... Figure 4 As shown. Figure 4In the diagram, "1D convolution" corresponds to the 1D convolutional network in the final technical solution, responsible for extracting basic features from the node representation sequence; "residual unit" corresponds to the enhanced feature extraction residual unit, which processes the basic features into enhanced features through internal operations; "fully connected layer" corresponds to the fully connected layer and activation function, which calculates the similarity of the higher-order tuples to be modeled based on the enhanced features. Meanwhile, the "higher-order tuple filter" represents the preprocessing part of the input higher-order tuples to be modeled; "embedding layer (center word)" and "embedding layer (surrounding words)" and "dot product" together constitute the basic components of pairwise relation optimization, used to capture local similarity between nodes; and "tuple relation optimization" and "pairwise relation optimization" respectively refer to the two collaborative learning paths in the technical solution that construct tuple relation loss functions and pairwise relation loss functions based on similarity to optimize the node representation vectors. Figure 4 The complete architecture and data processing flow of the HyReSim neural network model were also demonstrated. The model input consists of two parts: first, the high-order tuples to be modeled generated by the high-order tuple filter, whose example nodes are a1, a2, b1, b2, c1, c2, etc. These nodes obtain the representation vectors of the center node and the context node through the embedding layer (center word) and the embedding layer (surrounding words), respectively, and calculate the pairwise similarity between nodes through the dot product operation, thereby serving the pairwise relation optimization path; second, the node representation sequence of the high-order tuples to be modeled composed of the same batch of nodes is input into the tuple relation optimization path. This path first extracts the basic features of the sequence through a one-dimensional convolutional network. These basic features then enter a residual unit (an enhanced feature extraction residual unit) for deep feature enhancement. The resulting enhanced features are finally input into a fully connected layer and combined with an activation function to calculate the similarity of the current tuple. This similarity is compared with the true label for tuple relationship optimization. The model's output layer explicitly distinguishes between non-tuple and tuple relationships as supervisory signals. The entire HyReSim neural network model collaboratively optimizes the loss functions from these two paths, jointly driving the update and learning of the representation vectors of all nodes, as detailed below: ① Tuple relation optimization aims to enable the model to learn and predict whether a set of nodes (i.e., a higher-order tuple to be modeled) can form a meaningful hyperedge (i.e., a reasonable crop variety). Input a higher-order tuple to be modeled into the model. The corresponding node represents a sequence. ,in This represents the node's representation vector. The sequence is first processed through a one-dimensional convolutional network to extract basic features. Basic features It is then fed into the enhanced feature extraction residual unit for in-depth feature enhancement. This process is described by the following formula (1): in, This represents the convolution operation. This indicates a batch normalization operation. This represents a nonlinear activation function (such as ReLU). Equation (1) reflects the internal calculation process of the enhanced feature extraction residual unit: for the input... Perform the first convolution Normalization Then, after activation function The result undergoes a second convolution, batch normalization, and activation function processing to obtain deep features; finally, these deep features are connected to the original input via residual connections. Adding them together yields the initial enhanced features. To further integrate the information, the following operations are also performed: Formula (2) will use the initial basic features A quick transformation path With initial enhancement features Add and activate to generate the final feature representation. This design enhances the integration of characteristics from different paths. Ultimately, it will... Input a fully connected layer and combine The activation function is used to calculate the predicted similarity of the tuple, and the loss is constructed accordingly. in, and These are the weight matrix and bias vector of the fully connected layer, respectively. This specifically refers to Activation function. This is the actual label of the tuple, defined as follows: During training, the model is driven to accurately determine the authenticity of tuple relationships by minimizing the loss constituted by formula (3).

[0083] ② Besides higher-order tuple relations, pairwise similarity between nodes is also crucial for constructing accurate representation vectors. Pairwise relation optimization is based on node representation vectors. It learns by measuring the co-occurrence probability of nodes in a local context. Given a pair of nodes... and Its pairwise loss function is defined as: Among them, conditional probability Defined as: here, Represents the target node context node, It is a normalized partition function. The essence of formula (5) is to maximize the log-likelihood of a node co-occurring with its true context nodes. In actual training, to avoid calculating the entire node set... The high cost of normalization terms leads to the widespread use of negative sampling techniques for efficient approximate optimization.

