Agricultural multi-source knowledge migration method based on graph model

Abstract

The invention discloses an agricultural multi-source knowledge migration method based on a graph model, and belongs to the technical field of agricultural information, and the method comprises the steps: carrying out the reduction, feature selection and feature fusion of agricultural literature data; the data processing efficiency and accuracy are improved by utilizing agricultural data reduction based on self-organizing mapping and feature selection and fusion based on spatial mapping; constructing a knowledge migration network based on a graph model, and performing knowledge semantic extraction and knowledge migration relationship mining; agricultural multi-carrier meta-knowledge acquisition and information feedback are adjusted, and meta-knowledge acquisition and network weight adjustment are optimized. According to the method, large-scale and heterogeneous agricultural data can be effectively processed, effective migration of knowledge is realized, the performance of agricultural learning tasks is greatly improved, and the method is of great significance to improvement of agricultural production efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural information technology, and more specifically relates to an agricultural multi-source knowledge migration method based on a graph model. Background Art

[0002] In modern society, the acquisition and dissemination of agricultural knowledge has become an important part of social development, especially in terms of its impact on efficient agricultural production. The acquisition of agricultural knowledge mainly relies on the analysis and processing of agricultural-related data to accumulate experience. However, agricultural-related data usually come from a wide range of sources, have complex structures, and have poor homogeneity, which brings certain difficulties to data analysis. In addition, there is professional and regional knowledge in different agricultural fields, and the transfer and migration of this knowledge is not easy.

[0003] The agricultural multi-source knowledge transfer method based on the graph model can realize the migration between different sources, heterogeneous and large-scale agricultural knowledge, breaking through the boundaries of agricultural knowledge positions and professions. However, the existing knowledge transfer method based on the graph model still has some problems when dealing with agricultural multi-source knowledge migration, such as data redundancy, feature selection and fusion cannot meet the needs of multi-source heterogeneous data, and the graph model is difficult to determine the appropriate knowledge migration path in massive unlabeled agricultural data. To this end, the present invention proposes a more efficient and optimized method.

[0004] In addition, existing knowledge transfer methods also have problems in acquiring meta-knowledge and adjusting information feedback. The purpose of meta-knowledge acquisition is the principle of structural similarity, but existing methods still have problems in how to achieve high accuracy and strong extensibility in meta-knowledge acquisition. At the same time, information feedback adjustment is an important process, but existing methods are not ideal in adjusting network weights and improving the accuracy of learning tasks. Taking these problems into consideration, the present invention proposes a new method for meta-knowledge acquisition and information feedback adjustment. Summary of the invention

[0005] The present invention mainly solves key technical problems in agricultural data processing, such as efficient simplification of agricultural multi-source data to reduce redundant information, effective feature selection and fusion processing of heterogeneous data, construction of a graph model with knowledge transfer capabilities to express the relationship between agricultural data, and acquisition of meta-knowledge and efficient information feedback adjustment. Through these methods, the present invention provides an agricultural multi-source knowledge transfer method based on a graph model to improve data processing efficiency and accuracy and enhance the performance of agricultural learning tasks.

[0006] In order to solve the above problems, the present invention is implemented by adopting the following technical solutions: the method comprises:

[0007] S1. Data simplification, feature selection and feature fusion of agricultural literature, agricultural data simplification based on self-organizing maps; feature selection and fusion based on spatial mapping;

[0008] S2. Build a knowledge transfer network based on graph model, extract knowledge semantics, and mine knowledge transfer relationships;

[0009] S3. Adjust the agricultural multi-carrier meta-knowledge acquisition and information feedback, meta-knowledge acquisition, and information feedback adjustment.

