Self-learning data analysis model recommendation method and system based on domain adaptation

Abstract

The invention relates to the technical field of artificial intelligence and data analysis, and particularly provides a self-learning data analysis model recommendation method and system based on domain adaptation, and the method comprises the steps: dividing an input text into atomic analysis units through employing a syntactic tree analysis and paragraph topic clustering algorithm, pre-labeling entities in the atomic analysis unit based on a domain dictionary and a rule engine; according to the atomic analysis unit, outputting an analysis model recommendation result by adopting a semantic word segmentation optimization method, a multi-granularity retrieval matching method and a double-weight calculation sorting method; representing an association relationship between entities in the analysis model recommendation result based on an entity association graph; and performing structure updating on the entity association graph based on the user interaction behavior. According to the method, the text analysis accuracy, the semantic understanding depth, the recommendation precision and the dynamic updating capability of the knowledge base are remarkably improved, an efficient, accurate and self-adaptive solution is provided for data analysis in the vertical fields of medical treatment, finance, education and the like, and the method has important value and wide application prospects.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and data analysis, in particular to a self-learning data analysis model recommendation method and system based on domain adaptation. BACKGROUND

[0002] In today's digital era, vertical fields such as medicine and finance face core bottlenecks in deep data analysis. Although general large-scale language models can achieve basic functions in professional fields, they have three key limitations: first, the lack of domain-specificity, general models are difficult to accurately capture the disease-gene correlation in the medical field, the macro-micro indicator mapping in the financial field, and other deep professional knowledge contexts; second, the model has an illusion problem, which makes it difficult to guarantee the stability and reliability of the output results; third, the computing energy consumption is too high, and the real-time analysis of models with more than 70B parameters will cause significant response delay and excessive energy consumption, while smaller parameter models have the problem of poor output quality, which is difficult to meet the high real-time and high reliability requirements of business scenarios such as medical diagnosis and financial risk control, and is not conducive to low-configuration deployment in private environments. Finally, the dynamic adaptability is weak, and traditional methods lack online self-learning ability based on user feedback, which cannot continuously optimize with knowledge updates and demand changes.

[0003] Therefore, in fields with complex knowledge structures and significant statistical application values (such as medicine and finance), there is an urgent need for a lightweight, domain-adaptive, and self-learning data analysis model recommendation method that can significantly reduce computing overhead while ensuring reliability, and dynamically evolve the knowledge base and recommendation strategy, thereby efficiently supporting accurate decision-making and problem discovery. SUMMARY

[0004] In view of this, the present application proposes a self-learning data analysis model recommendation method and system based on domain adaptation to solve the above-mentioned problems of the prior art.

[0005] To achieve the above-mentioned purpose, the present application proposes a self-learning data analysis model recommendation method based on domain adaptation, comprising:

[0006] The input text is divided into atomic analysis units using syntax tree analysis and paragraph theme clustering algorithm, and the entities in the atomic analysis units are pre-labeled based on domain dictionary and rule engine;

[0007] According to the atomic analysis units, the semantic segmentation optimization, multi-granularity retrieval matching, and double-weight calculation sorting method are used to output the analysis model recommendation results;

[0008] The association relationships between entities in the analysis model recommendation results are represented based on entity association graphs;

[0009] The entity association graph is structurally updated based on user interaction behavior.

[0010] Furthermore, the semantic segmentation optimization includes employing an improved bidirectional long short-term memory network, combined with an attention mechanism, to identify the associations between entities across sentences;

[0011] The improved bidirectional long short-term memory network includes:

[0012] The traditional single-layer network structure is extended into a multi-layer structure, with the upper layer used to capture sentence-level local context information and the lower layer used to process document-level global context.

[0013] To set the L2 regularization parameter λ for the Bi-LSTM layer, a penalty term is added to the loss function. The complexity of the control model is given by λ∈[0.001,0.01]; the specific formula is as follows: ,

[0014] Add a Dropout layer between the Bi-LSTM layer and the attention layer, randomly dropping neurons at a ratio δ∈[0.2,0.5], and adjust according to practice;

[0015] During model training, a gradient clipping threshold θ∈[1,5] is set to prevent gradient explosion;

[0016] Forward and backward semantic features of the word sequence are obtained by forward LSTM and backward LSTM respectively;

[0017] A bidirectional attention mechanism is introduced to process the semantic features to identify cross-sentence entity associations, including:

[0018] Through formula Calculate the relevance score between the query vector and the key vector;

[0019] The attention weights are obtained by normalization using the softmax function: a ij =softmax(e ij );

[0020] Generate a context vector based on the weighted sum of the attention weights: ;

[0021] The above method is used to obtain a context vector with global semantic information.

[0022] Furthermore, the multi-granularity retrieval matching includes:

[0023] For a single entity word, it is directly mapped to a method template in the analysis knowledge base;

[0024] For compound words, the compound words are encoded into dense vectors, and schemes with similarity ≥ 0.75 are matched based on the ANN algorithm.

[0025] Furthermore, the dual-weighted ranking calculation includes:

[0026] Merge the multi-granularity search matching results and obtain the corresponding potential recommendation results;

[0027] The weight of each potential recommendation result is calculated based on the user's personalized habits, and the recommendation result of the analysis model is obtained based on the weight score.

[0028] Furthermore, the entity association graph includes a data layer, a pattern layer, and a domain adaptation layer;

[0029] The data layer stores entity instances and the relationships between entities;

[0030] The schema layer is used to define entity categories and relationship types, and to construct a conceptual hierarchy.

