Animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method

Abstract

The invention discloses an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method. The method comprises the steps of obtaining feeding knowledge resource data of different animal types and constructing a feeding knowledge meta-database; performing multi-granularity entity recognition and analysis on the feeding knowledge resources in the feeding knowledge meta-database to generate four types of entity sets, constructing an entity interaction graph, performing entity relationship reasoning to obtain a triple set, and constructing a feeding resource knowledge graph by using the triple set; constructing a personalized recommendation engine, carrying out user question answering and demand knowledge recommendation through the personalized recommendation engine, and carrying out interest resource recommendation based on user use behaviors; and building an animal nutrition and feed course resource sharing cloud platform by utilizing a micro-service architecture technology based on the personalized recommendation engine, the feeding knowledge meta-database and the feeding resource knowledge graph. Therefore, accurate, real-time and multi-dimensional knowledge service and help are provided for animal husbandry practitioners or interests and amateurs.

Description

Technical Field

[0001] The present invention relates to the technical field of animal breeding resource sharing, and in particular to an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method. Background Art

[0002] In the field of animal nutrition and feed, existing educational resources and knowledge service systems generally have problems such as data silos, weak knowledge relevance, and insufficient personalized services. Traditional databases mostly use structured tables to store basic data such as breeding standards and feed ingredients, which makes it difficult to effectively express the dynamic changes in animal growth stages, the complex relationship between nutritional needs and environmental factors, and the data standards of different research institutions or enterprises are not unified, which makes cross-platform resource integration difficult. Unstructured resources such as teaching courses and experimental cases are usually stored in the form of independent files, lacking semantic annotation and knowledge association, so users need to rely on keyword matching when searching, and it is difficult to accurately obtain knowledge content adapted to specific species, physiological stages or production scenarios. Existing recommendation systems are mostly based on collaborative filtering based on explicit user behavior (such as clicks and favorites). They lack an understanding of industry knowledge logic and cannot combine professional rules of animal nutrition (such as nutritional balance models and feed incompatibility taboos) to make inference recommendations, resulting in limited scientificity and practicality of the recommended solutions.

[0003] At the knowledge organization level, existing technologies usually use hierarchical classification or labeling systems to manage resources, which makes it impossible to deeply explore the potential connections between heterogeneous data such as course videos, scientific research literature, and feeding logs. For example, an instructional video on a certain feed formula may involve multi-dimensional knowledge elements such as nutritional requirements for a specific growth stage, regional raw material substitution options, and economic benefit analysis. However, traditional methods make it difficult to systematically present these connections, resulting in inefficient knowledge transfer. In addition, knowledge in the field of animal nutrition is highly time-sensitive, and new scientific research results (such as verification of the effectiveness of new additives) and policy standards are frequently updated. The existing static knowledge base lacks a dynamic evolution mechanism, making it difficult to timely reflect cutting-edge progress in the industry.

[0004] In response to the above technical bottlenecks, a cloud platform for sharing animal nutrition and feed course resources and a knowledge graph construction method are proposed to provide accurate, real-time, and multi-dimensional knowledge services and assistance to livestock industry practitioners or enthusiasts. Summary of the Invention

[0005] The present invention overcomes the defects of the prior art and provides an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method.

[0006] To achieve the above objectives, the first aspect of the present invention provides an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method, comprising:

[0007] Acquiring animal husbandry knowledge resource data of different animal species and performing data preprocessing, and constructing an animal husbandry knowledge metadata database based on the preprocessed knowledge resource data;

[0008] Performing multi-granularity entity recognition and analysis on the breeding knowledge resources in the breeding knowledge metadata database to generate four entity sets, constructing an entity interaction graph based on the four entity sets, performing entity relationship reasoning to obtain a triple set, and constructing a breeding resource knowledge graph using the triple set;

[0009] Build a personalized recommendation engine to answer user questions and recommend knowledge they need, and recommend resources of interest based on user usage behavior;

[0010] Based on the personalized recommendation engine, feeding knowledge metadata database and feeding resource knowledge graph, a cloud platform for sharing animal nutrition and feed course resources is built using microservice architecture technology.

[0011] In this solution, the step of obtaining the animal husbandry knowledge resource data of different animal species and performing data preprocessing, and constructing the animal husbandry knowledge metadata database based on the preprocessed knowledge resource data, specifically includes:

[0012] Use distributed web crawler clusters to collect data to obtain feeding knowledge resource data of different animal species, and perform data preprocessing on the collected feeding knowledge resource data;

[0013] The regular expression engine is used to match nutritional indicator patterns in unstructured text, and entity annotation is performed using a preset domain dictionary. The entity is converted into a standardized key-value pair structure, and the isolation forest algorithm is used to detect and filter data outliers.

[0014] After data cleaning is completed, a local sensitive hashing algorithm is introduced to remove duplicates from the cleaned breeding knowledge resource data across sources, extract the text features of each breeding knowledge resource data and perform word segmentation to generate a word set;

[0015] Perform hash mapping on each term in the term set based on a preset hash function, select the minimum hash value through a random permutation strategy to form a signature vector, input it into the SimHash algorithm for feature space compression, and generate a feature fingerprint through weighted summation and binarization operations;

[0016] Constructing a distributed hash table to store the characteristic fingerprints in buckets, traversing each characteristic fingerprint in the bucket to calculate the Hamming distance and make a judgment with a preset threshold, and deduplicating the cleaned breeding knowledge resource data according to the judgment result;

[0017] The pre-processed breeding knowledge resource data is obtained, and a breeding knowledge metadata database is constructed according to the pre-processed knowledge resource data.

[0018] In this solution, the breeding knowledge resources in the breeding knowledge metadata database are subjected to multi-granularity entity recognition and analysis to generate four types of entity sets, specifically including:

[0019] Obtaining a breeding knowledge meta-database, connecting the breeding knowledge meta-database to a pre-trained domain language model to perform contextual semantic encoding on the multimodal data in the breeding knowledge meta-database, and introducing a multi-head attention mechanism to obtain attention weights between different modal knowledge data for representing association weights, thereby obtaining a set of semantically enhanced feature vectors;

[0020] Constructing and training a dual-channel entity recognition network, and inputting the semantically enhanced feature vector into the trained dual-channel entity recognition network to extract knowledge units, wherein the dual-channel entity recognition network includes a character-level feature extraction channel and a word-level semantic embedding channel;

[0021] In the character-level feature extraction channel, a character sequence of the input semantic enhancement feature vector set is scanned through a three-layer convolution kernel and underlying feature extraction is performed, feature dimensions are compressed through a maximum pooling operation, and finally a character-level feature vector is output;

[0022] Embedding a word segmenter pre-trained based on domain-specific knowledge in the word-level semantic embedding channel, segmenting the semantic units through the word segmenter when the semantic feature enhanced vector set is input, and mapping the segmented semantic units to the initial embedding space;

[0023] In the initial embedding space, a mixed semantic basis vector is obtained by dynamic expansion operation, and is input into the Transformer encoder for contextual semantic modeling. Features at different abstraction levels are gradually refined to analyze basic lexical associations and logic, and generate word-level semantic feature vectors.

