A context-aware semantic-aware data preparation pipeline recommendation method and system

Abstract

The present application relates to the technical field of data preparation pipeline recommendation, in particular to a context-oriented semantic perception data preparation pipeline recommendation method and system. The method comprises data preprocessing on the obtained automated pipeline construction dataset; for the preprocessed dataset, double-view feature fusion based on a self-attention mechanism is performed; for the preprocessed dataset, context semantic information extraction based on a large language model is performed; based on feature fusion and information extraction, a pipeline recommendation result is obtained through a deep reinforcement learning network; the pipeline recommendation result is evaluated and optimized; and the pipeline is output. Through double-view feature fusion based on a self-attention mechanism, the statistical properties of the dataset and the historical pipeline component sequence are dynamically integrated end to end.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data preparation pipeline recommendation, in particular to a context-aware semantic perception-oriented data preparation pipeline recommendation method and system. BACKGROUND

[0002] In the field of machine learning and data science, data preparation is an important link to transform raw data into a form that can be efficiently utilized by algorithms. Data preparation usually includes a series of operations such as data cleaning, feature engineering, missing value processing, data normalization, encoding, etc., and its purpose is to transform raw data into structured input that can be efficiently utilized by downstream algorithms. However, the construction of an ideal data preparation pipeline faces a series of extremely complex and highly coupled challenges in actual engineering and scientific scenarios.

[0003] First, modern data exhibits characteristics such as large scale, strong heterogeneity, and prominent time-varying. Whether it is structured table data, semi-structured logs, or unstructured text and images, it may be mixed with problems such as missing, redundancy, abnormality, nesting, or distribution drift. Second, the data preparation process itself is highly dependent on the specific needs of the task and model. Different machine learning tasks (such as classification, regression, clustering, etc.) and their corresponding model structures (such as tree models, neural networks, ensemble learning, etc.) have different requirements for feature construction, data conversion, and class balancing. The diversity of business objectives and the evolution of model architectures further exacerbate the dynamic and personalized challenges of data preparation solutions. At the same time, data preparation involves a variety of operations, complex sequences, and a huge parameter space, and there may be strong dependencies or implicit constraints between different processing steps. The theoretical combination space doubles with each additional component operation, leading to an exponential explosion of candidate pipelines. More seriously, existing data preparation automation and human-machine collaborative optimization methods still have obvious shortcomings in dealing with the above challenges. Most methods rely on shallow feature expression or rule engines, lack comprehensive modeling of complex dependencies between operations, and cannot fully exploit historical experience, contextual information, and the global influence of operation sequences. In recent years, data preparation automation research has gradually shifted from traditional pipeline construction methods that rely on human experience and static rules to more intelligent and adaptive dynamic decision-making paradigms, emphasizing the use of contextual information inherent in data and processing, especially the complex dependencies between data features and pipeline history, to achieve automatic optimization and efficient search of data preparation processes. However, there are still some key problems that have not been solved: First, current methods rely on shallow feature concatenation or static rules in the expression and utilization of contextual information, making it difficult to effectively model the complex global dependencies between pipeline history, data statistical features, and task requirements. Second, there is a lack of dynamic and intelligent space compression mechanism in the automatic component selection process, leading to pipeline structure search still limited by combinatorial space explosion and low decision-making efficiency. Third, existing methods have limited adaptability to heterogeneous and dynamic data scenarios, making it difficult to achieve end-to-end high-quality pipeline automatic construction. Most methods fail to capture high-order global dependencies and long-term historical influences between operations, lack dynamic feedback mechanisms, and result in pipeline optimization processes lacking closed-loop adaptation and long-term effectiveness guarantees.

[0004] Therefore, how to fully leverage the advantages of large models and their self-attention mechanisms in context understanding and complex dependency modeling to achieve deep integration of multi-source information and efficient decision-making has become a key technical problem in promoting the continuous evolution of data preparation automation. In response to these unsolved problems, this paper (or this patent) proposes a context-aware data preparation pipeline automatic construction method based on large models and self-attention mechanisms, aiming to further improve the intelligence, generalization, and automatic optimization level of data preparation in complex scenarios. SUMMARY

[0005] To solve the above-mentioned problems, the application provides a context semantic perception-oriented data preparation pipeline recommendation method and system.

[0006] In a first aspect, the application provides a context semantic perception-oriented data preparation pipeline recommendation method, which adopts the following technical scheme:

[0007] A context semantic perception-oriented data preparation pipeline recommendation method comprises:

[0008] Obtaining an automated pipeline construction dataset;

[0009] Data preprocessing is performed on the obtained automated pipeline construction dataset;

[0010] Based on the self-attention mechanism, double-view feature fusion is performed on the preprocessed dataset;

[0011] Based on the large language model, context semantic information extraction is performed on the preprocessed dataset;

[0012] Based on feature fusion and information extraction, a pipeline recommendation result is obtained through a deep reinforcement learning network;

[0013] The pipeline recommendation result is evaluated and optimized;

[0014] The pipeline is output.