[0031] The domain adaptation layer is used to construct specific entity association graphs for different application domains.

[0032] Furthermore, the construction process of the entity association graph includes:

[0033] Entity recognition in text based on an improved Bi-LSTM model;

[0034] Identify relationships between entities using attention mechanisms and collaborative filtering algorithms;

[0035] Entities and relationships are categorized into predefined categories, and the hierarchical relationships between entity categories are defined.

[0036] By integrating domain-specific dictionaries into the entity recognition and classification process, feature extraction is performed based on domain characteristics, and domain-specific rules are applied.

[0037] Furthermore, the process of updating the structure of the entity association graph based on user interaction behavior includes:

[0038] Real-time collection of user interaction behavior, and encoding the user interaction behavior into time-series features;

[0039] Graph Neural Networks (GNNs) are used to update node features:

[0040]

[0041] Where A represents the adjacency matrix, which represents the connection relationships between nodes in the graph, and H... (l) H represents the node representation matrix of the l-th layer. (l-1) The node represents the matrix of the next higher level. W(l) Let represent the learnable weight matrix of the l-th layer. σ is a nonlinear activation function used to introduce a nonlinear transformation;

[0042] Based on the aforementioned temporal features, a graph neural network is used to update the weight matrix of the identified and categorized entities and their relationships:

[0043]

[0044] in, Represents the nodes in the adjacency matrix The final connection weight between node j and node j is a probability value used to measure the probability of connection between the two nodes in a given entity. Under the condition of , the strength of its association with entity j. This represents the updated base edge weights. Representing entities The semantic similarity between them is calculated using cosine similarity, with a value range of [-1, 1]. Representing entities respectively The feature vector of entity j comes from the output of the previous layer of the graph neural network. τ represents the temperature parameter, a hyperparameter greater than 0, used to adjust the smoothness of the output distribution of the Softmax function.

[0045] Based on the weight matrix, important entity associations are obtained. Based on the important entity associations, new entities and association relationships are added to the entity association graph, and the adjacency matrix of related entities is updated.

[0046] Furthermore, the method for updating the weight matrix is ​​as follows:

[0047]

[0048]

[0049] in, This represents the updated base edge weights (entity co-occurrence frequency weights). This represents the original edge weights (such as weights calculated based on co-occurrence frequency or semantic similarity). This represents the number of times entities i and j co-occur in user feedback. The co-occurrence count is calculated over a 6-hour time window, filtering out invalid interactions with a dwell time of less than 15 seconds. β∈[0.01,0.1] is the adaptive learning rate. Dynamic adjustment is made based on historical feedback errors, among which Let η be the prediction error loss function, η = 0.001 be the second-order learning rate, and ▽ represent the gradient calculation; α i This represents the updated node weights, where σ is the activation function ( h is used to compress node weights to the [0,1] interval; i W represents the feature vector of entity i. a Let b represent the learnable parameter matrix and bias term, and σ be the activation function.

[0050] Furthermore, the entity association graph is updated when one of the following conditions is met:

[0051] The number of newly added entities or relationships has reached a preset threshold;

[0052] User feedback indicated that the co-occurrence frequency exceeded the preset ratio.

[0053] On the other hand, to achieve the above objectives, this invention proposes a recommendation system based on a domain-adaptive self-learning data analysis model, including a text analysis module, a recommendation module, an analysis knowledge base, and a self-learning module;

[0054] The text analysis module is used to divide the input text into atomic analysis units using syntactic tree analysis and paragraph topic clustering algorithms, and to pre-annotate the entities in the atomic analysis units based on the domain dictionary and rule engine;

[0055] The recommendation module is used to output the analysis model recommendation results based on the atomic analysis unit, using semantic word segmentation optimization, multi-granularity retrieval matching, and dual-weight calculation and ranking methods.

[0056] The analytical knowledge base is used to represent the relationships between entities in the recommendation results of the analytical model based on an entity association graph.

[0057] The self-learning module is used to update the structure of the entity association graph based on user interaction behavior.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention employs syntactic tree analysis and paragraph topic clustering algorithms for text parsing, accurately decomposing complex text into atomic analysis units. It also utilizes a domain dictionary and rule engine for entity pre-annotation, effectively improving the accuracy and efficiency of text parsing and laying a solid foundation for subsequent analysis. This avoids domain terminology recognition errors caused by general word segmentation in traditional methods. Furthermore, the invention's knowledge base uses a knowledge graph-based template approach for entity storage, replacing large-model recommendation methods, thus saving computational resources while improving recommendation reliability.

[0060] In the semantic word segmentation optimization stage, this invention employs an improved bidirectional long short-term memory network (Bi-LSTM) combined with an attention mechanism, successfully overcoming the challenge of cross-sentence entity association recognition. Through a multi-layered network structure design, it simultaneously captures sentence-level local contextual information and document-level global context, enabling the model to understand more complex semantic relationships within the text and providing strong support for accurate recommendation analysis models.