[0024] The character-level feature vector and the word-level semantic feature vector are fused through a gated concatenation layer to generate a joint feature vector, which is then input into a bidirectional gated recurrent unit to analyze long-range context dependencies. A conditional random field is introduced to decode the entity type, and finally four types of entity sets are obtained, namely animal species, growth stage, nutritional indicators and feed composition factors.

[0025] In this solution, the entity interaction graph is constructed by the four types of entity sets to perform entity relationship reasoning to obtain a triple set, and the breeding resource knowledge graph is constructed using the triple set, which specifically includes:

[0026] Obtain four entity sets, and construct an entity interaction graph based on the four entity sets using four entity types, namely, animal species, growth stage, nutritional index, and feed composition factor, as nodes. Each entity node is assigned a time attribute, a data source attribute, and an entity type attribute. The co-occurrence frequency and syntactic dependency distance of each entity node in the text are used as edge weights between the entity nodes.

[0027] Importing the entity interaction graph into a graph attention network for relational reasoning, and embedding a time decay factor in the graph attention network for improvement, wherein the time decay factor is described by an exponential decay function;

[0028] The attention module set in the hidden layer of the network captures the association patterns of different semantic dimensions in parallel. The entity relationship probability distribution is obtained through softmax normalization in the output layer. Based on the entity relationship probability distribution, candidate triplets are generated. The confidence is calculated and the triplets exceeding the preset threshold are filtered to obtain the final triple set.

[0029] A cross-document path reasoning mechanism is set up to obtain isolated entity nodes through the entity interaction graph, and the breeding knowledge resources containing the target isolated entity nodes are retrieved from the breeding knowledge metadata database to generate multi-document indirect association paths, and valid paths are screened through credibility evaluation to generate supplementary triples;

[0030] Based on the final triple set being imported into the graph database, four types of entities are defined as graph core node types, and entity-node relationships are defined as edges to construct a breeding resource knowledge graph.

[0031] In this solution, the personalized recommendation engine is constructed to answer user questions and recommend required knowledge, and recommend resources of interest based on user usage behavior, specifically including:

[0032] Build a personalized recommendation engine, which includes an intelligent recommendation model and a user interest analysis model. The intelligent recommendation model is used to answer user questions and recommend required knowledge, and the user interest analysis model is used to analyze the interest trends of target users;

[0033] The intelligent recommendation model is constructed through a graph neural network and associated with the breeding resource knowledge graph. When the target user enters a query text, the pre-trained domain BERT model is used to identify the query text and generate a user query vector.

[0034] Input the user query vector into the breeding resource knowledge graph for cosine similarity matching, and select candidate resources with a similarity greater than a preset threshold and a preset resource quantity threshold to form an initial recommendation pool;

[0035] Feature extraction is performed on each candidate resource in the initial recommendation pool to obtain a candidate resource feature vector. A second-order subgraph centered on the candidate resource node is then constructed using a graph neural network. A three-layer graph convolutional network is used to aggregate neighborhood information. After three iterations, a graph embedding vector is output.

[0036] The candidate resource features and graph embedding vectors are input into the cross-attention fusion layer, and the dependency weights are obtained by linear projection alignment to the same dimension for weighted fusion. The weighted fused feature vectors are used as the input of the LambdaMART algorithm to sort the candidate resources, output the recommended resource sorting list, and make intelligent recommendations for feeding resources to the target users.

[0037] In this solution, the personalized recommendation engine is constructed to answer user questions and recommend required knowledge, and recommend resources of interest based on user usage behavior, and further includes:

[0038] A user interest analysis model is constructed based on a convolutional neural network. When a target user queries resources or asks for help, the target user's historical behavior data and real-time query information are obtained and input into the user interest analysis model for interest modeling.

[0039] When the historical behavior data and real-time query information of the target user are input, differential embedding processing is performed by inputting the historical behavior data and real-time query information of the target user into the embedding layer, discrete features are mapped into multi-dimensional dense vectors through the embedding table, and continuous features are compressed through quantile normalization processing to obtain embedded feature vectors;

[0040] The embedded feature vector is input into the hierarchical attention network. The bottom-level temporal attention module calculates the time decay weight of the behavior sequence, strengthens the semantic contribution of recent behavior, and uses the upper-level topic attention module to capture the user's focus of interest. Multi-scale convolution is used to parallelly process behavior sequences of different time granularities, extract local interest fluctuation patterns and long-term preference evolution trends, and finally output the target user's interest analysis information.

[0041] If there is no active knowledge or question query of the target user, the target user interest analysis information is imported into the breeding resource knowledge graph to match interest resources, generate an interest resource recommendation table and push interest resources;

[0042] If there is active knowledge or question query of the target user, an interest vector space is generated based on the target user interest analysis information, a ranked list of recommended resources is obtained, and each resource in the ranked list of recommended resources is mapped to the vector space using the TransR algorithm;

[0043] Calculate the multi-angle similarity between each resource item and the target user's interest vector, including cosine similarity, Jaccard index and KL divergence, to generate a comprehensive matching score. Based on the comprehensive matching score, use the MMR algorithm to re-rank the resource items in the recommended resource ranking list, and make intelligent recommendations for breeding resources based on the re-ranking results.

[0044] In this solution, the animal nutrition and feed course resource sharing cloud platform is built based on the personalized recommendation engine, feeding knowledge metadata database and feeding resource knowledge graph using microservice architecture technology, specifically including:

[0045] Build data collection management framework, knowledge graph query framework, and personalized recommendation framework based on personalized recommendation engine, breeding knowledge metadata database, and breeding resource knowledge graph;

[0046] The data collection and management framework includes a data collection unit and a data management unit, wherein the data collection unit is implemented by deploying a distributed crawler cluster, and the data management unit is associated with a breeding knowledge metadata database;

[0047] The knowledge graph query framework is built by the breeding resource knowledge graph, which is used to manage knowledge graph updates and maintenance, and realize knowledge graph query and interaction through visualization technology;

[0048] The personalized recommendation framework is built by the personalized recommendation engine and is used to analyze the real-time query text and historical behavior of the target user to recommend knowledge resources and resources of interest to the user;

[0049] Based on the microservice architecture technology, a data collection and management framework, knowledge graph query framework, and personalized recommendation framework are constructed to build an animal nutrition and feed course resource sharing cloud platform.