[0010] In one embodiment, the agricultural data reduction based on self-organizing maps includes:

[0011] (1) First, collect and organize agricultural multi-source data. Due to different agricultural characteristics or sources, there is a lot of duplicate or incomplete information;

[0012] (2) Use the self-organizing map model SOM to map the input high-dimensional data vector into a grid structure in a low-dimensional space;

[0013] (3) The “rule + exception” concept of rough sets is introduced into the SOM model. For rule theory, a set of objects that are consistent in all attributes is regarded as a concept or a rule, and other inconsistent objects are regarded as exceptions. By eliminating exception data and preserving rule data, the amount of data is reduced while retaining the main characteristics of the data.

[0014] In one embodiment, the feature selection and fusion based on spatial mapping includes:

[0015] Feature selection is the process of selecting a subset of features from the original data set; this is done by calculating the correlation between different features; for each agricultural multi-carrier data, these features are mapped to a unified feature space for comparison and selection;

[0016] Feature fusion maps all carrier data to the feature fusion space by constructing a mapping function; in the mapping space, redundant information and noise between features are eliminated, revealing the common features hidden in different data carriers.

[0017] In one embodiment, the knowledge semantic extraction includes:

[0018] Preprocessing: First, preprocess the agricultural multi-source feature data, including data cleaning and format conversion, so that the data can meet the subsequent processing requirements.

[0019] Feature extraction: Use the TF-IDF algorithm to extract useful features from the preprocessed data.

[0020] In one embodiment, the knowledge transfer relationship mining includes:

[0021] Initialize the graph model: Use features as nodes to initially build a graph model; each node represents a feature, and the edge represents the relationship between features;

[0022] Label propagation: Use semi-supervised learning to train the model using a small amount of labeled data, and build the model together with unlabeled data; if there is similarity between labeled samples and unlabeled samples, propagate the label of the labeled sample to similar unlabeled samples;

[0023] Optimize migration relationships: Based on label propagation, use the annotation information of the auxiliary carrier to optimize the migration relationship of the target unlabeled samples; by adjusting the edge weights, nodes with high similarity have the same or similar labels.

[0024] In one embodiment, the meta-knowledge acquisition includes:

[0025] Firstly, by following the principle of structural similarity, universal meta-knowledge is extracted from agricultural multi-carrier data, and a one-to-one mapping relationship is established between them in the knowledge transfer network and the meta-knowledge network; by applying the rough set attribute reduction theory, extension knowledge is obtained from the semantic network, and the meta-knowledge represented by extension transformation can be used to establish the corresponding meta-knowledge base;

[0026] Then, the weight relationship between different knowledge clusters is determined according to the Keeney-Raiffa algorithm; under the control of the knowledge cluster weights, the knowledge of different carriers will provide information support for the needs of agricultural knowledge transfer tasks.

[0027] In one embodiment, the information feedback adjustment includes:

[0028] Based on the characteristics of multi-source heterogeneity in agriculture, the information feedback adjustment process adopts a feedback algorithm of network weight adjustment;

[0029] Assuming that ωij(t) is the weight corresponding to the jth knowledge node when the tth round network learns the i-th carrier sample data, the learning process is expressed as ωij(t+1)=ωij(t)+α(t)dij,

[0030] Where α(t) is the decay learning rate, dij is the common metric between the jth knowledge and the ith carrier data, and the Euclidean distance is used here;

[0031] In this way, data resources irrelevant to public knowledge are gradually eliminated during the learning process, thus improving the accuracy of the learning task.

[0032] In one scheme, to avoid overfitting, excessive weights are pruned and ωij is between 0 and 1.

[0033] Beneficial effects of the present invention:

[0034] 1. By adopting the agricultural data reduction method based on self-organizing mapping, a large amount of agricultural data can be effectively processed, data redundancy and incompleteness can be effectively reduced, and the efficiency and accuracy of data processing can be improved.

[0035] 2. Feature selection and fusion methods based on spatial mapping can effectively process heterogeneous agricultural data, extract meaningful features, improve the comparability of features, and make feature analysis more accurate and valuable.