[0015] Further, the data preprocessing of the obtained automated pipeline construction dataset comprises identifying data fields closely related to the target label in multiple heterogeneous data sources according to task definition, and after completing preliminary screening of features and labels, class distribution statistics are performed through explicit label column identification. In the extraction stage, the context information of tabular data is captured synchronously, and accurate labeling of data fields is realized through the correspondence mechanism of column name and semantics. Finally, a consistency and standardization mechanism is introduced, different source label encodings are uniformly mapped to a consistent class set by aligning the labels, the uniqueness of classification semantics is ensured, and uniform numerical operation is performed on the features. The normalization or bucketing strategy is used for numerical fields, and the discrete fields are converted into dictionary index or embedding vector to ensure the comparability of different features in the same space. For missing data, the completeness of the feature matrix is ensured through filling or rejection method. After the process, the final classification task representation is obtained:

[0016] ,

[0017] wherein and are standardized features and labels, ​This is contextual metadata that is bound to the data.

[0018] Furthermore, the self-attention-based dual-view feature fusion includes dataset statistical view feature extraction and pipeline history component feature extraction. For the dataset statistical view feature extraction, a detailed statistical feature extraction is first performed on the input dataset. For cases where the numerical feature distribution spans a large range, a combination of fractional and exponential parts is used to represent all numerical features. by Encoding is performed in the following manner, where The fractional part representing the feature. Indicates its exponential part, both are expressed through The original features are jointly represented; for feature extraction of pipeline history components, a sequence modeling neural network is introduced as the core encoding module, which extracts features from all executed data processing components in historical order. It is transformed into a dense vector sequence through the embedding layer. ,in For the number of components, For the embedding dimension, then the dense vector sequence The input is fed into a multi-layered stacked Transformer layer, which models the global dependencies between all components in the sequence in parallel, as follows:

[0019] ,

[0020] Then, a weighted pooling approach is used to extract global features from the sequence to represent the historical context information of the current pipeline.

[0021] Furthermore, the dual-view feature fusion based on a self-attention mechanism also includes introducing a feature fusion method based on a cross-view attention mechanism, which dynamically captures the interaction between the data distribution view and the pipeline structure view in an end-to-end trainable manner. Here, let the statistical features of the dataset be... Historical global characteristics of pipelines Using these as input sources for different modalities, and by using the statistical features of the dataset as query vectors, As the key, the attention weights of statistical features under different pipeline historical representations are automatically calculated and expressed as follows:

[0022] ,

[0023] in, , , As a trainable reference, the global representation of the pipeline components is obtained after aggregation through an attention mechanism. , represented as: .

[0024] Furthermore, the extraction of contextual semantic information based on a large language model includes introducing a contextual semantic information extraction mechanism based on a large language model. This involves automated deep mining of multi-source contextual information from the dataset. Specifically, by converting the original data table into CSV format text with headers and data samples, the large language model's ability to handle extremely long contexts and its token-level attention mechanism are utilized to capture key features such as field naming, value distribution patterns, and synonym categorization. This is represented as follows:

[0025] ,

[0026] in This represents the context vector generated by the large language model. This represents the embedding of inference functions into a large language model, followed by a context-based decision-making process that simulates the process by which experts narrow down the range of operational choices using domain knowledge. This achieves dynamic pruning and intelligent filtering of the candidate component space, and is represented as follows:

[0027] ,

[0028] Among them, when candidate components In the given context information conditional probability Greater than the preset threshold At that time, the component is selected into the final component set. By introducing contextual embedding features into each operational decision, the system automatically focuses on the component combinations most likely to improve data processing performance within the action space. Simultaneously, it proposes an optimized objective loss function, considering the impact of recommendation accuracy and contextual information to optimize the decision-making process in pipeline construction, expressed as:

[0029] ,

[0030] in, It is a recommended piping component. These are the components that are actually used. The regularization coefficient controls the context information. Its role in the optimization process.

[0031] Furthermore, the pipeline recommendation result obtained through a deep reinforcement learning network based on feature fusion and information extraction includes embedding all information from the pipeline construction process as a state. In the decision-making process, through in-depth The network predicts all possible actions. The value of the Bellman equation takes into account the expected future multi-step cumulative reward, realizes the dynamic balance of short-term and long-term benefits, and is expressed as:

[0032] ,

[0033] wherein, represents the expected value of long-term accumulation that can be obtained by taking action in a given state and following the policy of the current parameter , represents the transition to the next state after performing action , represents a discount factor, and further based on the policy selects the component that can maximize the reward, that is, , by inputting the current state s, outputs the priority or probability of each possible action, The calculation of the value follows the Bellman equation, which evaluates the expected value of future cumulative rewards when selecting component at the current step, and is expressed as:

[0034] ,

[0035] wherein represents the immediate reward value obtained when selecting component at the current step, is a discount factor for the reward, if takes a value close to 1, it means that the agent will pay high attention to the benefits brought by each step in the future, and will take long-term effects into account; if takes a value less than a set value, the agent will focus more on the immediate returns obtained by the current action, and the degree of attention to future rewards will be reduced.