[0050] The second aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method program. When the program applicable to the animal nutrition and feed course resource sharing cloud platform and the knowledge graph construction method is executed by a processor, the steps of the animal nutrition and feed course resource sharing cloud platform and the knowledge graph construction method as described above are implemented.

[0051] The present invention discloses an animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method, comprising the following steps: obtaining feeding knowledge resource data for different animal species and constructing a feeding knowledge metadata database; performing multi-granular entity recognition and analysis on the feeding knowledge resources in the feeding knowledge metadata database to generate four entity sets, constructing an entity interaction graph to perform entity relationship reasoning to obtain a triple set, and using the triple set to construct a feeding resource knowledge graph; constructing a personalized recommendation engine to answer user questions and recommend required knowledge, and recommend resources of interest based on user usage behavior; and constructing an animal nutrition and feed course resource sharing cloud platform based on the personalized recommendation engine, the feeding knowledge metadata database, and the feeding resource knowledge graph using microservice architecture technology. This provides accurate, real-time, and multi-dimensional knowledge services and assistance to livestock industry practitioners or enthusiasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0053] Figure 1 A flow chart of an animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method provided by one embodiment of the present invention;

[0054] Figure 2 A flow chart of a first method for constructing a breeding resource knowledge graph according to an embodiment of the present invention;

[0055] Figure 3 A flow chart of a second construction method of a breeding resource knowledge graph provided by a real-time example of the present invention;

[0056] Figure 4 A block diagram of an animal nutrition and feed course resource sharing cloud platform provided in one embodiment of the present invention;

[0057] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0058] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0059] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0060] Figure 1 A flow chart of an animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method provided by one embodiment of the present invention;

[0061] like Figure 1 As shown, the present invention provides an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method flow chart, including:

[0062] S102, obtaining animal husbandry knowledge resource data of different animal species and performing data preprocessing, and constructing an animal husbandry knowledge metadata database based on the preprocessed knowledge resource data;

[0063] S104, performing multi-granularity entity recognition and analysis on the breeding knowledge resources in the breeding knowledge metadata database to generate four entity sets, constructing an entity interaction graph based on the four entity sets, performing entity relationship reasoning to obtain a triple set, and constructing a breeding resource knowledge graph using the triple set;

[0064] S106, building a personalized recommendation engine, answering user questions and recommending required knowledge through the personalized recommendation engine, and recommending resources of interest based on user usage behavior;

[0065] S108, based on the personalized recommendation engine, feeding knowledge metadata database and feeding resource knowledge graph, uses microservice architecture technology to build an animal nutrition and feed course resource sharing cloud platform.

[0066] Furthermore, in a preferred embodiment of the present invention, the step of obtaining the animal husbandry knowledge resource data of different animal species and performing data preprocessing, and constructing the animal husbandry knowledge metadata database based on the preprocessed knowledge resource data, specifically includes:

[0067] Use distributed web crawler clusters to collect data to obtain feeding knowledge resource data of different animal species, and perform data preprocessing on the collected feeding knowledge resource data;

[0068] The regular expression engine is used to match nutritional indicator patterns in unstructured text, and entity annotation is performed using a preset domain dictionary. The entity is converted into a standardized key-value pair structure, and the isolation forest algorithm is used to detect and filter data outliers.

[0069] After data cleaning is completed, a local sensitive hashing algorithm is introduced to remove duplicates from the cleaned breeding knowledge resource data across sources, extract the text features of each breeding knowledge resource data and perform word segmentation to generate a word set;

[0070] Perform hash mapping on each term in the term set based on a preset hash function, select the minimum hash value through a random permutation strategy to form a signature vector, input it into the SimHash algorithm for feature space compression, and generate a feature fingerprint through weighted summation and binarization operations;

[0071] Constructing a distributed hash table to store the characteristic fingerprints in buckets, traversing each characteristic fingerprint in the bucket to calculate the Hamming distance and make a judgment with a preset threshold, and deduplicating the cleaned breeding knowledge resource data according to the judgment result;

[0072] The pre-processed breeding knowledge resource data is obtained, and a breeding knowledge metadata database is constructed according to the pre-processed knowledge resource data.

[0073] It should be noted that this approach first uses a distributed web crawler cluster to automatically crawl heterogeneous data from multiple sources. A dynamic crawler node scheduling strategy adapts to the anti-crawling mechanisms of different websites, collecting structured and unstructured data from sources such as scientific research literature databases, aquaculture enterprise technical manuals, and government public standard documents. For numerical nutritional indicators embedded in unstructured text, a regular expression engine is used to match feature patterns in a pre-set pattern rule library. Entity boundaries are identified and type annotations are performed using a domain-specific dictionary, converting the data into a standardized key-value pair structure. To eliminate noise or outliers that may have been introduced during data collection, an unsupervised detection model is constructed using the Isolation Forest algorithm. Random subspace sampling generates 100 isolation trees, calculates anomaly scores for each data point, and filters suspected anomalies with scores above a threshold. After initial cleaning, the cross-source data deduplication process is initiated. Near-duplicate detection of text resources is performed using the Locality Sensitive Hashing (LSH) algorithm: First, TF-IDF-weighted bag-of-words model features are extracted from the document. A signature matrix is ​​generated using the MinHash algorithm. A minimum hash operation is performed 128 times on the set of terms in each document to form a compact signature vector. The SimHash algorithm is then used to compress the signature vector's feature space, calculating the hash value of each feature dimension using a weighted summation method. A 64-bit feature fingerprint is then generated through binarization. A distributed storage architecture based on a consistent hashing ring is constructed, mapping the feature fingerprints to 1024 virtual buckets. Within each bucket, the Hamming distance is calculated in parallel using multiple threads. When the distance between two fingerprints falls below a preset threshold (set to 3, corresponding to 95% similarity), the content is identified as duplicate, triggering a deduplication operation that retains only the authoritative source version.

[0074] Subsequently, a feeding knowledge metadata database was constructed based on the high-quality data after cleaning and deduplication. The database adopts a hybrid storage architecture: structured data (such as nutritional standard tables and feed composition matrices) are stored in a PostgreSQL relational database, and dynamic attributes (such as environmental correction factors) are stored using JSONB type fields; unstructured documents (such as PDF technical manuals and experimental videos) are stored in a distributed object storage system, and full-text retrieval is achieved by establishing an inverted index; semi-structured data (such as feeding logs in XML format) are parsed and stored in a MongoDB document database, retaining the original hierarchical relationship. The metadata database is designed with a unified semantic identification specification, and each data entry is attached with source metadata (collection time, data source credibility score) and preprocessing logs (cleaning and deduplication operation records), forming a traceable and scalable basic data storage system to provide high-quality data support for subsequent knowledge graph construction and intelligent services.