[0036] 3. The knowledge transfer network based on the graph model constructed by the present invention can effectively simulate the transfer relationship of agricultural knowledge, realize the effective transfer of knowledge through label propagation and optimization of transfer relationship, and greatly improve the efficiency and accuracy of agricultural knowledge transfer.

[0037] 4. Through meta-knowledge acquisition and information feedback adjustment, agricultural meta-knowledge can be captured and applied more accurately. By adjusting the network weights, the accuracy of learning tasks is improved, making agricultural knowledge transfer more adaptable and robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the method of the present invention;

[0039] Figure 2 This is a flow chart of knowledge semantic extraction of the present invention;

[0040] Figure 3 This is a flowchart for realizing the knowledge transfer relationship mining of the present invention;

[0041] Figure 4 This is a flow chart for extracting meta-knowledge based on the rough set attribute reduction theory of the present invention. DETAILED DESCRIPTION

[0042] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Typical embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0043] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0044] like Figure 1 As shown, a method for agricultural multi-source knowledge transfer based on a graph model includes:

[0045] S1. Perform data simplification, feature selection and feature fusion of agricultural literature;

[0046] 1) Agricultural data reduction based on self-organizing maps

[0047] In view of the redundant and incomplete characteristics of agricultural multi-source data, the data reduction method using self-organizing map model network and rough set is studied. After data crawling, parsing, analysis, indexing and searching, vector quantization is used to merge data resource objects into a weighted set of two-dimensional array nodes. This set is used to represent each object, eliminating redundant information and integrating incomplete information. The "rule + exception" idea of ​​rough set is introduced into the self-organizing map model, which retains the "rule" data to represent the data that will be encountered in the future in this field, and eliminates the "exceptions".

[0048] (1) First, collect and organize agricultural multi-source data. Due to different agricultural characteristics or sources, these data may contain a large amount of duplicate or incomplete information.

[0049] (2) Use the self-organizing map model (SOM). SOM is an unsupervised learning method that does not require predetermined categories. It can map the input high-dimensional data vector to a grid structure in a low-dimensional space. The specific process is to use competitive learning to input a sample set, and the network iteratively adjusts the node weights. Finally, the weights of all nodes converge, generating an ordered feature map. Suppose there are M data points x1, x2, …, xM. These data are mapped to N nodes through the self-organizing map model, where N < < M, and each node corresponds to a weight vector w1, …, wN. After iterative calculation, the node closest to each data point will be found, and the weights of the node and its surrounding nodes will be updated.

[0050] In the self-organizing map, the update rule of node weights is expressed as:

[0051] w j (t+1)=w j (t)+η(t)h j (x(t)-w j (t))

[0052] Among them, w j is the weight of node j, η(t) is the learning rate, h j is the neighborhood function, h j is the input vector.

[0053] (3) The "rule + exception" concept of rough sets is introduced into the SOM model. Rough sets are used to process data containing uncertainty and contradictions. For rule theory, a set of objects that are consistent in all attributes is regarded as a concept or a rule, and other inconsistent objects are regarded as exceptions. By eliminating exception data and preserving rule data, the amount of data can be effectively reduced while retaining the main characteristics of the data.

[0054] 2) Feature selection and fusion based on spatial mapping

[0055] In view of the multimodal characteristics of agricultural multi-carrier data, this paper studies the method of mapping various carrier data hierarchically to the feature fusion space, and gives a reasonable explanation of this method from the geometric aspect. It eliminates the redundant information in the features and realizes the organic fusion of multimodal data features.

[0056] (1) Feature selection, also known as feature subset selection or attribute selection, is the process of selecting a subset of features from the original dataset. This is usually done by calculating the correlation between different features or the correlation between features and the target variable. For each type of agricultural multi-carrier data, these features are mapped to a unified feature space for comparison and selection.