[0084] In summary, the HyReSim neural network model collaboratively optimizes the tuple relation loss (Equation (3)) and pairwise relation loss (a practical variant of Equation (5)) using a joint loss function. This joint loss function takes the form of: ,in It is a preset hyperparameter used to flexibly control the relative weights of the two losses in the final optimization objective, thereby fully capturing information in the hypernetwork for downstream link prediction and variety prediction tasks.

[0085] To comprehensively evaluate the performance of the proposed HyReSim neural network model in capturing complex relationships, this invention was experimentally validated on several publicly available hypernetwork datasets. Four representative hypernetwork datasets—GPS Network, Drug Network, MovieLens Social Network, and WordNet Semantic Network—were used for comprehensive evaluation. These datasets can all be naturally modeled as heterogeneous hypernetworks, where each higher-order tuple constitutes a hyperedge. Detailed statistical information for the datasets is shown in Table 1.

[0086] Table 1: The GPS dataset records user trajectory information. A high-order tuple (user, location, activity) is used as a hyperedge to construct a hypernetwork, simulating the overall event of "a user performing an activity at a certain location." The drug dataset records adverse events and drug error reports. A high-order tuple (user, drug, reaction) is used as a hyperedge to construct a hypernetwork, describing the complete record of "a user experiencing a specific reaction after taking a certain drug." The MovieLens dataset records user movie tagging activities. A high-order tuple (user, movie, tag) is used as a hyperedge to construct a hypernetwork, expressing the comprehensive behavior of "a user tagging a movie." The wordnet dataset records WordNet 3.0 semantic network information. A high-order tuple (head entity, relation, tail entity) is used as a hyperedge to construct a hypernetwork, encoding a complete triple fact.

[0087] In the analogy of crop breeding, the structure of these datasets is consistent with the model where "crop varieties (hyperedges) consist of multiple phenotypic traits (nodes)". Each dataset provides a clear set of node types and hyperedges for training and evaluating the model's ability to learn node representation vectors.

[0088] To ensure fair comparison, this invention selects several advanced network and hypernetwork representation learning methods as baselines, including random walk-based and deep learning-based methods, specifically: ① Deepwalk: This method applies a Skip-gram model to node sequences generated by random walks on a regular graph to learn low-dimensional embedding vectors of nodes in the network. ② Node2Vec: This method obtains more expressive node representation vectors by designing a biased random walk strategy on a regular graph that combines depth-first and breadth-first search. ③ HPSG: This method is a hypernetwork representation learning method that uses node sequences generated by hyperpath-based random walks as input to a Skip-gram model to learn node representation vectors. ④ Hyper2Vec: This method is a hypernetwork representation learning method that captures higher-order relationships between nodes by performing biased second-order random walks on a hypernetwork, thereby learning node representation vectors. ⑤ HyperS2V: This method is a hypernetwork representation learning method that quantifies structural similarity between nodes based on node hyperdegree information, thereby learning node representation vectors. ⑥DHNE: This method is a hypernetwork representation learning approach that directly processes hyperedge information through a neural network model incorporating a multilayer perceptron, thereby learning node representation vectors. These baseline methods cover different design approaches from ordinary graphs to hypernetworks, and from shallow to deep models, providing a comprehensive reference for evaluating the overall performance of the HyReSim neural network model.

[0089] Link prediction is a key task in evaluating the performance of representation learning models, aiming to predict missing or potentially future hyperedges in a hypernetwork. In this experiment, four hypernetwork datasets—GPS, drug, MovieLens, and WordNet—were selected as the research objects. For each dataset, 80% of the hyperedges were randomly selected as the training set for learning node representation vectors, and the remaining 20% ​​were used for testing. Ten independent random training / test split experiments were conducted on the HyReSim neural network model and all baseline methods. The average AUC (Area Under Curve) of the prediction results was then calculated as the performance metric. Detailed experimental comparison data are shown in Table 2.