[0075] Furthermore, in a preferred embodiment of the present invention, the personalized recommendation engine is constructed to answer user questions and recommend required knowledge, and recommend resources of interest based on user usage behavior, specifically including:

[0076] Build a personalized recommendation engine, which includes an intelligent recommendation model and a user interest analysis model. The intelligent recommendation model is used to answer user questions and recommend required knowledge, and the user interest analysis model is used to analyze the interest trends of target users;

[0077] The intelligent recommendation model is constructed through a graph neural network and associated with the breeding resource knowledge graph. When the target user enters a query text, the pre-trained domain BERT model is used to identify the query text and generate a user query vector.

[0078] Input the user query vector into the breeding resource knowledge graph for cosine similarity matching, and select candidate resources with a similarity greater than a preset threshold and a preset resource quantity threshold to form an initial recommendation pool;

[0079] Feature extraction is performed on each candidate resource in the initial recommendation pool to obtain a candidate resource feature vector. A second-order subgraph centered on the candidate resource node is then constructed using a graph neural network. A three-layer graph convolutional network is used to aggregate neighborhood information. After three iterations, a graph embedding vector is output.

[0080] The candidate resource features and graph embedding vectors are input into the cross-attention fusion layer, and the dependency weights are obtained by linear projection alignment to the same dimension for weighted fusion. The weighted fused feature vectors are used as the input of the LambdaMART algorithm to sort the candidate resources, output the recommended resource sorting list, and make intelligent recommendations for feeding resources to the target users.

[0081] It should be noted that by building a personalized recommendation engine, target users can obtain knowledge resources that are more in line with their expectations when using the resource sharing cloud platform, and further provide interest-based knowledge supplements through the user interest analysis model, thereby improving the user's understanding of preferred knowledge. When the user submits a query request, the pre-trained domain BERT model performs multi-granularity analysis on the input text: the underlying Transformer encoder extracts word-level context features, and the top-level classifier identifies intent categories (such as nutritional needs queries, disease diagnosis consultations), and generates a 768-dimensional query semantic vector. The cosine similarity between this vector and the entity vectors of course resources, literature abstracts, etc. stored in the knowledge graph is calculated, and the top 200 candidate resources with a similarity exceeding 0.82 and meeting resource type constraints (such as videos, papers, and datasets) are selected to form the initial recommendation pool.

[0082] For each resource node in the candidate pool, a local subgraph of the knowledge graph is expanded: Starting from the resource node, two layers of neighborhoods are expanded outward along relationship edges (e.g., "associated nutritional standards" and "cited experimental cases") to capture contextual entities such as associated animal species, growth stages, and environmental parameters. A three-layer graph convolutional network (GCN) is used to aggregate neighborhood information. Each layer integrates the type characteristics of one-hop neighboring nodes (e.g., nutritional indicator value ranges, feed composition properties) through a message passing mechanism. After three iterations, a 256-dimensional graph embedding vector is generated, encoding the structural association characteristics of the resource in the knowledge network. Simultaneously, multimodal features are extracted from the candidate resources themselves: semantic vectors are generated for text descriptions using Sentence-BERT, Inception-v3 visual features are extracted from keyframes of video resources, and structured data (e.g., feed formula tables) are converted into numerical feature vectors. The graph embedding vector and the resource's native features are fed into a cross-attention fusion layer. After linear projection to align them to the same dimension, a bidirectional attention weight matrix is ​​calculated to reveal potential dependencies between features (e.g., the strength of theoretical support for video content and related literature). Subsequently, the weighted fused multi-dimensional joint feature vector is input into the LambdaMART sorting algorithm, and the gradient boosting decision tree framework is used to gradually optimize the NDCG index through 50 regression trees, learn the nonlinear correlation law between resource characteristics and user click behavior, and finally output a recommendation list sorted in descending order of relevance, thereby completing the recommendation of breeding knowledge resources for the target user's query information.

[0083] Furthermore, in a preferred embodiment of the present invention, the construction of a personalized recommendation engine, answering user questions and recommending required knowledge through the personalized recommendation engine, and recommending resources of interest based on user usage behavior, further includes:

[0084] A user interest analysis model is constructed based on a convolutional neural network. When a target user queries resources or asks for help, the target user's historical behavior data and real-time query information are obtained and input into the user interest analysis model for interest modeling.

[0085] When the historical behavior data and real-time query information of the target user are input, differential embedding processing is performed by inputting the historical behavior data and real-time query information of the target user into the embedding layer, discrete features are mapped into multi-dimensional dense vectors through the embedding table, and continuous features are compressed through quantile normalization processing to obtain embedded feature vectors;

[0086] The embedded feature vector is input into the hierarchical attention network. The bottom-level temporal attention module calculates the time decay weight of the behavior sequence, strengthens the semantic contribution of recent behavior, and uses the upper-level topic attention module to capture the user's focus of interest. Multi-scale convolution is used to parallelly process behavior sequences of different time granularities, extract local interest fluctuation patterns and long-term preference evolution trends, and finally output the target user's interest analysis information.

[0087] If there is no active knowledge or question query of the target user, the target user interest analysis information is imported into the breeding resource knowledge graph to match interest resources, generate an interest resource recommendation table and push interest resources;

[0088] If there is active knowledge or question query of the target user, an interest vector space is generated based on the target user interest analysis information, a ranked list of recommended resources is obtained, and each resource in the ranked list of recommended resources is mapped to the vector space using the TransR algorithm;

[0089] Calculate the multi-angle similarity between each resource item and the target user's interest vector, including cosine similarity, Jaccard index and KL divergence, to generate a comprehensive matching score. Based on the comprehensive matching score, use the MMR algorithm to re-rank the resource items in the recommended resource ranking list, and make intelligent recommendations for breeding resources based on the re-ranking results.