[0061] The introduction of multi-granularity retrieval matching and dual-weighted ranking methods in this invention further enhances recommendation performance, enabling the system to handle not only single entity words but also combined words effectively. By using the ANN algorithm to match highly similar schemes, the retrieval scope is expanded and retrieval accuracy is improved. Dual-weighted calculation comprehensively considers users' personalized habits and group preferences, making the recommendation results more aligned with users' actual needs and improving the accuracy and reliability of recommendations. The construction and dynamic update mechanism of the entity association graph is another major highlight of this invention. The multi-level architecture of the entity association graph can comprehensively and accurately represent the complex relationships between entities, providing rich semantic information for data analysis. The graph structure update based on user interaction behavior allows the system to learn and adapt to changes in user needs in real time, continuously optimizing recommendation results and realizing the dynamic evolution of the knowledge base, effectively solving the problem of lagging updates in traditional knowledge bases. Attached Figure Description

[0062] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings:

[0063] Figure 1 This is a schematic diagram illustrating the architecture and working principle of a domain-adaptive self-learning data analysis model recommendation system in an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the system connection relationship in an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram of name association analysis in an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of medical terminology association analysis in an embodiment of the present invention. Detailed Implementation

[0067] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0068] This embodiment proposes a recommendation method and system based on a domain-adaptive self-learning data analysis model, such as... Figure 1 As shown, this embodiment uses a text analysis module to divide the input text into atomic analysis units using syntactic tree analysis and paragraph topic clustering algorithms, and pre-labels the entities in the atomic analysis units based on a domain dictionary and a rule engine.

[0069] The recommendation module outputs the recommendation results of the analysis model based on the atomic analysis units, using semantic word segmentation optimization, multi-granularity retrieval matching, and dual-weight calculation and ranking methods.

[0070] An analytical knowledge base is used to represent the relationships between entities in the recommendation results of the analytical model based on an entity association graph.

[0071] The self-learning module is used to update the structure of the entity association graph based on user interaction behavior. The functions of each module are as follows:

[0072] 1. Text parsing module:

[0073] The core functions of this module include using syntactic tree analysis and paragraph topic clustering algorithms (such as TextTiling) to segment the input text into atomic analysis units according to semantic units, or using natural language processing methods such as large models to classify the output text until it meets expectations; and using domain dictionaries (such as the ICH International Medical Terminology Dictionary and commonly used clinical medical terms) and rule engines to pre-label entities in atomic units (such as labeling "lung cancer" as a disease entity).

[0074] 2. Recommendation Module:

[0075] It is used to receive data from the text parsing module and output recommendation results using semantic word segmentation optimization, multi-granularity retrieval matching, dual-weight calculation and ranking methods;

[0076] (1) Semantic word segmentation optimization:

[0077] The semantic segmentation optimization uses an improved bidirectional long short-term memory network (Bi-LSTM) combined with an attention mechanism to identify cross-sentence entity associations (such as the correspondence between "a certain treatment" and the preceding "disease").

[0078] Improved bidirectional long short-term memory network structure

[0079] The improved Bi-LSTM network in this embodiment adopts a multi-layered architecture, including an input layer, a bidirectional LSTM layer, an attention layer, and an output layer. Compared to the traditional Bi-LSTM, this improved structure is optimized in the following aspects:

[0080] Network structure adjustment:

[0081] The improved Bi-LSTM employs a hierarchical structure, extending the single-layer Bi-LSTM into a multi-layer structure. The first layer of Bi-LSTM is responsible for capturing sentence-level local contextual information, while the second layer processes document-level global context. This design allows the model to simultaneously focus on short-distance dependencies within sentences and long-distance dependencies across sentences.

[0082] In its implementation, the improved Bi-LSTM employs an adaptive gating mechanism. By adjusting the initial values ​​of the gating parameters and optimizing the gradient propagation path, it addresses the gradient vanishing problem that traditional Bi-LSTM often encounters when handling long-range dependencies. Simultaneously, the network introduces residual connections, establishing direct connections between different layers, further enhancing its ability to capture long-range dependencies.

[0083] Regularization mechanism implementation:

[0084] To address the overfitting problem, this embodiment integrates multiple regularization techniques into Bi-LSTM:

[0085] L2 regularization: Adding an L2 penalty term to the weight matrix to control model complexity, specifically implemented through the loss function.

[0086]

[0087] Where W represents the weight matrix of the Bi-LSTM layer, λ represents the L2 norm square of the weights (the sum of squares of all weight components), and λ is the weight decay coefficient, ranging from 0.001 to 0.01.

[0088] Specifically, this is achieved by setting the L2 regularization parameter when defining the Bi-LSTM layer, such as setting the weight decay coefficient of the LSTM layer to between 0.001 and 0.01.

[0089] Dropout regularization: A Dropout layer is added between the Bi-LSTM layer and the attention layer to randomly drop some neurons, enhancing the model's generalization ability. The proportion of neurons randomly dropped is δ∈[0.2,0.5], and is adjusted according to practice.

[0090] Gradient clipping: Set a gradient clipping threshold (usually 1-5) during training to prevent gradient explosion and improve training stability.

[0091] The gradient clipping uses the ReLU activation function:

[0092] The synergistic effect of regularization techniques effectively improves the model's generalization ability under small sample conditions, making it particularly suitable for scenarios with scarce data in vertical fields such as healthcare and finance.

[0093] Attention mechanism implementation:

[0094] This embodiment introduces a bidirectional attention mechanism based on Bi-LSTM. By calculating the correlation between different time steps and dynamically allocating weights, it enhances the ability to capture the association of key entities. The implementation process of the attention mechanism is as follows:

[0095] Bidirectional feature encoding: Forward and backward semantic features of the word sequence are obtained by forward LSTM and backward LSTM respectively, and then concatenated to obtain a bidirectional feature vector.