[0090] Table 2: As shown in Table 2, the HyReSim neural network model outperforms the network representation learning methods deepwalk and node2vec, and the hypernetwork representation learning methods HPSG and Hyper2vec in link prediction on the GPS and WordNet datasets. Furthermore, on the drug and MovieLens datasets, the HyReSim neural network model demonstrates experimental performance advantages that rival the state-of-the-art hypernetwork representation learning method DHNE. These results are primarily attributed to the fact that the HyReSim neural network model not only captures pairwise relationships between nodes through the pairwise relationship loss function derived from Equation (5), but also strengthens the capture of higher-order tuple relationships through the tuple relationship optimization path constructed by Equations (1), (2), and (3). Therefore, the HyReSim neural network model not only exhibits strong performance in capturing pairwise relationships between hypernetwork nodes, but also effectively applies to and improves the link prediction accuracy in hypernetworks with complex higher-order tuple relationships.

[0091] Hypernetwork reconstruction is another important evaluation metric, used to examine the extent to which the node representation vectors learned by the model can reconstruct the overall structure of the original network. Hypernetwork reconstruction experiments were conducted on the GPS and drug datasets. The evaluation metric for hypernetwork reconstruction is defined as: in, It is an indicator variable: This indicates the first [unit / item] reconstructed based on the model score. The hyperedges must actually exist in the original hypernetwork; otherwise... . The size of the original hyperedge set. The super-network reconstruction ratio is the proportion of reconstructed superedges to the total number of original superedges. Measured before reconstruction The accuracy value when the scale exceeds the margin.

[0092] The results of the hypernetic network reconstruction are as follows Figure 5 and Figure 6 As shown in the figure, experimental results demonstrate that the HyReSim neural network model outperforms the network representation learning methods deepwalk, node2vec, and the hypernetwork representation learning method HyperS2V on both datasets in terms of reconstruction accuracy. This result directly indicates that the proposed HyReSim neural network model, by incorporating hyperedge information into hypernetwork representation learning and co-optimizing pairwise and tuple relationships, obtains higher quality and more structurally faithful node representation vectors. These vectors can better support the task of reconstructing the global network structure from local representations.

[0093] This invention addresses the problem of how to more efficiently characterize both pairwise and higher-order tuple relationships between nodes in hypernetworks, proposing the HyReSim neural network model. The model defines the tuple relationship loss function and the pairwise relationship loss function using formulas (3) and (5), respectively, and performs collaborative optimization through a parameter-weighted joint loss function. The model uses an enhanced feature extraction residual unit as a key technical component, achieving a balance between model depth, feature purity, and computational efficiency. Experiments show that on four heterogeneous hypernetworks—GPS, drug, MovieLens, and WordNet—the HyReSim neural network model outperforms most baseline methods in link prediction and hypernetwork reconstruction tasks, validating the effectiveness of its design. Specifically, this invention enhances the ability to capture higher-order tuple relationships by constructing a deep neural network model containing a one-dimensional convolutional network, an enhanced feature extraction residual unit, and fully connected layers. This model inputs the node representation sequence corresponding to the high-order tuple to be modeled into the network, performs basic feature extraction, deep feature enhancement and fusion in sequence, and finally calculates the similarity representing the possibility of the existence of the tuple, thereby achieving refined modeling of the overall semantics of the hyperedge. Moreover, this invention adopts a structurally clear and functionally independent enhanced feature extraction residual unit as the core module. This unit is designed as a separate unit, and its internal calculation process is shown in formulas (1) and (2). Enhanced features are generated by two convolutions, batch normalization, activation function processing and residual connection with the input features. This modular design allows for adjustment and customization of the type, number or connection method of convolution kernels according to different hypernetwork characteristics or computational resource constraints, improving the adaptability and scalability of the model. Moreover, this invention retains the original feature information in the deep network of the model through the residual connection mechanism. In the design of the enhanced feature extraction residual unit and the model broader, the residual connection allows the input features to be directly passed to the subsequent layers and added to the features after complex nonlinear transformation. This approach retains the rich information obtained from the powerful multi-level feature extraction capabilities of deep neural networks while mitigating the gradient degradation or irrelevant noise interference that may be caused by excessively deep networks, ensuring that the learned node representation vectors are both deeply abstract and stable and reliable.