[0036] Further, the pipeline recommendation result obtained by the deep reinforcement learning network based on feature fusion and information extraction further includes training the deep reinforcement learning network, first randomly extracting a data set from a large-scale training data set pool containing different fields and diversified data distribution, obtaining current data , constructing a state based on the current statistical characteristics of the data set and the existing pipeline historical information, selecting an action to automatically generate an adaptive data preparation pipeline, and evaluating the effect of the pipeline through the performance indicators of the downstream model, the evaluation result as a reward signal feedback to the agent for optimizing the decision-making policy, and is expressed as:

[0037] ,

[0038] Under the continuous training of cross-domain heterogeneous data, the optimal pipeline for the sample is obtained by minimizing the loss function, which is represented as:

[0039] ,

[0040] and updating the parameters with a learning rate , which ensures that the automated pipeline recommendation system can continuously output high-quality pipeline combinations across diverse data environments and in the inference phase, the optimal pipeline for new task data is represented as:

[0041] ,

[0042] where represents the pipeline configuration that maximizes the expected cumulative reward in the set of all candidate pipelines ; represents the mathematical expectation of a random process, which measures the average performance.

[0043] Further, the evaluation and optimization of the pipeline recommendation results include input inference for new task data and optional labels for supervised tasks ; with field type recognition and consistency checking mechanism to generate multi-dimensional labels such as numerical, categorical, text, time series and geographic information or identifiers, representative sample rows are selected through hierarchical sampling method, and value domain sketch is constructed for data features, while statistical views are extracted for the whole data . Finally, similar tasks are found in the task library based on embedding retrieval mechanism, if the similarity exceeds the set threshold, the corresponding pipeline prefix is extracted as the initial pipeline history , and and are fused to form the initial state vector for inference:

[0044] ,

[0045] After the generation of the initial state vector, candidate regulations are carried out under the constraints of the component library , and the conditional prior probability is calculated based on the context semantic representation . Then, the comprehensive score of each candidate component is calculated:

[0046] ,

[0047] the pre-trained strategy is called​ and its corresponding action value function Under the mild exploration mechanism, the optimal component sequence is gradually selected and the pipeline is constructed.

[0048] Further, the evaluation and optimization of the pipeline recommendation result further comprises updating the current environment state in real time according to the state transition mechanism of reinforcement learning after the candidate component is selected and executed In each transition, the newly generated intermediate result feature is embedded into the semantic space, fused with the previous context representation, and aggregated into a new context vector through weighting At the same time, the pipeline history trajectory is updated to include the sequence description of the current executed component, so that subsequent candidate selection can be performed under more complete context information:

[0049] ,

[0050] Through the above recursive state transition, the optimal solution of the target task is gradually approached; In the pipeline construction process, a global evaluation mechanism is introduced to dynamically score the candidate pipeline in multiple dimensions, when the iteration is completed and the final pipeline is generated, the pipeline is taken as the output result of inference, and based on the feedback mechanism, the inference output is compared with the verification result after actual execution, if the prediction deviation is found, the deviation information is returned to the policy network and the value function network, so that the model parameters are gradually corrected in future inference, represented as:

[0051] ,

[0052] Wherein represents the difference between the actual performance and the predicted performance, which is used to measure the gap between the prediction model and the true result, represents the model parameters, represents the learning rate used to control the step size of each parameter update, represents the performance prediction error.

[0053] Secondly, a context semantic perception oriented data preparation pipeline recommendation system comprises:

[0054] The data acquisition module is configured to acquire an automated pipeline construction dataset;

[0055] The preprocessing module is configured to perform data preprocessing on the acquired automated pipeline construction dataset;

[0056] The feature fusion module is configured to perform double-view feature fusion based on the self-attention mechanism for the preprocessed dataset;

[0057] The information extraction module is configured to perform large language model-based context semantic information extraction on the preprocessed data set.

[0058] The pipeline recommendation module is configured to obtain a pipeline recommendation result through a deep reinforcement learning network based on feature fusion and information extraction.

[0059] The optimization module is configured to evaluate and optimize the pipeline recommendation result.

[0060] The output module is configured to output the pipeline.

[0061] In a third aspect, the present application provides a computer-readable storage medium, wherein a plurality of instructions are stored, the instructions being suitable for being loaded by a processor of a terminal device and executing the context semantic perception-oriented data preparation pipeline recommendation method.

[0062] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer-readable storage medium, the processor being used to implement instructions; and the computer-readable storage medium being used to store a plurality of instructions, the instructions being suitable for being loaded by the processor and executing the context semantic perception-oriented data preparation pipeline recommendation method.

[0063] In summary, the present application has the following beneficial technical effects:

[0064] The present application dynamically integrates the statistical properties of the data set and the historical pipeline component sequence through the double-view feature fusion based on the self-attention mechanism. The self-attention mechanism can automatically capture the deep dependency relationship and nonlinear interaction between features without human intervention, effectively enhancing the discrimination of global state representation, and making the recommendation system maintain higher robustness and generalization ability when facing complex and variable data distribution.