[0090] It should be noted that the user interest analysis model based on convolutional neural networks first jointly processes user historical behavior data and real-time query information through a feature engineering module. Discrete features (user ID, resource type, operation type, etc.) in historical behavior data (such as resource browsing records and question-and-answer interaction logs in the past 30 days) are mapped into 64-dimensional dense vectors through an embedding table; continuous features (single session duration, daily visit frequency, etc.) are quantile normalized to eliminate dimensional differences and then compressed to 32 dimensions by inputting the fully connected layer. The real-time query text generates a 128-dimensional semantic vector through a lightweight BERT model, which is then concatenated with the historical feature vector to form a fused input vector to construct a temporal behavior feature matrix. This is then fed into a hierarchical attention network for deep modeling. The underlying temporal attention module uses an exponential decay function (decay factor λ = 0.1) to calculate the temporal weight of behavioral sequences, increasing the contribution of behaviors from the past seven days. The upper-level topic attention module predefines eight domain topic vectors (e.g., feed ratio optimization, disease prevention and control techniques). Using a cross-attention mechanism, it dynamically calculates the correlation strength between each topic and user behavior, screening out core topics of interest. Multi-scale convolutional layers are then used to process temporal features in parallel. Three convolution kernels (widths 4, 8, and 12) are used to extract hourly interest fluctuations (e.g., frequent nutritional index queries during lunch), daily preference evolution (e.g., a focus on economic analysis on weekends), and weekly trend changes (e.g., a gradual focus on specific animal species). A gating mechanism then fuses these multi-granular features to output a 256-dimensional user interest vector.

[0091] When the user is in a passive interaction state (i.e., not performing specific knowledge retrieval or querying), the interest vector is matched against the entity embeddings in the animal husbandry resource knowledge graph using cosine similarity. A two-hop graph traversal is performed along the "user interest-resource topic-related knowledge point" path. Resource items with a match degree greater than 0.78 are selected to generate an interest recommendation table, thereby automatically recommending knowledge of interest and providing potential content of interest to the target user. When the user initiates an active query (i.e., performing specific knowledge retrieval or querying), the intelligent recommendation model prioritizes the user's query content and obtains a ranked list of recommended resources. Next, an interest vector space is generated based on the target user's interest analysis information. The ranked list of recommended resources is projected into a unified vector space using the TransR algorithm. Multi-dimensional similarity between each resource item and the target user's interest vector is calculated: cosine similarity measures directional consistency, Jaccard index assesses label overlap, and KL divergence measures distributional differences. A weighted fusion results in a comprehensive matching score. The MMR (Maximum Margin Relevance) algorithm is used to balance the relevance and diversity of recommendation results. The resource items with the highest comprehensive score and the largest distance from the boundary of the selected results are iteratively selected through a greedy strategy, so that the recommendation list covers 3-5 subtopics while maintaining core relevance. Finally, an optimized intelligent recommendation queue is output, which integrates user interest knowledge on the basis of query knowledge-based resource recommendation, thereby providing more comprehensive knowledge resource recommendations.

[0092] Furthermore, in a preferred embodiment of the present invention, the animal nutrition and feed course resource sharing cloud platform is built based on the personalized recommendation engine, the feeding knowledge metadata database and the feeding resource knowledge graph using microservice architecture technology, specifically including:

[0093] Build data collection management framework, knowledge graph query framework, and personalized recommendation framework based on personalized recommendation engine, breeding knowledge metadata database, and breeding resource knowledge graph;

[0094] The data collection and management framework includes a data collection unit and a data management unit, wherein the data collection unit is implemented by deploying a distributed crawler cluster, and the data management unit is associated with a breeding knowledge metadata database;

[0095] The knowledge graph query framework is built by the breeding resource knowledge graph, which is used to manage knowledge graph updates and maintenance, and realize knowledge graph query and interaction through visualization technology;

[0096] The personalized recommendation framework is built by the personalized recommendation engine and is used to analyze the real-time query text and historical behavior of the target user to recommend knowledge resources and resources of interest to the user;

[0097] Based on the microservice architecture technology, a data collection and management framework, knowledge graph query framework, and personalized recommendation framework are constructed to build an animal nutrition and feed course resource sharing cloud platform.

[0098] It should be noted that the data collection and management framework, knowledge graph query framework, and personalized recommendation framework are constructed based on the personalized recommendation engine, the husbandry knowledge metadata database, and the husbandry resource knowledge graph. The data collection and management framework dynamically schedules node resources through a distributed crawler cluster and designs differentiated crawling strategies for target sites, such as scientific research literature repositories, husbandry enterprise knowledge bases, and educational institution course repositories. The collected raw data is input into the data management unit for pipeline processing. Structured fields are extracted using regular expression matching with domain dictionaries. An isolation forest algorithm is used to detect outliers (e.g., indicator values ​​outside the reasonable range for a species' physiological stage). A hybrid MinHash-SimHash deduplication algorithm is used to eliminate cross-source redundant data. Finally, the cleaned and standardized data is persistently stored in the husbandry knowledge metadata database, forming a multimodal resource pool consisting of text, tables, videos, and experimental datasets. The knowledge graph query framework is built around the husbandry resource knowledge graph, storing the entity-relationship network in the Neo4j graph database. Graph computing engines (such as Apache TinkerPop) are used to support complex traversal queries, and visualization techniques are used to enable knowledge graph query and interaction. The personalized recommendation framework is built on the personalized recommendation engine. When a user initiates a query request, the API gateway routes the request to the recommendation computing service, which uses a pre-trained domain-specific BERT model to parse the query text and generate an intent vector. It also extracts the last 30 days of interaction records (such as resource browsing and favorites) from the user behavior log library and feeds them into a convolutional neural network interest model to generate a multi-dimensional dynamic interest vector. After retrieving candidate resources based on the knowledge graph, it generates a recommendation list sorted in descending order of overall match. Recommendations are then optimized based on user interests to provide more comprehensive resource recommendations.

[0099] Based on the microservice architecture, the three major frameworks are integrated and divided into independent modules such as data collection service, graph management service, recommendation computing service, and resource storage service. Each service communicates through the RESTful API and gRPC protocol, and the Kubernetes cluster implements containerized orchestration and elastic scaling. At the data flow level, user query requests are distributed to the corresponding services after load balancing by the API gateway. Asynchronous messages connect the data processing and recommendation generation links through the Kafka queue to ensure system stability in high-concurrency scenarios. At the operation and maintenance level, the Prometheus monitoring stack is integrated to collect container indicators (CPU / memory usage, request latency) in real time. When the service load exceeds the threshold, horizontal expansion is automatically triggered. At the same time, the entire link behavior is tracked through the ELK log analysis platform to provide data support for fault location and performance optimization. A closed loop is achieved from data collection, knowledge construction to intelligent services, thereby providing efficient knowledge retrieval and personalized resource recommendations, and promoting the accurate dissemination and collaborative innovation of knowledge in the field of animal nutrition.