[0057] In feature selection, the calculation formula is:

[0058]

[0059] Mutual information is used to measure the mutual dependence between X and Y;

[0060] I(X;Y) is the mutual information between random variables X and Y, which is used to measure the mutual dependence between X and Y;

[0061] X is a feature, Y is a fusion feature;

[0062] p(x,y) represents the joint probability distribution function of X and Y. p(x,y) gives the probability of the joint event that X takes the value x and Y takes the value y.

[0063] p(x) and p(y) represent the marginal probability distribution functions of X and Y respectively.

[0064] and It means summing up all possible values ​​of x and y.

[0065] In the calculation formula of mutual information, if X and Y are independent, then p(x,y) will be equal to p(x)p(y), and the sum of the whole formula will be 0. If X and Y are completely dependent, that is, knowing the value of X can determine the value of Y, then the sum will be 1. Therefore, mutual information can well quantify the correlation between two random variables.

[0066] (2) Feature fusion is to fuse the dynamically selected feature sets together to form a new feature expression. The specific process is to construct a mapping function to map all carrier data to the feature fusion space. This mapping function is linear. In the mapping space, redundant information and noise between features will be eliminated, thereby revealing the common features hidden in different data carriers.

[0067] In the feature fusion process, the linear mapping function is in the form of:

[0068] Y=AX+B

[0069] Among them, Y is the fusion feature, X is the input feature, and A and B are the linear mapping parameters.

[0070] S2. Build a knowledge transfer network based on graph model;

[0071] like Figure 2 As shown, 1) Knowledge semantic extraction

[0072] Agricultural knowledge semantics is regarded as nodes in the graph model. In order to determine the nodes in the network, it is necessary to study the method of extracting common features from multi-source features, especially to find data features in the auxiliary carrier that contribute to the learning task of the target carrier. The complementarity and ambiguity processing mechanism of the knowledge semantic concepts of the auxiliary carrier and the target carrier are studied.

[0073] (1) Preprocessing: First, the agricultural multi-source feature data is preprocessed, including data cleaning and format conversion, so that the data can meet the needs of subsequent processing.

[0074] (2) Feature extraction: The TF-IDF algorithm is used to extract useful features from the preprocessed data. In this process, attention is paid to those data features that contribute to the target carrier learning task.

[0075] Using the TF-IDF algorithm, the calculation formula for TF-IDF is:

[0076]

[0077] Among them, n i,j is the number of occurrences of word i in document j, is the sum of the number of occurrences of all words in document j, that is, the total number of words in document j, |D| is the total number of documents in the corpus, |j:t i ∈d j | is the total number of documents containing word i.

[0078] (3) Semantic representation: The word2vec algorithm is used to convert the extracted features into semantic information, which can simplify subsequent semantic calculations.

[0079] 2) Knowledge transfer relationship mining

[0080] Based on semi-supervised learning, we mine knowledge transfer relationships and analyze the migration algorithm of a large number of target unlabeled samples by making full use of a large amount of known auxiliary carrier annotation information on the basis of a small number of labeled target carrier samples. At the same time, we use a hierarchical graph model to analyze the relationship between knowledge at different granularity levels, which is more persuasive.

[0081] like Figure 3 The detailed implementation steps of knowledge transfer relationship mining are as follows:

[0082] 1) Initialize the graph model: Use the extracted features as nodes to initially build a graph model. In this model, each node represents a feature and the edge represents the relationship between features.

[0083] 2) Label propagation: In addition to using a small amount of labeled data to train the model, semi-supervised learning can also be used to build the model together with unlabeled data. If there is similarity between labeled samples and unlabeled samples, the label of the labeled sample can be propagated to similar unlabeled samples.

[0084] The calculation formula of the label propagation algorithm (LPA) is:

[0085] Y (t) =αY (0) +(1-α)SY (t-1)

[0086] Among them, Y (t) is the label prediction value of the t-th iteration, αY (0) is the initial label value (the label of the labeled sample is known, and the label of the unlabeled sample is generally initialized to 0), S is the similarity matrix between samples, and α is the balance coefficient, which determines the speed of label propagation.