[0065] Secondly, by introducing the context semantic information extraction based on the large language model, the present application can realize high-dimensional embedding expression of multi-source heterogeneous information such as data table header, field name and sample content. This module not only makes up for the deficiency of traditional statistical features in modeling the field context, but also significantly reduces the pipeline component search space, improves the pertinence, accuracy and reasoning efficiency of the recommendation process.

[0066] Thirdly, the deep reinforcement learning network is used for automatic search and optimal decision of the pipeline structure. The system can dynamically balance the immediate and long-term benefits by combining the multi-dimensional features of the current state, realize the optimal recommendation of the pipeline global structure, and effectively avoid the local optimum and empiricism trap.

[0067] A plurality of experiments and evaluations show that the present application has significant improvement in pipeline generation accuracy, decision delay, adaptability to new fields and generalization, and has wide engineering application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The pipeline construction framework of the present application is shown in the figure.

[0069] Figure 2 The automatic pipeline construction method flowchart proposed by the present application is shown in the figure.

[0070] Figure 3 The step 1 double-view feature fusion flowchart based on self-attention mechanism of the present application is shown in the figure.

[0071] Figure 4 The step 3 deep reinforcement learning network training flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0072] The present application will be further described in detail below with reference to the accompanying drawings.

[0073] Example 1

[0074] Reference Figure 1 The data preparation pipeline recommendation method for context-aware semantic perception of the present embodiment includes:

[0075] S1. Automatic pipeline construction data set extraction

[0076] In the construction process of the automatic data preparation pipeline, the acquisition and representation of the data set are the front link of the whole process, and their quality and expression ability directly affect the performance of the subsequent feature fusion, self-attention modeling and pipeline optimization. The present application proposes a systematic extraction and modeling method for the data set of the machine learning classification task, provides rich and stable context input for pipeline optimization through the joint mechanism of distribution feature modeling, semantic dependency relationship mining and global consistency representation. The overall idea is to closely combine the input characteristics of the classification data set with the task requirements, to ensure that the subsequent component recommendation and structure optimization can not only capture a single statistical pattern, but also perceive the potential class distribution law and context interaction effect.

[0077] S1.1 Data set statistical view feature extraction

[0078] In the classification task scenario, the system first defines the target task as:

[0079] ,

[0080] Among them, represents the sample feature matrix, represents the class label, represents the number of categories. According to the task definition, the extraction process needs to identify the data fields closely related to the target label in multiple heterogeneous data sources, such as structured records in relational databases, event attributes in log streams, or indicator parameters in experimental result sets. After completing the preliminary screening of features and labels, the system explicitly identifies the label column and performs statistics on its category distribution, specifically by calculating:

[0081] ,

[0082] to depict the frequency of each category. This can reveal potential class imbalance and long-tail distribution characteristics in the data, such as the phenomenon that some categories have very low sample proportions. This distribution awareness process not only constitutes the core contextual information of the classification dataset, but also serves as an important reference for model selection, resampling strategies, and regularization methods in the subsequent automated pipeline.

[0083] S1.2 Context information capture and semantic annotation

[0084] In the construction process of the classification dataset, relying solely on the feature matrix and label vector is often insufficient to fully express the actual context of the data. Therefore, the present invention proposes to capture the contextual information of tabular data simultaneously in the extraction phase, and achieve accurate annotation of data fields through the correspondence mechanism of column name and semantics. Let the field set contained in the original data table be:

[0085] ,

[0086] where each field corresponds to a column of the data table. The system generates semantic description for each field in the extraction process, and then forms a mapping relationship:

[0087] ,

[0088] This column name-semantic correspondence not only depicts the structure of the data, but also preserves the external environment and business logic. For example, in a medical classification task, the column name "blood_pressure" is annotated as "clinical detection indicator", and the column name "collection_date" is annotated as "experimental timestamp"; in a log stream classification scenario, the column name "event_id" is annotated as "event identifier", and the column name "timestamp" is annotated as "time series anchor point". Through this mechanism, the system not only obtains numerical-level features and labels after extraction, but also understands the semantic role of each column in the task. This structured representation enables the automated pipeline to fully utilize the business meaning and task constraints of the fields in the subsequent model and component recommendation process, thereby significantly improving the adaptability and discriminability of the classification task.