[0100] Figure 2 A first construction method process of a breeding resource knowledge graph provided by an embodiment of the present invention;

[0101] like Figure 2 As shown, the present invention provides a flow chart of a first construction method of a breeding resource knowledge graph, including:

[0102] S202, obtaining a breeding knowledge meta-database, connecting the breeding knowledge meta-database to a pre-trained domain language model, performing contextual semantic encoding on the multimodal data in the breeding knowledge meta-database, and introducing a multi-head attention mechanism to obtain attention weights between different modal knowledge data for representing association weights, thereby obtaining a set of semantically enhanced feature vectors;

[0103] S204, constructing and training a dual-channel entity recognition network, inputting the semantic enhancement feature vector into the trained dual-channel entity recognition network to extract knowledge units, wherein the dual-channel entity recognition network includes a character-level feature extraction channel and a word-level semantic embedding channel;

[0104] S206, in the character-level feature extraction channel, scanning the character sequence of the input semantic enhancement feature vector set through three layers of convolution kernels and performing underlying feature extraction, compressing the feature dimension through a maximum pooling operation, and finally outputting a character-level feature vector;

[0105] S208, embedding a word segmenter pre-trained based on domain-specific knowledge in the word-level semantic embedding channel, segmenting the semantic units through the word segmenter when the semantic feature enhanced vector set is input, and mapping the segmented semantic units to the initial embedding space;

[0106] S210, obtaining a hybrid semantic basis vector by a dynamic expansion operation in the initial embedding space, inputting the hybrid semantic basis vector into a Transformer encoder for contextual semantic modeling, and gradually refining features at different abstraction levels to analyze basic lexical associations and logic, thereby generating a word-level semantic feature vector;

[0107] In S212, the character-level feature vector and the word-level semantic feature vector are fused through a gated concatenation layer to generate a joint feature vector, which is then input into a bidirectional gated recurrent unit to analyze long-range context dependencies. A conditional random field is introduced to decode the entity type, and finally four entity sets are obtained, namely, animal species, growth stage, nutritional indicators, and feed composition factors.

[0108] It should be noted that multimodal data from the animal husbandry knowledge metadata database is first fed into a pre-trained domain language model for deep semantic encoding. A multi-head attention mechanism dynamically calculates the association weights between data from different modalities. For example, a cross-modal attention matrix is ​​used to construct feature maps between text paragraphs and numerical indicators in corresponding experimental tables. Keywords in video descriptions are aligned with time series monitoring data using a spatiotemporal attention layer. After iterative processing through a multi-layer Transformer encoder, heterogeneous data such as text, tables, and video subtitles are uniformly encoded into multi-dimensional semantically enhanced feature vectors. Each vector carries cross-modal association weights and contextual semantic information, forming a unified feature representation that can be used for downstream tasks. Based on the semantically enhanced feature vectors, a dual-channel entity recognition network is used for joint feature learning. The character-level feature extraction channel utilizes a three-layer convolutional architecture to process the raw character sequence: the first layer with 3×50 convolutional kernels captures short-range character combination patterns, the second layer with 5×100 convolutional kernels identifies medium-span morphological features (e.g., cross-character associations in "rumen protein"), and the third layer with 7×150 convolutional kernels extracts long-range term structure patterns (e.g., the negative prefix modification relationship in "non-starch polysaccharides"). Each convolution layer is followed by a max-pooling operation to compress redundant features, ultimately outputting a character-level feature vector that prioritizes the regularity of the specialized terminology. A concurrently running word-level semantic embedding pipeline processes the input text using a domain-specific tokenizer, forcing specialized terminology to be retained as complete semantic units. The segmented word sequence is mapped to a multidimensional vector space via a dynamically expanded embedding layer, where pre-trained BERT weights are loaded for common vocabulary and new domain terms are initialized using a learnable parameter matrix. The embedded word vectors are fed into a 12-layer Transformer encoder for contextual modeling. The bottom-level network learns basic lexical associations (e.g., the numerical unit binding of "metabolic energy" to "kcal / kg"). The middle-level network parses complex modifier relationships (e.g., the spatiotemporal constraints on nutritional requirements during the third week after weaning). The high-level network establishes cross-sentence logical reasoning (e.g., the causal relationship between "high temperature environment" and "increased nutrient requirements"). The final output word-level semantic vector integrates multi-granular semantic features to provide a context-sensitive distributed representation for entity recognition. Subsequently, a gated concatenation layer was used to achieve dual-channel feature fusion, and a dynamic weight allocation mechanism was designed: when processing morphologically complex terms, the gating signal favors character-level features (weight 0.7); when facing semantically ambiguous scenarios, the weight of word-level semantic features (0.8) was enhanced. The fused multidimensional joint feature vector was input into a bidirectional gated recurrent unit (BiGRU), and the multidimensional hidden layer was used to capture long-range dependencies through time-step memory gating. Finally, the conditional random field decoding layer was connected. Domain constraints were hard-coded in the label transfer matrix (for example, prohibiting illegal transfers from "animal species to feed ingredients"). The optimal label sequence was solved using the Viterbi algorithm, outputting four types of entities: animal species, growth stage, nutritional index, and feed ingredient factor.

[0109] Figure 3 A flow chart of a second method for constructing a breeding resource knowledge graph according to an embodiment of the present invention;

[0110] like Figure 3 As shown, the present invention provides a flow chart of a second construction method of a breeding resource knowledge graph, including:

[0111] S302: Acquire four entity sets, and construct an entity interaction graph based on the four entity sets using four entity types, namely, animal species, growth stage, nutritional index, and feed ingredient factor, as nodes. Each entity node is assigned a time attribute, a data source attribute, and an entity type attribute. The co-occurrence frequency and syntactic dependency distance of each entity node in the text are used as edge weights between the entity nodes.

[0112] S304: importing the entity interaction graph into a graph attention network for relational reasoning, and embedding a time decay factor in the graph attention network for improvement, wherein the time decay factor is described by an exponential decay function;

[0113] S306, capturing association patterns of different semantic dimensions in parallel through the attention module set in the network hidden layer, obtaining the entity relationship probability distribution through softmax normalization in the output layer, generating candidate triples based on the entity relationship probability distribution, calculating the confidence and screening the triples that exceed the preset threshold to obtain the final triple set;

[0114] S308, setting a cross-document path reasoning mechanism, obtaining isolated entity nodes through the entity interaction graph, searching the breeding knowledge resources containing the target isolated entity nodes in the breeding knowledge metadata database to generate multi-document indirect association paths, and screening valid paths through credibility evaluation to generate supplementary triples;

[0115] S310, based on the final triple set imported into the graph database, four types of entities are defined as graph core node types, and entity node relationships are defined as edges to construct a breeding resource knowledge graph.