[0087] 3) Optimize the migration relationship: Based on label propagation, the label information of the auxiliary carrier is used to optimize the migration relationship of the target unlabeled samples. The goal of the optimization is to adjust the edge weights so that nodes with high similarity have the same or similar labels.

[0088] Using support vector machine (SVM) as the classifier, the optimization objective function of SVM is:

[0089]

[0090] The constraints are

[0091] y i (w T φ(x i )+b)≥1-ξ i ,1≤i≤n

[0092] Among them, w is the normal vector of the decision boundary, b is the displacement term, C is the penalty coefficient, and ξ is the slack variable used to control the degree of penalty for misclassification.

[0093] S3. Adjust the acquisition of agricultural multi-carrier meta-knowledge and information feedback.

[0094] 1) Meta-knowledge acquisition

[0095] In the process of meta-knowledge acquisition, first of all, the acquisition method of agricultural multi-carrier meta-knowledge and its structure is studied to form a learning mechanism of meta-knowledge. By following the principle of structural similarity, universal meta-knowledge can be extracted from agricultural multi-carrier data, and a one-to-one mapping relationship can be established between them in the knowledge transfer network and the meta-knowledge network. For example, by applying the rough set attribute reduction theory, extension knowledge can be obtained from the semantic web, and the meta-knowledge represented by extension transformation can establish a corresponding meta-knowledge base. Then, the weight relationship between different knowledge clusters can be determined according to the Keeney-Raiffa algorithm. Under the control of the knowledge cluster weight, the knowledge of different carriers will provide better information support for the needs of agricultural knowledge transfer tasks.

[0096] like Figure 4 As shown in the figure, when using rough set attribute reduction theory to extract meta-knowledge, it is achieved through the following steps:

[0097] S301, construct a decision table: Assume that there is a set of agricultural carrier data X = {x1, x2, ..., xn}, where each element xi has some attributes A = {a1, a2, ..., am}. Then a decision table D = {X, A, V, f} can be constructed, where V = {V(a1), V(a2), ..., V(am)} represents the possible value set of each attribute, and f: X×A→V is an information function.

[0098] S302, calculate the importance of attributes: for each attribute ai in the decision table D, calculate its importance SIG(ai). This can be done by the following steps:

[0099] S3021, calculating the roughness R(POS(X)) of the entire set POS(X);

[0100] S3022, calculate the roughness R(POS(X)-{ai}) of the whole set POS(X) after eliminating the attribute ai;

[0101] S3023. Compare the two roughnesses to obtain the importance of attribute ai: SIG(ai)=R(POS(X))-R(POS(X)-{ai}).

[0102] S303, attribute simplification: Sort the attributes by importance, and the attributes at the top are considered to be more important. Then, a threshold is selected, and the attributes whose importance is less than the threshold are discarded. The remaining attributes are the "meta-knowledge" obtained after simplification.

[0103] The calculation formula for the roughness involved in the above process is: Where BND(X) is the boundary area, which represents the sample set that cannot be accurately divided under the division of this information system.

[0104] The Keeney-Raiffa algorithm is a weight-solving method, which is mainly used to solve the weights between different targets in a decision tree. In the process of meta-knowledge acquisition, each knowledge cluster can be regarded as a decision target, and the knowledge of each carrier can be regarded as the benefit of the corresponding target. Using the Keeney-Raiffa algorithm, the weight of each knowledge cluster can be determined to provide better information support in the agricultural knowledge transfer task.

[0105] Assuming there are n knowledge clusters, i.e., decision goals, the benefit of the decision on each goal can be set as R1, R2, ... Rn, and the weight of each goal is w1, w2, ... wn. According to the Keeney-Raiffa algorithm, the sum of the product of the weight and the benefit should be equal to the total decision utility U, i.e.: U = w1R1 + w2R2 + .... + wn*Rn.