[0089] S1.3 Data consistency and standardization mechanism

[0090] Since data often comes from heterogeneous environments, there are obvious differences in format, encoding method and label system between different sources. If these data are directly used, it will lead to uncontrollability of the pipeline optimization process. Therefore, in the extraction link, the invention introduces a consistency and standardization mechanism to realize the unified representation of cross-source data. The mechanism first aligns the labels, and uniformly maps the label encoding of different sources to a consistent category set , ensuring the uniqueness of the classification semantics. Secondly, the features are uniformly numerically operated, the numerical fields use normalization or bucketing strategy, and the discrete fields are converted into dictionary index or embedding vector to ensure the comparability of different features in the same space. For missing data, the system ensures the integrity of the feature matrix through filling or rejection method. After this process, the final classification task representation is obtained:

[0091] ,

[0092] where and are the standardized features and labels respectively, is the context meta information bound with the data. Through the above mechanism, the classification data has a unified format and stable semantics before entering the automatic pipeline optimization link, so that it can seamlessly interface with the subsequent feature modeling and component search.

[0093] S2. Dual-view feature fusion based on self-attention mechanism

[0094] S2.1 Statistical view feature extraction of dataset

[0095] In the automatic construction process of the data preparation pipeline, the invention first performs detailed statistical feature extraction on the input dataset. Specifically, for each intermediate dataset , its statistical features are composed of column-wise field feature vectors, denoted as:

[0096]

[0097] where, represents the statistical feature of the th column, is the maximum column number. When the actual field number is less than , a truncation or zero padding operation is performed to ensure the consistency of the vector length of the statistical features under all tasks, so as to facilitate the construction of the state vector representation in the subsequent reinforcement learning network.

[0098] Each column feature A set of statistics that can effectively reflect the data properties of the field are included, which not only includes basic properties such as column type, missing value ratio, and unique value number, but also can include higher-order numerical statistics such as mean, standard deviation, and median to help the model distinguish and judge the applicability and priority of different components. These column features provide rich data information support for pipeline component selection.

[0099] In view of the fact that numerical features may have a large distribution span, extreme values, or uneven distribution, the present application further proposes an innovative expression method: using a combination of fractional part and exponential part, i.e. scientific notation. Specifically, all numerical features are encoded in the form of , where represents the fractional part of the feature, represents the exponential part, and both represent the original feature through . This design fully retains the order of magnitude information of the feature, effectively avoids the loss of resolution that may be caused by simple normalization, and ensures that the model still has excellent discriminability and expression stability when processing large-scale span data. At the same time, this expression method helps to strengthen the learning model to adaptively focus on key information of different orders of magnitude, improving the sensitivity of the model to extreme values and change amplitudes.

[0100] In summary, systematic statistical feature extraction and scientific notation style numerical feature encoding not only lay a solid data foundation for subsequent self-attention mechanism deep feature fusion, but also significantly improve the generalization ability and practical application effect of pipeline automation recommendation.

[0101] S2.2 Pipeline history component feature extraction

[0102] In the automatic data preparation pipeline modeling framework of the present application, pipeline history component feature extraction focuses on the accurate expression of the context semantics, component interaction relationship and sequence characteristics of the executed operation sequence. Specifically, for each data set to be optimized, the system continuously tracks the order and type information of all selected operation components in the current pipeline when selecting the next component. This order not only includes basic data processing steps, but can also cover feature engineering, model evaluation, etc., and its arrangement directly affects the final data processing effect.

[0103] To fully extract the rich temporal information, dependency relationships and potential impacts on subsequent decisions implied in the execution order of data preparation pipeline components, the invention innovatively introduces a sequence modeling neural network as the core encoding module. Specifically, the system first converts all executed data processing components (including but not limited to missing value filling, standardization, feature selection, feature generation, dimensionality reduction, etc.) in accordance with their historical order into a dense vector sequence , where is the number of components, is the embedding dimension. In this process, each type of component has an independent and trainable embedding representation in the vector space, allowing the model to perceive and distinguish the semantic differences between different operation types.

[0104] The dense vector sequence is then input into a multi-layer stacked Transformer module. Each layer of the Transformer consists of multi-head attention mechanisms and feedforward networks, which can model the global dependencies between all components in the sequence in parallel, effectively capturing the interactions and long-term effects between different historical steps.

[0105] ,

[0106] After the Transformer encoding is completed, a weighted pooling method is used to extract global features from the sequence to represent the historical context information of the current pipeline. The weighted pooling method assigns a learnable weight to the hidden state of each position in the sequence, and sums all the encoded vectors weighted by the weights, thereby obtaining a global feature vector that comprehensively reflects the influence of all historical components. The formula is as follows:

[0107] ,

[0108] ,

[0109] where is the hidden state vector of each component position in the pipeline history at time , and w is a learnable parameter variable. is the weight of each position, is the global pipeline history component feature.

[0110] S2.3 Dual-view feature fusion

[0111] To fully exploit and integrate the complementary potential of data distribution features and pipeline history structure information in automated pipeline optimization, the invention innovatively introduces a feature fusion method based on cross-view attention. Unlike traditional simple concatenation or static weighted fusion methods, cross-view attention can dynamically capture the deep dependencies and complex interactions between "data distribution views" and "pipeline structure views" in an end-to-end trainable manner, thereby significantly improving the expression and discrimination ability of global state representation.