[0116] It should be noted that the entity interaction graph is first constructed based on the four entity sets of animal species, growth stages, nutritional indicators and feed composition factors. Each entity node is attached with multi-dimensional attributes: the time attribute records the effective period of the entity (such as the start / end timestamp of "90-150 days of beef cattle fattening period"), the data source attribute marks the document type and authoritative score, and the entity type attribute is encoded as a classification identifier (01-animal species, 02-growth stage, etc.). The edge weight is generated through composite calculation: the co-occurrence frequency of the entity in the text is given a basic weight, the path length extracted by the syntactic dependency analysis (+0.3 when the dependency distance ≤ 3) reflects the semantic closeness, and finally normalized to a comprehensive correlation strength in the range of 0-1. Among them, the syntactic dependency distance is defined by calculating the difference in the linear position numbers of two words with a dependency relationship in the sentence (that is, the absolute difference in the word order index in the sentence). The entity interaction graph is imported into the graph attention network for relational reasoning. The time decay factor is embedded in the network structure, and the exponential function e is used. -λΔt Edge weights are dynamically adjusted, with λ = 0.15 controlling the decay rate and Δt being the difference between the midpoints of the two entities' time periods (e.g., if the time difference between "week 2 of lactation" and "lysine requirement" is 0, the decay factor is 1). In the attention module, eight independent attention heads are used to capture patterns in different semantic dimensions, such as nutritional requirement associations, component inhibition effects, and environmental correction relationships. A hierarchical attention fusion module (weighting: 0.4 for the nutritional requirement head, 0.3 for the inhibition effect head, and 0.05 for each of the other heads) is used to enhance core relationship recognition. The output layer is softmax-normalized to generate a probability distribution of 10 relationships. Candidate triplets with a confidence score > 0.85 are screened and automatically annotated with spatiotemporal constraints to form a preliminary set of triplets.

[0117] Furthermore, for isolated entity nodes in the graph (such as the new additive "phytase"), the cross-document path reasoning engine is activated. First, a full-text search is performed in the breeding knowledge metadata database to extract the document set containing the target entity. The potential association path (such as "phytase → phosphorus utilization → laying hen egg production rate") is constructed through co-occurrence network analysis, the path credibility is calculated, and the path with a preset threshold is retained to generate supplementary triples. For indirect association paths, the TransH algorithm is used to verify the semantic rationality: the embedding similarity of the path endpoint entities is calculated. If it is greater than the preset embedding similarity threshold, the path is accepted as a valid association. Finally, the final triple set is imported into the graph database, and the core node types are defined as four categories: animal species, growth stage, nutritional indicators, and feed ingredient factors. The entity node relationship is the edge to construct the breeding resource knowledge graph.

[0118] The second aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method program. When the program applicable to the animal nutrition and feed course resource sharing cloud platform and the knowledge graph construction method is executed by a processor, the steps of the animal nutrition and feed course resource sharing cloud platform and the knowledge graph construction method as described above are implemented.

[0119] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0120] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0121] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0122] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0123] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method, characterized in that: include: Acquiring animal husbandry knowledge resource data of different animal species and performing data preprocessing, and constructing an animal husbandry knowledge metadata database based on the preprocessed knowledge resource data; Performing multi-granularity entity recognition and analysis on the breeding knowledge resources in the breeding knowledge metadata database to generate four entity sets, constructing an entity interaction graph based on the four entity sets, performing entity relationship reasoning to obtain a triple set, and constructing a breeding resource knowledge graph using the triple set; Build a personalized recommendation engine to answer user questions and recommend knowledge they need, and recommend resources of interest based on user usage behavior; Based on the personalized recommendation engine, feeding knowledge metadata database and feeding resource knowledge graph, a cloud platform for sharing animal nutrition and feed course resources is built using microservice architecture technology.

2. The animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method according to claim 1, characterized in that: The step of obtaining the breeding knowledge resource data of different animal species and performing data preprocessing, and constructing a breeding knowledge metadata database based on the preprocessed knowledge resource data, specifically includes: Use distributed web crawler clusters to collect data to obtain feeding knowledge resource data of different animal species, and perform data preprocessing on the collected feeding knowledge resource data; The regular expression engine is used to match nutritional indicator patterns in unstructured text, and entity annotation is performed using a preset domain dictionary. The entity is converted into a standardized key-value pair structure, and the isolation forest algorithm is used to detect and filter data outliers. After data cleaning is completed, a local sensitive hashing algorithm is introduced to remove duplicates from the cleaned breeding knowledge resource data across sources, extract the text features of each breeding knowledge resource data and perform word segmentation to generate a word set; Perform hash mapping on each term in the term set based on a preset hash function, select the minimum hash value through a random permutation strategy to form a signature vector, input it into the SimHash algorithm for feature space compression, and generate a feature fingerprint through weighted summation and binarization operations; Constructing a distributed hash table to store the characteristic fingerprints in buckets, traversing each characteristic fingerprint in the bucket to calculate the Hamming distance and make a judgment with a preset threshold, and deduplicating the cleaned breeding knowledge resource data according to the judgment result; The pre-processed breeding knowledge resource data is obtained, and a breeding knowledge metadata database is constructed according to the pre-processed knowledge resource data.

3. The animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method according to claim 1, characterized in that: The multi-granularity entity recognition and analysis is performed on the breeding knowledge resources in the breeding knowledge metadata database to generate four types of entity sets, specifically including: Obtaining a breeding knowledge meta-database, connecting the breeding knowledge meta-database to a pre-trained domain language model to perform contextual semantic encoding on the multimodal data in the breeding knowledge meta-database, and introducing a multi-head attention mechanism to obtain attention weights between different modal knowledge data for representing association weights, thereby obtaining a set of semantically enhanced feature vectors; Constructing and training a dual-channel entity recognition network, and inputting the semantically enhanced feature vector into the trained dual-channel entity recognition network to extract knowledge units, wherein the dual-channel entity recognition network includes a character-level feature extraction channel and a word-level semantic embedding channel; In the character-level feature extraction channel, a character sequence of the input semantic enhancement feature vector set is scanned through a three-layer convolution kernel and underlying feature extraction is performed, feature dimensions are compressed through a maximum pooling operation, and finally a character-level feature vector is output; Embedding a word segmenter pre-trained based on domain-specific knowledge in the word-level semantic embedding channel, segmenting the semantic units through the word segmenter when the semantic feature enhanced vector set is input, and mapping the segmented semantic units to the initial embedding space; In the initial embedding space, a mixed semantic basis vector is obtained by dynamic expansion operation, and is input into the Transformer encoder for contextual semantic modeling. Features at different abstraction levels are gradually refined to analyze basic lexical associations and logic, and generate word-level semantic feature vectors. The character-level feature vector and the word-level semantic feature vector are fused through a gated concatenation layer to generate a joint feature vector, which is then input into a bidirectional gated recurrent unit to analyze long-range context dependencies. A conditional random field is introduced to decode the entity type, and finally four types of entity sets are obtained, namely animal species, growth stage, nutritional indicators and feed composition factors.