[0106] On this basis, the Keeney-Raiffa algorithm mainly determines the weight of each goal by comparing the changes in the decision utility (i.e. marginal utility) caused by different goals. Assume that the change in the benefit of a decision on a certain goal i is ΔRi, and the change in the total utility of the decision is ΔU, then the weight wi can be expressed as:

[0107] By solving the above formula, the weights of different knowledge clusters (i.e., decision-making objectives) can be obtained, thus achieving the goal of acquiring agricultural multi-carrier meta-knowledge.

[0108] 2) Information feedback adjustment

[0109] In view of the multi-source heterogeneity of agriculture, the feedback algorithm of network weight adjustment is studied. By reducing the weight, data resources irrelevant to public knowledge are gradually eliminated during the learning process to improve the accuracy of the learning task.

[0110] Based on the characteristics of multi-source heterogeneity in agriculture, the feedback algorithm for adjusting network weights is studied in the information feedback adjustment process. Assuming that ωij(t) is the weight corresponding to the jth knowledge node when the tth round network learns the i-th carrier sample data, the learning process can be expressed as ωij(t+1)=ωij(t)+α(t)dij, where α(t) is the decaying learning rate, and dij is the common metric between the j-th knowledge and the i-th carrier data, where the Euclidean distance is used. In this way, data resources that are not related to public knowledge are gradually eliminated during the learning process, thereby improving the accuracy of the learning task. It should be noted that in order to avoid overfitting, it is necessary to prune excessive weights, such as limiting ωij between 0 and 1.

[0111] Embodiment 1:

[0112] Assume that there are four agricultural data sources, namely A, B, C and D. A and D come from agricultural technology articles, B comes from farm experiment data, and C comes from farmers' experience sharing. The initial data contains the following information:

[0113] A: "In order to prevent and control diseases, 1 kg of pesticide should be applied per acre of land..."

[0114] B: "In 2019, the increase in hamster numbers caused a 5% decrease in wheat production..."

[0115] C: "Generally speaking, 2 days after it rains is a good time to sow soybeans..."

[0116] D: "When farmers plant rice in the field, the distance between each seedling should be kept at 40cm..."

[0117] First, self-organizing maps are used to simplify data A and D to reduce language redundancy. Then, feature selection based on spatial mapping selects meaningful features from A, B, C, and D and fuses them. For example, the amount of pesticide used is selected from A, the relationship between the number of hamsters and wheat yield is selected from B, the sowing time is selected from C, and the spacing between seedlings is selected from D.

[0118] Then, these features are mapped to a unified feature space for comparison and selection to form a feature fusion space, in which redundant information and noise are eliminated and common features are highlighted.

[0119] Next, we build a knowledge transfer network based on the graph model, use the feature model as a node, and initialize the graph model. Then, we transfer the knowledge based on label propagation and optimization of the migration relationship. For example, we find that both A and D have knowledge about reasonable planting measures, so we transfer the labels of "pesticide usage" and "seedling spacing" that have been marked in A to D.

[0120] Finally, through meta-knowledge acquisition and information feedback adjustment, meta-knowledge acquisition extracts meta-knowledge from agricultural multi-carrier data by following the principle of structural similarity, such as trying to extract general experience of sowing time from C. Information feedback adjustment is dynamically adjusted according to the feedback results in the learning process by adjusting the network weights.

[0121] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0122] It should be understood that the detailed description of the technical solutions of the present invention by means of the preferred embodiments is illustrative rather than restrictive. A person skilled in the art may modify the technical solutions described in the embodiments, or replace some of the technical features by equivalents, based on reading the specification of the present invention; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for agricultural multi-source knowledge transfer based on graph model, Features: The method includes: S1. Data simplification, feature selection and feature fusion of agricultural literature, agricultural data simplification based on self-organizing maps; feature selection and fusion based on spatial mapping; S2. Build a knowledge transfer network based on graph model, extract knowledge semantics, and mine knowledge transfer relationships; S3. Adjust the agricultural multi-carrier meta-knowledge acquisition and information feedback, meta-knowledge acquisition, and information feedback adjustment.