[0112] Specifically, let the data set statistical features and the pipeline history global features be the input sources of different modalities, respectively. By taking the statistical features as the query vector (Query), and the pipeline history global features as the key / value vector (Key / Value), the cross-view attention mechanism can automatically calculate the attention weights of the statistical features under different pipeline history representations, achieving efficient extraction and aggregation of key information. Its attention score is given by the following formula:

[0113] ,

[0114] where, , , is a trainable reference. Compared with traditional concatenation or static weighted fusion methods, the introduction of cross-view attention mechanism greatly enhances the context adaptive ability, nonlinear expression and high-order interaction ability of global state representation, and the generalization of the model under different data distributions, task transfer and unknown scenarios.

[0115] ,

[0116] where, is the global representation of the pipeline component after aggregation by the attention mechanism, representing the deep interaction between data distribution features and pipeline history features.

[0117] The attention mechanism can automatically learn and dynamically adjust the attention weight between different information sources according to the actual task and the input distribution, so that the system is no longer limited to the simple information superposition of static channels when facing diversified data and complex pipeline structures, but realizes the organic unification of "global perception" and "local focus". This mechanism effectively captures the complex and dynamic coupling relationship between key components and statistical characteristics, improves the interaction depth and expression level between different modal views, and enriches the semantic dimension and discrimination ability of the state space. At the same time, the information reweighting ability of the cross-view attention mechanism enables the model to dynamically focus on the features and historical paths that contribute most to the current decision, which not only helps the model to exhibit stronger robustness in new data, unknown scenarios, distribution drift or task migration, but also proves the effectiveness and adaptability of the attention mechanism in improving the generalization ability of pipeline optimization and component recommendation system from both theoretical and empirical aspects.

[0118] Therefore, the dual-view feature fusion method based on the cross-view attention mechanism provides a solid theoretical foundation and innovative solution for the automated data preparation system in a complex, dynamic and variable data environment, and greatly enriches and expands the technical boundaries of data-driven pipeline modeling.

[0119] S3. Contextual semantic information extraction based on large language model

[0120] In the field of structure search and component recommendation of automated data preparation pipeline, although a large number of related works have tried to use genetic programming, Bayesian optimization, reinforcement learning and differentiable search to automatically select and sort pipeline components, there are still problems of exponential expansion of search space and huge reasoning overhead. The fundamental reason for this limitation is that most methods only rely on schema and basic statistical features to guide pipeline construction, lacking systematic utilization of domain context and deep data semantic information, making it difficult to effectively draw on the domain knowledge and practical experience relied on by data scientists when manually constructing the process. To address this deficiency, the present invention proposes a contextual semantic information extraction mechanism based on a large language model, which automatically mines multi-source contextual information of the dataset, including but not limited to table header column names, typical field values, domain terms, type and value distribution, etc., and converts these key information into high-dimensional vector representation, injecting domain priori and rich semantics into the pipeline component selection and search process.

[0121] Unlike traditional shallow representations that rely solely on table structure and statistics, large language models can encode table headers, representative sample rows, and field contents into a unified contextual input in the form of structured text, and achieve end-to-end embedding inference through deep neural networks. Large language models, exemplified by DeepSeek, not only possess powerful cross-domain semantic understanding capabilities but also adaptively model entity relationships, implicit contexts, and domain conventions in complex data tables. In practical applications, we transform the original data table into CSV format text with headers and data samples. Leveraging the efficient processing of extremely long contexts and token-level attention mechanisms of large language models, we effectively capture key features such as field naming, value distribution patterns, and synonym classification, ensuring that the embedding vectors simultaneously contain both structural information and domain semantics.

[0122] ,

[0123] in This represents the context vector generated by the large language model. This indicates that the large language model embeds an inference function. Furthermore, large models like DeepSeek enhance their robustness and generalization ability to dynamically changing table structures through input sequence truncation, prioritization of key fields, and comprehensive representation of multi-granular semantic information within the context window. Compared to traditional table-specific models that only predict from individual cells, large language models are not only unaffected by physical structure perturbations but also flexibly adapt to highly dynamic data environments. They effectively simulate the deep reasoning of field attributes and data semantics by human experts during pipeline construction, significantly improving the generalization, interpretability, and intelligence of automated pipeline recommendations.

[0124] More importantly, this invention fully leverages the powerful contextual understanding and representation capabilities of large language models, directly incorporating rich semantic information into the decision-making process of pipeline component selection and structure search. This context-based decision-making approach not only effectively simulates the process by which human experts narrow down the range of operational choices through domain knowledge, but also enables dynamic pruning and intelligent filtering of the candidate component space.