4. The animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method according to claim 1, characterized in that: The method of constructing an entity interaction graph through the four types of entity sets and performing entity relationship reasoning to obtain a triple set, and constructing a breeding resource knowledge graph using the triple set, specifically includes: Obtain four entity sets, and construct an entity interaction graph based on the four entity sets using four entity types, namely, animal species, growth stage, nutritional index, and feed composition factor, as nodes. Each entity node is assigned a time attribute, a data source attribute, and an entity type attribute. The co-occurrence frequency and syntactic dependency distance of each entity node in the text are used as edge weights between the entity nodes. Importing the entity interaction graph into a graph attention network for relational reasoning, and embedding a time decay factor in the graph attention network for improvement, wherein the time decay factor is described by an exponential decay function; The attention module set in the hidden layer of the network captures the association patterns of different semantic dimensions in parallel. The entity relationship probability distribution is obtained through softmax normalization in the output layer. Based on the entity relationship probability distribution, candidate triplets are generated. The confidence is calculated and the triplets exceeding the preset threshold are filtered to obtain the final triple set. A cross-document path reasoning mechanism is set up to obtain isolated entity nodes through the entity interaction graph, and the breeding knowledge resources containing the target isolated entity nodes are retrieved from the breeding knowledge metadata database to generate multi-document indirect association paths, and valid paths are screened through credibility evaluation to generate supplementary triples; Based on the final triple set being imported into the graph database, four types of entities are defined as graph core node types, and entity-node relationships are defined as edges to construct a breeding resource knowledge graph.

5. The animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method according to claim 1, characterized in that: The personalized recommendation engine is constructed to answer user questions and recommend required knowledge, and recommend resources of interest based on user usage behavior, specifically including: Build a personalized recommendation engine, which includes an intelligent recommendation model and a user interest analysis model. The intelligent recommendation model is used to answer user questions and recommend required knowledge, and the user interest analysis model is used to analyze the interest trends of target users; The intelligent recommendation model is constructed through a graph neural network and associated with the breeding resource knowledge graph. When the target user enters a query text, the pre-trained domain BERT model is used to identify the query text and generate a user query vector. Input the user query vector into the breeding resource knowledge graph for cosine similarity matching, and select candidate resources with a similarity greater than a preset threshold and a preset resource quantity threshold to form an initial recommendation pool; Feature extraction is performed on each candidate resource in the initial recommendation pool to obtain a candidate resource feature vector. A second-order subgraph centered on the candidate resource node is then constructed using a graph neural network. A three-layer graph convolutional network is used to aggregate neighborhood information. After three iterations, a graph embedding vector is output. The candidate resource features and graph embedding vectors are input into the cross-attention fusion layer, and the dependency weights are obtained by linear projection alignment to the same dimension for weighted fusion. The weighted fused feature vectors are used as the input of the LambdaMART algorithm to sort the candidate resources, output the recommended resource sorting list, and make intelligent recommendations for feeding resources to the target users.

6. The animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method according to claim 1, characterized in that: The said building of a personalized recommendation engine, answering user questions and recommending required knowledge through the said personalized recommendation engine, and recommending resources of interest based on user usage behavior, further includes; A user interest analysis model is constructed based on a convolutional neural network. When a target user queries resources or asks for help, the target user's historical behavior data and real-time query information are obtained and input into the user interest analysis model for interest modeling. When the historical behavior data and real-time query information of the target user are input, differential embedding processing is performed by inputting the historical behavior data and real-time query information of the target user into the embedding layer, discrete features are mapped into multi-dimensional dense vectors through the embedding table, and continuous features are compressed through quantile normalization processing to obtain embedded feature vectors; The embedded feature vector is input into the hierarchical attention network. The bottom-level temporal attention module calculates the time decay weight of the behavior sequence, strengthens the semantic contribution of recent behavior, and uses the upper-level topic attention module to capture the user's focus of interest. Multi-scale convolution is used to parallelly process behavior sequences of different time granularities, extract local interest fluctuation patterns and long-term preference evolution trends, and finally output the target user's interest analysis information. If there is no active knowledge or question query of the target user, the target user interest analysis information is imported into the breeding resource knowledge graph to match interest resources, generate an interest resource recommendation table and push interest resources; If there is active knowledge or question query of the target user, an interest vector space is generated based on the target user interest analysis information, a ranked list of recommended resources is obtained, and each resource in the ranked list of recommended resources is mapped to the vector space using the TransR algorithm; Calculate the multi-angle similarity between each resource item and the target user's interest vector, including cosine similarity, Jaccard index and KL divergence, to generate a comprehensive matching score. Based on the comprehensive matching score, use the MMR algorithm to re-rank the resource items in the recommended resource ranking list, and make intelligent recommendations for breeding resources based on the re-ranking results.

7. The animal nutrition and feed course resource sharing cloud platform and knowledge graph construction method according to claim 1, characterized in that: The animal nutrition and feed course resource sharing cloud platform is built based on the personalized recommendation engine, feeding knowledge metadata database and feeding resource knowledge graph using microservice architecture technology, specifically including: Build data collection management framework, knowledge graph query framework, and personalized recommendation framework based on personalized recommendation engine, breeding knowledge metadata database, and breeding resource knowledge graph; The data collection and management framework includes a data collection unit and a data management unit, wherein the data collection unit is implemented by deploying a distributed crawler cluster, and the data management unit is associated with a breeding knowledge metadata database; The knowledge graph query framework is built by the breeding resource knowledge graph, which is used to manage knowledge graph updates and maintenance, and realize knowledge graph query and interaction through visualization technology; The personalized recommendation framework is built by the personalized recommendation engine and is used to analyze the real-time query text and historical behavior of the target user to recommend knowledge resources and resources of interest to the user; Based on the microservice architecture technology, a data collection and management framework, knowledge graph query framework, and personalized recommendation framework are constructed to build an animal nutrition and feed course resource sharing cloud platform.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium includes an animal nutrition and feed course resource sharing cloud platform and a knowledge graph construction method program. When the animal nutrition and feed course resource sharing cloud platform and the knowledge graph construction method program is executed by a processor, the steps of the animal nutrition and feed course resource sharing cloud platform and the knowledge graph construction method as described in any one of claims 1 to 7 are implemented.

Images