2. According to claim 1, a method for agricultural multi-source knowledge transfer based on a graph model, Features: The agricultural data reduction based on self-organizing mapping includes: (1) First, collect and organize agricultural multi-source data. Due to different agricultural characteristics or sources, there is a lot of duplicate or incomplete information; (2) Use the self-organizing map model SOM to map the input high-dimensional data vector into a grid structure in a low-dimensional space; (3) The "rule + exception" concept of rough sets is introduced into the SOM model. For rule theory, a set of objects that are consistent in all attributes is regarded as a concept or a rule, and other inconsistent objects are regarded as exceptions. By eliminating exception data and preserving rule data, the amount of data is reduced while retaining the main characteristics of the data.

3. According to claim 1, a method for agricultural multi-source knowledge transfer based on a graph model, Features: The feature selection and fusion based on spatial mapping includes: Feature selection is the process of selecting a subset of features from the original data set; this is done by calculating the correlation between different features; for each agricultural multi-carrier data, these features are mapped to a unified feature space for comparison and selection; Feature fusion maps all carrier data to the feature fusion space by constructing a mapping function; in the mapping space, redundant information and noise between features are eliminated, revealing the common features hidden in different data carriers.

4. The method for agricultural multi-source knowledge transfer based on a graph model according to claim 1, Features: The knowledge semantic extraction includes: Preprocessing: First, preprocess the agricultural multi-source feature data, including data cleaning and format conversion, so that the data can meet the needs of subsequent processing; Feature extraction: Use the TF-IDF algorithm to extract useful features from the preprocessed data.

5. The method for agricultural multi-source knowledge transfer based on a graph model according to claim 1, Features: The knowledge transfer relationship mining includes: Initialize the graph model: Use features as nodes to initially build a graph model; each node represents a feature, and the edge represents the relationship between features; Label propagation: Use semi-supervised learning to train the model using a small amount of labeled data, and build the model together with unlabeled data; if there is similarity between labeled samples and unlabeled samples, propagate the label of the labeled sample to similar unlabeled samples; Optimize migration relationships: Based on label propagation, use the annotation information of the auxiliary carrier to optimize the migration relationship of the target unlabeled samples; by adjusting the edge weights, nodes with high similarity have the same or similar labels.

6. The method for agricultural multi-source knowledge transfer based on a graph model according to claim 1, Features: The meta-knowledge acquisition includes: Firstly, by following the principle of structural similarity, universal meta-knowledge is extracted from agricultural multi-carrier data, and a one-to-one mapping relationship is established between them in the knowledge transfer network and the meta-knowledge network; by applying the rough set attribute reduction theory, extension knowledge is obtained from the semantic web, and the corresponding meta-knowledge base is established based on the meta-knowledge represented by extension transformation; Then, the weight relationship between different knowledge clusters is determined according to the Keeney-Raiffa algorithm; under the control of the knowledge cluster weights, the knowledge of different carriers will provide information support for the needs of agricultural knowledge transfer tasks.

7. The method for agricultural multi-source knowledge transfer based on a graph model according to claim 1, Features: The information feedback adjustment includes: Based on the characteristics of multi-source heterogeneity in agriculture, the information feedback adjustment process adopts a feedback algorithm of network weight adjustment; Assuming that ωij(t) is the weight corresponding to the jth knowledge node when the tth round network learns the i-th carrier sample data, the learning process is expressed as ωij(t+1)=ωij(t)+α(t)dij, Where α(t) is the decay learning rate, dij is the common metric between the jth knowledge and the ith carrier data, and the Euclidean distance is used here; In this way, data resources irrelevant to public knowledge are gradually eliminated during the learning process, thus improving the accuracy of the learning task.

8. The method for agricultural multi-source knowledge transfer based on a graph model according to claim 7, Features: To avoid overfitting, excessive weights are pruned and ωij is between 0 and 1.

Images