[0125] ,

[0126] Specifically, only when candidate components In the given context information conditional probability Greater than the preset threshold Only then will the component be selected into the final component set. This process dynamically narrows down the range of component selection by introducing contextual information, helping the system efficiently focus on those components that are most likely to improve data processing effectiveness. By introducing contextual embedding features for each step of the operation decision, the system can automatically focus on the most likely combination of components that can improve data processing effectiveness in a vast action space, significantly reducing the automated component search space, improving structural search efficiency, and reducing ineffective attempts and resource consumption. In this way, even in the face of highly complex and diverse data sets and task scenarios, efficient and accurate pipeline structure recommendation and optimization can be achieved, laying a solid semantic foundation and intelligent support for data preparation automation in complex business scenarios. To further improve the accuracy and efficiency of the pipeline component recommendation process, we propose an optimization target loss function that considers the impact of recommendation accuracy and contextual information to optimize the decision-making process in pipeline construction. The optimization target loss function is as follows:

[0127] ,

[0128] where, is the recommended pipeline component, is the actual component used, the first term is used to minimize the difference between the recommended component and the actual component, improving the accuracy of the recommendation. The second term is a regularization term, is the regularization coefficient, controlling the role of contextual information in the optimization process, ensuring that the impact of semantic information on the recommendation result is not too large, avoiding overfitting.

[0129] This technical path can automatically simulate the prior judgment and contextual reasoning of human experts in the pipeline construction process, significantly compressing the automated component search space and significantly improving the accuracy and explainability of the recommendation decision, and laying a solid semantic foundation and intelligent support for data preparation automation in complex business scenarios.

[0130] S4. Deep reinforcement learning network

[0131] In this method, the automatic construction of the data preparation pipeline is modeled as a reinforcement learning problem, aiming to achieve the optimal combination of processing components through the continuous learning of the agent. First, the system embeds all the key information in the current pipeline construction process, including the statistical characteristics of the current data set (such as the data type of each field, the proportion of missing values, the distribution, etc.), the history sequence of the executed pipeline, and the current available data preparation operations, to form the "state" This state comprehensively reflects the overall situation of the pipeline system at the time of decision-making, and is the basis for the agent to make each step of operation decision.

[0132] In each decision-making process, the agent will take the state as input, and predict the value of all available actions (i.e., candidate components) through a deep Q-network (DQN). The value not only measures the immediate reward brought by the current action, but also takes into account the expected future multi-step cumulative reward using the Bellman equation, achieving a dynamic balance between short-term and long-term benefits.

[0133]

[0134] Specifically, the action represents the selection of one data preparation component (such as missing value filling, normalization, encoding, outlier detection, etc.) from all available components to execute in the current pipeline. Each action selection will change the state of the dataset, thereby affecting the construction effect of the subsequent pipeline.

[0135] The policy is the decision rule used by the agent to determine which action to choose at each step. It is essentially a mapping controlled by parameters θ (usually neural network weights) that selects the component that maximizes the reward, i.e. . The input is the current state s, and the output is the priority or probability of each possible action. The calculation of the value follows the Bellman equation, which evaluates the expected value of future cumulative rewards when selecting component at the current step:

[0136]

[0137] where represents the immediate reward value obtained when selecting component at the current step, such as the improvement in downstream model performance. is the discount factor for balancing the immediate reward of the current decision and the cumulative reward that may be obtained in the future. The value of determines the agent's emphasis on "long-term returns" during the decision-making process. If is close to 1, the agent will pay close attention to the benefits brought by each step in the future, and will consider long-term effects as a major consideration. If is less than a certain value, the agent will focus more on the immediate reward obtained by the current action, and the attention to future rewards will decrease. By introducing the discount factor , the balance between "quick results" and "pursuit of long-term optimization" in the decision-making strategy can be effectively adjusted, so that the automatic construction process of the pipeline can not only consider the direct effect of each step selection, but also fully consider the global optimality of the overall structure. ​

[0138] Through end-to-end neural network training, DQN can automatically learn how to combine different data distributions and historical processing processes, dynamically adapt to environmental changes, and continuously optimize component selection without the need for human experience or fixed rules. This effectively avoids empiricism and the dilemma of local optima, ultimately achieving globally optimal recommendations for the data preparation pipeline structure. The core goal of reinforcement learning is to enable the agent to make optimal decisions under different data and scenarios through continuous trial and error and policy optimization, balancing immediate effects and long-term performance, and continuously improving the overall intelligence level of the automated pipeline.

[0139] In the actual training phase, to enable the pipeline's automatic model building to have strong generalization capabilities, this method learns a global decision-making strategy through continuous iteration. Specifically, the system first randomly selects a dataset from a large-scale training dataset pool containing different domains and diverse data distributions. (Get current data) Subsequently, the reinforcement learning agent constructs a state based on the current statistical features of the dataset and existing pipeline historical information. and select action Automatically generate an adapted data preparation pipeline. The generated pipeline will be applied to the dataset, and its effectiveness will be evaluated using metrics such as the performance of downstream models. Evaluation results. As a reward signal Feedback is given to the agent to optimize its decision-making strategy, enabling it to continuously generate higher-performing data preparation pipelines in subsequent training and inference.