[0096] Attention weight calculation: Calculate the self-attention weights for the bidirectional feature vectors using the following formula:

[0097] e ij =v T ˙tanh(W q h i +W k h j +b)

[0098] Where eij represents vector h i and h j The correlation score W between them q W k Represents the learnable weight matrix, used to adjust h respectively. i and h j A linear transformation is performed, where hi and hj represent the input vectors (such as word vectors or hidden states). b is the bias term. tanh represents the hyperbolic tangent activation function, used for nonlinear transformations. T This represents a learnable weight vector used to map the output of tanh to a scalar score.

[0099] α ij =softmax(eij )

[0100] Where, α ij The attention weight is a probability value (∑jα). ij =1), used to measure the degree of attention the i-th element pays to the j-th element, and softmax represents the normalization function, which transforms any real vector into a probability distribution:

[0101]

[0102] Where h i and h j These are the bidirectional feature vectors of the i-th and j-th words, respectively, W. q and W k is the weight matrix for query and key, v is the attention vector, and b is the bias term.

[0103] Context vector generation: Based on the calculated attention weights, the feature vectors at all time steps are weighted and summed to generate a context vector with global semantic information.

[0104]

[0105] Among them, c i The context vector representing the i-th position is the model's "comprehensive understanding" of the current task; α ij The attention weights (softmax normalized values) represent the degree of attention the i-th target gives to the j-th input, satisfying the following condition: h j Let represent the feature vector (such as word vector, hidden state, etc.) of the j-th input. ∑j represents the weighted summation over all input positions j.

[0106] Understandably, the core functions of improving Bi-LSTM include:

[0107] Dynamic focusing: via α ij Weights, the model can dynamically determine which inputs (h) j More importantly, in machine translation, when generating target words, the focus is on related words in the source sentence, and in text summarization, the most critical sentences in the original text are selected and combined.

[0108] Information fusion: integrating disparate input information {h} j Weighted fusion based on importance generates a more robust representation c. i .

[0109] The bidirectional attention mechanism can effectively capture entity associations across sentences, such as the correspondence between "a certain treatment" and the preceding "disease".

[0110] Dynamic combination word segmentation: Generate 2-5 word combinations (such as "lung cancer EGFR mutation") through a sliding window, and calculate the TF-IDF weight of the combination to screen high-frequency effective combinations.

[0111] (2) Multi-granularity retrieval matching:

[0112] Specifically, multi-granularity retrieval matching includes:

[0113] Word-level retrieval: A method template that directly maps a single entity (such as "lung cancer") to an analytical knowledge base (such as "disease association analysis").

[0114] Combinatorial retrieval: Encode combination words (such as "lung cancer EGFR mutation") into dense vectors (such as Sentence-BERT), and match schemes with similarity ≥ 0.75 using ANN algorithms (such as Faiss).

[0115] It is understandable that this embodiment saves computational resources and improves the reliability of recommendations by using templates to save entity associations instead of recommending them through a large model.

[0116] (3) Double-weighted calculation and sorting:

[0117] The search results from the two methods described above are combined to obtain all potential recommendation results from the analysis knowledge base. The weight of each recommendation analysis scheme is calculated based on the user's personalized habits.

[0118] Recommendation weight score = λ1 × user personalized preference score + λ2 × group preference score, where λ1 + λ2 = 1

[0119] A user's own preferences are reflected in the graph of the knowledge base analysis in the form of edge weights. These edge weights are the parameters of the user preference scores for the corresponding recommendation results. The group preference scores are the common weight parameters of the graph.

[0120] Based on the recommendation weight score, the top n recommendations are retained in each recommendation category. n is adjusted according to system settings.

[0121] (4) Output recommendation results:

[0122] Send structured candidate solutions to the rendering module, which should include at least the following metadata fields: ID, name, keywords, method, API name called, API parameters, and recommendation weight vector parameters.

[0123] 3. Analyze the knowledge base:

[0124] Domain Adapter: Preset templates for medical, financial, legal, etc., and supports user-defined entity categories (such as adding "teaching methods" to the education domain).

[0125] (1) Entity Relationship Graph:

[0126] This embodiment constructs an entity association graph using a multi-layered architecture, including a data layer, a schema layer, and a domain adaptation layer, which can effectively represent the complex relationships between entities.

[0127] The edge weights and calculation methods of the entity association graph are as follows:

[0128] Co-occurrence frequency weight:

[0129]

[0130] Where, N ij This represents the number of times entities i and j appear together in the same context. N total This indicates the total number of text units (such as the total number of documents or sentences).

[0131] Semantic similarity weight:

[0132]

[0133] h i ,h j The semantic vectors representing entities i and j (generated using Word2Vec, BERT, etc.).

[0134] Domain knowledge weight:

[0135]

[0136] Among them, s ij This represents the association strength (such as frequency or human rating) between entities i and j in the knowledge base. k represents the slope coefficient of the Sigmoid function (usually taken as 1~5).

[0137] Overall edge weights:

[0138]

[0139] Where λ1, λ2, and λ3 are weighting coefficients, and λ1 + λ2 + λ3 = 1.

[0140] Minimize the loss function using the gradient descent algorithm. Training λ k :

[0141]

[0142]

[0143] in, The weight training parameters, η=0.01 is the learning rate. To train the edge set.

[0144] Understandably, co-occurrence frequency weights can quickly capture high-frequency co-occurring entity pairs, solving the problem of superficial associations, and are suitable for preliminary relationship mining (such as "person-place" associations in news events). Semantic similarity weights can identify entities that are not co-occurring but semantically related (such as "smoking" and "lung cancer"), and are suitable for vertical domain knowledge graphs. Domain knowledge weights utilize prior knowledge to strengthen professional domain relationships (such as "drug-side effects"), effectively solving cold start and professional issues, and are suitable for scenarios such as healthcare and finance. Finally, comprehensive weights balance statistics, semantics, and knowledge to adapt to different scenario needs.