[0094] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0095] like Figure 7As shown, an embodiment of the present invention provides a supernetwork representation learning system 200 based on residual unit optimization for use in breeding, comprising a construction module 201, an acquisition module 202, a similarity acquisition module 203, an optimization module 204, and a prediction module 205. Module 201 is used to: construct a heterogeneous supernetwork of crops, where nodes represent phenotypic traits of crop varieties and superedges represent crop varieties; The acquisition module 202 is used to: acquire the representation vector of each node in the crop heterogeneous supernetwork, and acquire at least one high-order tuple to be modeled, wherein each hyperedge corresponds to a high-order tuple to be modeled; The similarity acquisition module 203 is used to: input the node representation sequence corresponding to each higher-order tuple to be modeled into the target model, and obtain the similarity of each higher-order tuple to be modeled; The optimization module 204 is used to optimize the representation vector of each node in the crop heterogeneous supernetwork based on the similarity of each high-order tuple to be modeled. The prediction module 205 is used to predict crop varieties using the optimized representation vectors of all nodes.

[0096] Optionally, in the above technical solution, the target model sequentially includes a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function. The one-dimensional convolutional network is used to extract basic features from the node representation sequence, the enhanced feature extraction residual unit is used to enhance the basic features into enhanced features, and the fully connected layer and activation function are used to calculate the similarity of the higher-order tuples to be modeled based on the enhanced features.

[0097] Optionally, in the above technical solution, the process by which the enhanced feature extraction residual unit processes the basic features into enhanced features includes: performing a first convolution operation on the input basic features to obtain a first intermediate feature; performing a first batch normalization and a first activation function on the first intermediate feature to obtain a second intermediate feature; performing a second convolution operation on the second intermediate feature to obtain a third intermediate feature; performing a second batch normalization and a second activation function on the third intermediate feature to obtain a fourth intermediate feature; and performing a residual connection between the fourth intermediate feature and the basic features to obtain the enhanced feature.

[0098] Optionally, in the above technical solution, the optimization module 204 is specifically used to: construct a tuple relation loss function based on the similarity of all the higher-order tuples to be modeled and their corresponding true labels; construct a pairwise relation loss function based on the representation vectors of nodes in the crop heterogeneous supernetwork; combine the tuple relation loss function and the pairwise relation loss function according to preset weights to obtain a joint loss function; and update the representation vector of each node by minimizing the joint loss function.

[0099] It should be noted that the beneficial effects of the supernetwork representation learning system 200 based on residual unit optimization for breeding provided in the above embodiments are the same as the beneficial effects of the supernetwork representation learning method based on residual unit optimization for breeding described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0100] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned supernetwork representation learning methods based on residual unit optimization applied to breeding.

[0101] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned supernetwork representation learning methods based on residual unit optimization applied to breeding.

[0102] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0103] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A hypernetwork representation learning method based on residual unit optimization applied to breeding, characterized in that, include: Construct a heterogeneous supernetwork for crops, where nodes represent phenotypic traits of crop varieties and superedges represent crop varieties; Obtain the representation vector of each node in the crop heterogeneous supernetwork, and obtain at least one high-order tuple to be modeled, wherein each hyperedge corresponds to a high-order tuple to be modeled. Input the node representation sequence corresponding to each of the higher-order tuples to be modeled into the target model to obtain the similarity of each higher-order tuple to be modeled; Based on the similarity of each higher-order tuple to be modeled, the representation vector of each node in the crop heterogeneous supernetwork is optimized; Crop varieties are predicted using the optimized representation vectors of all nodes.

2. The hypernetwork representation learning method based on residual unit optimization applied to breeding, as described in claim 1, is characterized in that, The target model sequentially includes a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function. The one-dimensional convolutional network is used to extract basic features from the node representation sequence, and the enhanced feature extraction residual unit is used to enhance the basic features into enhanced features. Fully connected layers and activation functions are used to calculate the similarity of the higher-order tuples to be modeled based on the enhanced features.