[0145] Multi-layer graph structure definition:

[0146] Data Layer: Stores concrete entity instances and direct relationships. Nodes represent specific entities (such as "EGFR gene," "lung cancer," etc.), and edges represent direct relationships between entities. This layer is constructed through the following steps:

[0147] Entity recognition: Identifying entities in text using an improved Bi-LSTM model;

[0148] Relationship extraction: Identifying relationships between entities through attention mechanisms and collaborative filtering algorithms;

[0149] Schema layer: Defines entity categories and relationship types, constructing the conceptual hierarchy. This layer is built through the following steps:

[0150] Entity classification: Classifying entities into predefined categories (such as "disease", "gene", "treatment" etc.);

[0151] Relationship classification: Categorize relationships into predefined types (such as "related diseases", "treatment plans", etc.);

[0152] Hierarchical structure: Defines the hierarchical relationship between entity categories (e.g., "cancer" → "lung cancer" → "EGFR-mutant lung cancer", etc.);

[0153] Domain Adaptation Layer: This layer constructs a domain-specific entity relationship graph tailored to the characteristics of different domains (such as healthcare and finance). It is built through the following steps:

[0154] Domain-specific dictionary integration: Integrating domain-specific dictionaries (such as the ICH dictionary in the medical field) into the entity recognition and classification process;

[0155] Domain feature extraction: Extracting specific features based on the characteristics of a domain (such as gene mutation features in the medical field, macroeconomic indicator features in the financial field, etc.);

[0156] Domain rule application: Apply domain-specific rules (such as synonym entity merging, domain knowledge constraints, etc.).

[0157] (2) Construction and storage of classification templates

[0158] Template structure design:

[0159] In the analysis knowledge base, predefine common entity categories (such as "disease", "treatment method", "gene", etc.) and bind association analysis methods to each category combination (such as "disease + treatment method" → "chi-square test").

[0160] Example template format: Json{"Entity Category": ["Disease", "Treatment Method"], "Recommended Method": ["Chi-square Test", "Association Analysis", "Survival Curve Analysis"]};

[0161] Template dynamic expansion: Allows for the automatic addition of entity categories (such as "Teaching Methods" under "Education") and association with specific analysis methods (such as "Teaching Methods + Student Performance" → "Regression Analysis"). Specifically, the self-learning module records data and triggers updates.

[0162] Rule filtering: By learning patterns through preset rules (such as "disease + gene" → "association analysis"), low-correlation combinations are filtered to avoid mismatches.

[0163] The above methods enable high-quality recommendation generalization capabilities. New entities are rapidly matched and analyzed using semantic similarity calculation and classification templates, without waiting for the accumulation of full datasets. Furthermore, the pre-set templates and dynamic expansion of multi-layer graph structures support rapid migration across various fields such as healthcare, finance, and education.

[0164] 4. Self-learning module:

[0165] By closely integrating the behavioral data acquisition module with the incremental update mechanism, a closed-loop learning system is formed.

[0166] Behavioral data collection:

[0167] Behavioral data collection → Incremental update mechanism → Graph structure update

[0168] Record user interaction behavior: click-through rate, dwell time, and secondary modification actions (such as manually adjusting recommendation parameters).

[0169] Generate behavior vectors: Encode user operation sequences into temporal features (e.g., dwell time > threshold is considered strong correlation).

[0170] Incremental update mechanism: New entities and relationships are received in real time from the self-learning module, and entity weights and classification boundaries are dynamically adjusted using a graph neural network (GNN), as shown in the following formula:

[0171]

[0172] in, Represents the nodes in the adjacency matrix The final connection weight between node j and node j is a probability value used to measure the probability of connection between the two nodes in a given entity. Under the condition of , the strength of its association with entity j. This represents the updated base edge weights (entity co-occurrence frequency weights). Representing entities The semantic similarity between them is calculated using cosine similarity, with a value range of [-1, 1]. Representing entities respectively The feature vector of entity j originates from the output H of the previous layer of the graph neural network. (l-1) τ represents the temperature parameter, a hyperparameter greater than 0, used to adjust the smoothness of the output distribution of the Softmax function.

[0173] The propagation formula for a graph neural network (GNN) layer is shown below:

[0174]

[0175] Where A represents the adjacency matrix, indicating the connection relationships (such as co-occurrence relationships) between nodes in the graph. H (l) The node representation matrix (node ​​features) represents the nodes in the l-th layer. H (l-1) The node represents the matrix of the next higher level. W (l) Let represent the learnable weight matrix of the l-th layer. σ is a non-linear activation function (such as ReLU or Sigmoid) used to introduce a non-linear transformation.

[0176] Understandably, graph neural networks aggregate information about neighboring nodes through the adjacency matrix A and the weight matrix W. (l) A linear transformation is applied to the aggregated features. An activation function σ is used to enable the model to learn complex patterns.

[0177] By solving the graph structure learning problem, the topological relationships between nodes can be automatically captured, making it suitable for scenarios such as social networks and knowledge graphs.

[0178] Edge weight adjustment based on user feedback:

[0179]

[0180]

[0181] in, This represents the updated base edge weights (entity co-occurrence frequency weights). This represents the original edge weights (such as weights calculated based on co-occurrence frequency or semantic similarity). This represents the number of times entities i and j co-occur in user feedback. The co-occurrence count is calculated over a 6-hour time window, filtering out invalid interactions with a dwell time of less than 15 seconds. β∈[0.01,0.1] is the adaptive learning rate. Dynamic adjustment is made based on historical feedback errors, among which Let η be the prediction error loss function, η = 0.001 be the second-order learning rate, and ▽ represent the gradient calculation; α i This represents the updated node weights, where σ is the activation function ( h is used to compress node weights to the [0,1] interval; i W represents the feature vector of entity i. a Let b represent the learnable parameter matrix and bias term, and σ be the activation function.

[0182] It is understandable that this embodiment updates edge weights in real time based on user behavior (such as clicks and searches), strengthens important associations, effectively solves the problems of cold start and data sparsity, supplements statistical deficiencies with user feedback, and improves the performance of recommendation systems and search ranking.

[0183] Attention-based node weight adjustment:

[0184] α i= σ(W) a h i +b)

[0185] Among them, h i The feature vector representing entity i (such as the node representation output by a GNN). W a and b represent the learnable parameter matrix and bias term. σ is the activation function (such as Sigmoid) that compresses the output to the range [0,1].

[0186] Graph structure updates include the following two parts: adding new entities and relationships to the graph, and updating the adjacency matrix of related entities.

[0187] An incremental update of the graph is triggered when one of the following conditions is met, updating the group preference score parameters:

[0188] The number of newly added entities or relationships has reached a preset threshold;

[0189] User feedback indicated that the co-occurrence frequency exceeded the preset ratio.

[0190] It is understandable that this embodiment calculates the attention weight α of node i. i This reflects its importance in the current task. The model can dynamically focus on key nodes (such as popular keywords and key figures), suppress irrelevant nodes, enhance key information, and solve noise filtering problems (better analysis methods are recommended).

[0191] In the self-learning module, the attention mechanism optimizes learnable parameters through the following training process:

[0192] (1) Initialization:

[0193] Wq and Wk are initialized using a Xavier normal distribution:

[0194]

[0195] Where n in and n out Let v and b represent the input and output dimensions, respectively. v and b are initialized to zero vectors. This initialization strategy helps to keep gradients stable in the early stages of training and prevent gradient explosion or vanishing.

[0196] (2) Forward propagation:

[0197]

[0198] α ij =softmax(e ij )

[0199]

[0200] The context vector ci, as the final output of the forward propagation, will be used as an enhanced feature representation and directly input to the downstream task-specific output layer to perform the final entity recognition or relationship classification prediction task. Its role is to aggregate information from the entire sequence, providing the most relevant context for the current decoding position.

[0201] (3) Back propagation:

[0202] Minimize the cross-entropy loss function using the gradient descent algorithm. The loss function is then optimized by calculating the gradient of the loss with respect to each learnable parameter using the chain rule. The loss function is defined as follows:

[0203]

[0204] Among them Authentic labels To predict the probability distribution, taking parameter Wq as an example, its gradient is calculated as follows:

[0205]

[0206] Chain rule differentiation process:

[0207]

[0208] Where sech2(·) is the derivative of the hyperbolic tangent function tanh(·). Similarly, the loss can be calculated. The gradients of v and b.

[0209] (4) Parameter update:

[0210] Based on the gradient calculated through backpropagation, the Adam optimizer is used to update all learnable parameters in the attention mechanism, including the matrix. , The hyperparameters of the Adam optimizer are configured differently according to the application domain. See Table 1 for specific learning rate settings. The values ​​in Table 1 are experimental reference values ​​and should be adjusted according to the actual application scenario.

[0211] Table 1

[0212] Domain Learning rate Number of convergent iterations Medical 0.001 12000 Finance 0.005 8000 Education 0.004 9,000 E-commerce 0.007 6,000

[0213] 5. Rendering module

[0214] This module presents the analysis data from the recommendation module in a structured format to assist users in decision-making. The module's input is a list of analysis methods and their metadata (such as applicable conditions, expected effects, and side effects) from the recommendation module. The output is an interactive menu-driven interface that displays basic information and relevance ranking of candidate analysis methods. It also supports filtering and sorting by keywords, relevance, and tags, improving usability and user experience, and enhancing the interpretability and operability of the recommended content.

[0215] The implementation process of the rendering module is as follows:

[0216] Receive recommendation data: Obtain a structured set of candidate analysis methods from the recommendation engine. Each method contains metadata fields: ID, name, keywords, method, API name called, API parameters, and recommendation weight vector parameters.

[0217] Generate visual menu items: Encapsulate each analysis method as an independent menu item component. Includes visual elements such as title, brief description, icon / label (e.g., "Highly Recommended", "New Solution"), and rating bar.

[0218] Embedded interactive controls: Add buttons for clicking to expand details, add to favorites, and compare multiple solutions. Support user feedback mechanisms (such as likes and marking as irrelevant) for subsequent model optimization.

[0219] Responding to user actions: When a user selects a method, a detailed information pop-up or redirection to the corresponding page is triggered. This can be integrated with other modules (such as knowledge graphs and document retrieval) to provide deeper support.

[0220] Example 2

[0221] This embodiment proposes the implementation process of a domain-adaptive self-learning data analysis model recommendation method in the medical field, including the following steps:

[0222] Step 1: Real-time data acquisition and preprocessing:

[0223] Input text parsing: The user submits the text "Zhang San studies the efficacy of EGFR mutation and targeted therapy in lung cancer". The system breaks it down into atomic units through syntactic tree analysis and calls the medical domain dictionary and rule engine to identify entities such as "Zhang San", "lung cancer", "EGFR mutation" and "targeted therapy".