3. The hypernetwork representation learning method based on residual unit optimization applied to breeding, as described in claim 2, is characterized in that... The process by which the enhanced feature extraction residual unit processes the basic features into the enhanced features includes: Perform a first convolution operation on the input basic features to obtain the first intermediate features; The first intermediate feature is subjected to a first batch normalization and a first activation function process to obtain the second intermediate feature; Perform a second convolution operation on the second intermediate feature to obtain the third intermediate feature; The third intermediate feature is subjected to a second batch normalization and a second activation function to obtain the fourth intermediate feature; The enhanced feature is obtained by performing a residual connection between the fourth intermediate feature and the basic feature.

4. A hypernetwork representation learning method based on residual unit optimization applied to breeding, as described in any one of claims 1 to 3, characterized in that, Based on the similarity of each higher-order tuple to be modeled, the representation vector of each node in the crop heterogeneous supernetwork is optimized, including: Based on the similarity of all higher-order tuples to be modeled and their corresponding true labels, a tuple relation loss function is constructed. Based on the representation vectors of nodes in the heterogeneous supernetwork of crops, a pairwise relation loss function is constructed; The tuple relation loss function and the pairwise relation loss function are combined according to preset weights to obtain a joint loss function; The representation vector of each node is updated by minimizing the joint loss function.

5. A hypernetwork representation learning system based on residual unit optimization for use in breeding, characterized in that, It includes a construction module, an acquisition module, a similarity acquisition module, an optimization module, and a prediction module; The construction module is used to: construct a heterogeneous supernetwork of crops, wherein nodes represent phenotypic traits of crop varieties and superedges represent crop varieties; The acquisition module is used to: acquire the representation vector of each node in the crop heterogeneous supernetwork, and acquire at least one high-order tuple to be modeled, wherein each hyperedge corresponds to one high-order tuple to be modeled. The similarity acquisition module is used to: input the node representation sequence corresponding to each of the higher-order tuples to be modeled into the target model to obtain the similarity of each higher-order tuple to be modeled; The similarity acquisition module is used to optimize the representation vector of each node in the crop heterogeneous supernetwork based on the similarity of each high-order tuple to be modeled. The prediction module is used to predict crop varieties using the optimized representation vectors of all nodes.

6. A hypernetwork representation learning system based on residual unit optimization for breeding, as described in claim 5, is characterized in that... The target model sequentially includes a one-dimensional convolutional network, an enhanced feature extraction residual unit, a fully connected layer, and an activation function. The one-dimensional convolutional network is used to extract basic features from the node representation sequence, and the enhanced feature extraction residual unit is used to enhance the basic features into enhanced features. Fully connected layers and activation functions are used to calculate the similarity of the higher-order tuples to be modeled based on the enhanced features.

7. A hypernetwork representation learning system based on residual unit optimization for breeding, as described in claim 6, is characterized in that, The process by which the enhanced feature extraction residual unit processes the basic features into the enhanced features includes: Perform a first convolution operation on the input basic features to obtain the first intermediate features; The first intermediate feature is subjected to a first batch normalization and a first activation function process to obtain the second intermediate feature; Perform a second convolution operation on the second intermediate feature to obtain the third intermediate feature; The third intermediate feature is subjected to a second batch normalization and a second activation function to obtain the fourth intermediate feature; The enhanced feature is obtained by performing a residual connection between the fourth intermediate feature and the basic feature.

8. A hypernetwork representation learning system based on residual unit optimization for breeding, as described in any one of claims 5 to 7, characterized in that, The optimization module is specifically used for: Based on the similarity of all higher-order tuples to be modeled and their corresponding true labels, a tuple relation loss function is constructed. Based on the representation vectors of nodes in the heterogeneous supernetwork of crops, a pairwise relation loss function is constructed; The tuple relation loss function and the pairwise relation loss function are combined according to preset weights to obtain a joint loss function; The representation vector of each node is updated by minimizing the joint loss function.

9. An electronic device, characterized in that, The invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the supernetwork representation learning method based on residual unit optimization for breeding as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the supernetwork representation learning method based on residual unit optimization for breeding, as described in any one of claims 1 to 4.

Images