[0224] Step 2: Combining semantic word segmentation and vector retrieval:

[0225] A sliding window generates 2-5 word combinations. A statistical model is used to filter high-frequency combinations, calculate combination weights, and retain valid candidates. Furthermore, the classification criteria are optimized; categories such as "name," "disease," "gene," and "treatment method" are abstracted into entity categories within the medical field to facilitate more accurate analysis.

[0226] Semantic features are extracted using deep learning models to generate dense vector representations.

[0227] As a preferred embodiment, when the similarity of word combinations is low, or when there is no corresponding relationship that results in the inability to associate entity relationships (when the recommendation weight score is lower than the threshold, the knowledge base recommendation is considered to have failed), the model is triggered to call the analysis method of entity information analysis in the knowledge base that is similar to the above-mentioned entity association type, and the model recommendation analysis method is adopted to adaptively recommend and optimize the scenario where there is an inability to associate entity relationships.

[0228] Rendering output example:

[0229] Zhang San is a person's name, so the menu analysis is as follows: Figure 3 As shown. EGFR mutation is a medical term; analysis recommendations are as follows. Figure 4 As shown.

[0230] Step 3: Self-learning feedback and optimization:

[0231] Behavioral data collection: Records user interaction behaviors such as dwell time on clicked modules, secondary modification operations, and whether additional keywords have been added; detects outliers through a sliding window. (Consider it as noise) and record it as vector data.

[0232] Keyword classification abstraction: "Lung cancer" and "EGFR mutation" are abstracted into entity classifications: "disease" and "gene". The recommended weight calculation parameters are used for the analysis of similar entities or combinations of entities.

[0233] The above content is combined to form the relationship of recommendation word analysis.

[0234] Dual weighting calculation: Personalized weighting: User preferences are mined based on collaborative filtering algorithms.

[0235] Recommendation priority adjustment: Added association with the analysis knowledge base, recommendation weight is dynamically increased based on user feedback.

[0236] Step 4: Real-time updates of the entity association graph:

[0237] Incremental update process: If the node description is too long, the model is called to generate a summary, historical attributes are retained and new information is merged; the similarity between the new edge and the existing edge is determined by the model, and if they are similar, they are merged into a composite relationship.

[0238] Node update strategy: Attribute matching: Query the first information table to verify the legality of the new entity; Association construction: Obtain the target object through the second information table and establish the association between the new entity and the existing node.

[0239] Side update strategy: Dynamically adjust classification boundaries.

[0240] Step 5: Quality Monitoring and Version Control

[0241] Consistency verification: Verify the logical consistency of the updated graph through subgraph isomorphism detection.

[0242] Version management: Each update generates a unique version identifier and supports rollback to any historical state.

[0243] Performance evaluation: Calculate recommendation accuracy and response latency, and automatically trigger performance tuning mechanisms.

[0244] Example 3

[0245] This embodiment proposes the implementation process of a domain-adaptive self-learning data analysis model recommendation method in the financial field, including the following steps:

[0246] Step 1: Real-time data acquisition and preprocessing:

[0247] Input text parsing: The user submits the text "A certain new energy vehicle company's stock is affected by the carbon neutrality policy". The system breaks it down into atomic units through syntactic tree analysis and calls the financial domain dictionary and rule engine to identify entities such as "new energy vehicle company stock" and "carbon neutrality policy".

[0248] Step 2: Combining semantic word segmentation and vector retrieval:

[0249] A sliding window generates 2-5 word combinations (such as "new energy vehicle company stocks carbon neutrality policy"), and a statistical model is used to screen high-frequency combinations, calculate the combination weights, and retain valid candidates.

[0250] Semantic features are extracted using deep learning models to generate dense vector representations.

[0251] Step 3: Real-time updates of the entity association graph:

[0252] Cluster analysis was used to identify a strong correlation between "carbon neutrality policy" and "new energy industry" and new classification labels were dynamically added.

[0253] The graph neural network is invoked to recalculate the node embeddings and adjust the edge weights between entities.

[0254] By linking the "carbon neutrality policy" with historical stock price data, we recommend a policy sensitivity analysis model.

[0255] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A recommendation method based on a domain-adaptive self-learning data analysis model, characterized in that, include: The input text is divided into atomic analysis units by using syntactic tree analysis and paragraph topic clustering algorithms, and the entities in the atomic analysis units are pre-labeled based on the domain dictionary and rule engine; Based on the atomic analysis unit, the analysis model outputs recommendation results using semantic word segmentation optimization, multi-granularity retrieval matching, and dual-weight calculation and ranking methods. The entity association graph represents the association relationships between entities in the recommendation results of the analysis model. The entity association graph is structurally updated based on user interaction behavior.

2. The method according to claim 1, characterized in that, The semantic word segmentation optimization includes using an improved bidirectional long short-term memory network, combined with an attention mechanism, to identify the associations between entities across sentences; The improved bidirectional long short-term memory network includes: The traditional single-layer network structure is extended into a multi-layer structure, with the upper layer used to capture sentence-level local context information and the lower layer used to process document-level global context. To set the L2 regularization parameter λ for the Bi-LSTM layer, a penalty term is added to the loss function. The complexity of the control model is given by λ∈[0.001,0.01]; the specific formula is as follows: , Add a Dropout layer between the Bi-LSTM layer and the attention layer, randomly dropping neurons at a ratio δ∈[0.2,0.5], and adjust according to practice; During model training, a gradient clipping threshold θ∈[1,5] is set to prevent gradient explosion; Forward and backward semantic features of the word sequence are obtained by forward LSTM and backward LSTM respectively; A bidirectional attention mechanism is introduced to process the semantic features to identify cross-sentence entity associations, including: Through formula Calculate the relevance score between the query vector and the key vector; The attention weights are obtained by normalization using the softmax function: a ij =softmax(e ij ); Generate a context vector based on the weighted sum of the attention weights: ; The above method is used to obtain a context vector with global semantic information.

3. The method according to claim 1, characterized in that, The multi-granularity retrieval matching includes: For a single entity word, it is directly mapped to a method template in the analysis knowledge base; For compound words, the compound words are encoded into dense vectors, and schemes with similarity ≥ 0.75 are matched based on the ANN algorithm.

4. The method according to claim 1, characterized in that, The dual-weighted sorting calculation includes: Merge the multi-granularity search matching results and obtain the corresponding potential recommendation results; The weight of each potential recommendation result is calculated based on the user's personalized habits, and the recommendation result of the analysis model is obtained based on the weight score.

5. The method according to claim 1, characterized in that, The entity association graph includes a data layer, a pattern layer, and a domain adaptation layer; The data layer stores entity instances and the relationships between entities; The schema layer is used to define entity categories and relationship types, and to construct a conceptual hierarchy. The domain adaptation layer is used to construct specific entity association graphs for different application domains.

6. The method according to claim 1, characterized in that, The process of constructing the entity association graph includes: Entity recognition in text based on an improved Bi-LSTM model; Identify relationships between entities using attention mechanisms and collaborative filtering algorithms; Entities and relationships are categorized into predefined categories, and the hierarchical relationships between entity categories are defined. By integrating domain-specific dictionaries into the entity recognition and classification process, feature extraction is performed based on domain characteristics, and domain-specific rules are applied.

7. The method according to claim 1, characterized in that, The process of updating the structure of the entity association graph based on user interaction behavior includes: Real-time collection of user interaction behavior, and encoding the user interaction behavior into time-series features; Graph Neural Networks (GNNs) are used to update node features: ， Where A represents the adjacency matrix, which represents the connection relationships between nodes in the graph, and H... (l) H represents the node representation matrix of the l-th layer. (l-1) The node representation of the upper layer is a matrix; W (l) Let represent the learnable weight matrix of the l-th layer. σ is a nonlinear activation function used to introduce a nonlinear transformation; Based on the aforementioned temporal features, a graph neural network is used to update the weight matrix of the identified and categorized entities and their relationships: ， in, Represents the nodes in the adjacency matrix The final connection weight between node j and node j is a probability value used to measure the probability of connection between the two nodes in a given entity. Under the condition that, its association strength with entity j; This represents the updated base edge weights; Representing entities The semantic similarity between them is calculated using cosine similarity, with a value range of [-1, 1]. Representing entities respectively The feature vector of entity j comes from the output of the previous layer of the graph neural network. τ represents the temperature parameter, which is a hyperparameter greater than 0 and is used to adjust the smoothness of the output distribution of the Softmax function. Based on the weight matrix, important entity associations are obtained. Based on the important entity associations, new entities and association relationships are added to the entity association graph, and the adjacency matrix of related entities is updated.

8. The method according to claim 7, characterized in that, The method for updating the weight matrix is ​​as follows: ， ， in, This represents the updated base edge weights (entity co-occurrence frequency weights). Represents the original edge weights (such as weights calculated based on co-occurrence frequency or semantic similarity); This represents the number of times entities i and j co-occur in user feedback. The number of co-occurrences is counted over a 6-hour time window, and invalid interactions with a dwell time of less than 15 seconds are filtered out; β∈[0.01,0.1] is the adaptive learning rate. Dynamic adjustment is made based on historical feedback errors, among which Let η be the prediction error loss function, η = 0.001 be the second-order learning rate, and ▽ represent the gradient calculation; α i This represents the updated node weights, where σ is the activation function ( h is used to compress node weights to the [0,1] interval; i W represents the feature vector of entity i. a Let b represent the learnable parameter matrix and bias term, and σ be the activation function.

9. The method according to claim 1, characterized in that, The entity association graph is updated when one of the following conditions is met: The number of newly added entities or relationships has reached a preset threshold; User feedback indicated that the co-occurrence frequency exceeded the preset ratio.

10. A self-learning data analysis model recommendation system based on the method of any one of claims 1-9, characterized in that, It includes a text analysis module, a recommendation module, an analysis knowledge base, and a self-learning module; The text analysis module is used to divide the input text into atomic analysis units using syntactic tree analysis and paragraph topic clustering algorithms, and to pre-annotate the entities in the atomic analysis units based on the domain dictionary and rule engine; The recommendation module is used to output the analysis model recommendation results based on the atomic analysis unit, using semantic word segmentation optimization, multi-granularity retrieval matching, and dual-weight calculation and ranking methods. The analytical knowledge base is used to represent the relationships between entities in the recommendation results of the analytical model based on an entity association graph. The self-learning module is used to update the structure of the entity association graph based on user interaction behavior; The dual-weighted ranking method includes: Calculate the recommendation weight score based on user personalized preference scores and group preference scores: Recommendation weight score = ω1 × user personalized preference score + ω2 × group preference score; Where ω1+ω2=1, the recommendation results of the analysis model are sorted according to the recommendation weight score, and the recommendation results are sorted according to the comprehensive weight.

Images