[0140] ,

[0141] With continuous training on heterogeneous cross-domain data, the learned global strategy not only improves pipeline performance on a single dataset but also maintains good generalization and robustness in new tasks and unknown domains. DQN employs a target network and empirical replay for stable training on samples. Minimize loss:

[0142] ,

[0143] and with learning rate Update This training mechanism ensures that the automated pipeline recommendation system can continuously output high-quality pipeline combinations across diverse data environments, fully meeting the intelligent needs of large-scale, heterogeneous data scenarios. During the inference phase, new task data is processed... The optimal pipeline can be expressed as:

[0144] ,

[0145] The present application is directed to the field of automated data preparation pipeline construction, which addresses the core bottlenecks of search space expansion, insufficient generalization, and limited intelligence. A three-layer structured scheme is proposed, which combines large language models and self-attention mechanisms. This method is driven by theory and guided by engineering feasibility, breaking the traditional pipeline optimization paradigm that relies solely on statistical features and table structures. It achieves deep integration of data distribution characteristics, historical pipeline information, and contextual semantic knowledge, greatly improving the intelligence, adaptability, and generalization of pipeline automation recommendations. Overall, the present application not only enriches the technical means in the field of automated data preparation, but also provides a new paradigm for efficient and robust data pipeline construction in multiple tasks and scenarios.

[0146] S5. Reasoning and executable output

[0147] To enable the application of the trained strategy in actual business scenarios, the present application sets up a reasoning and delivery link after "S4. Deep reinforcement learning network", achieving automatic pipeline generation, compliance explanation, and executable output for new data / new tasks. This link completes the end-to-end closed loop from context perception to pipeline landing without relying on human participation.

[0148] S5.1. Reasoning input and initialization

[0149] Let the new task data be , and the optional label of the supervised task be (null in unsupervised / weakly supervised scenarios) for input reasoning; with field type identification and consistency checking mechanism to generate multi-dimensional labels such as numerical, categorical, text, time series, geographic information, or identifier. At the same time, the system identifies fields that may contain sensitive information based on regularization expressions and domain dictionary rules, and performs security processing such as masking and hashing to ensure that privacy is not leaked during the reasoning process. Then, representative sample rows are selected through hierarchical sampling method, and value domain sketching is performed on data features, such as extracting quantiles, mean, variance, and missing rate in numerical fields, extracting cardinality, main category distribution, and long tail proportion in categorical fields, extracting key n-grams and keywords in text fields, and extracting span, periodicity, and trend information in time fields; at the same time, field units are automatically identified and scaled to establish a highly comparable feature description.

[0150] The above processing results will be encoded as structured semantic text, and on this basis, synonym merging and term unification will be performed to ensure semantic consistency of fields such as "gender", "sex", and "gender". Subsequently, the system maps table headers, field types, value domain sketches, and representative samples to embedding space through a semantic encoding model, obtaining field-level and table-level embedding vectors, and generating context semantic representations through attention aggregation Meanwhile, statistical views are extracted from the whole data including at least field type encoding, missing rate, unique value ratio, zero variance flag, numerical distribution points, dominant category proportion of sub-type, and time column periodicity and trend factor. The statistical characteristics are encoded in a scientific counting method to retain the order of magnitude information and proportional information. Finally, the system searches for similar tasks in the task library based on the embedding retrieval mechanism, and if the similarity exceeds the set threshold, the corresponding pipeline prefix is extracted as the initial pipeline history , and fused to form the initial state vector of reasoning :

[0151] .

[0152] S5.2 Candidate specification and strategy inference

[0153] After the initial state vector is generated, the reasoning process enters the candidate component specification and strategy inference phase. The present invention carries out candidate specification under the constraint conditions of the component library The component library maintains input-output contract, preconditions, mutual exclusion relationship, dependency graph, and resource and security level metadata for each component. The system first calculates the conditional prior probability based on the context semantic representation , which comprehensively considers the field matching degree, the historical marginal contribution of similar tasks, and the potential risk deduction, and is calibrated by temperature to control the confidence interval. Then, the system calculates the comprehensive score of each candidate component:

[0154] ,

[0155] Among them, the compatibility term reflects the semantic consistency with the current state, and the resource adaptation term reflects the feasibility under the computing power, storage, and time delay budget. If the component has hard violations in compliance, security, or interpretability indicators, it will be directly excluded, and the rejection reason is recorded.

[0156] After the candidate set specification is completed, the reasoning engine enters the decision cycle. This stage calls the pre-trained strategy and its corresponding action value function under the mild exploration mechanism The system gradually selects the optimal component sequence and constructs the pipeline by combining greedy selection with temperature softening. After each action selection, the agent generates a local candidate pipeline and performs rapid verification through the agent evaluation mechanism, including task fitness simulation, resource consumption estimation, and explainability threshold testing. If the local pipeline meets the threshold in multiple dimensions, it proceeds to the next iteration; otherwise, the system reverts and reselects the candidate. When the iteration reaches the upper limit or the evaluation result converges, the reasoning process terminates, and the system outputs the final pipeline design as the model's reasoning result.

[0157] S5.3 State transition and action execution

[0158] After the candidate components are selected and executed, the system updates the current environment state in real time according to the state transition mechanism of reinforcement learning . This state update not only reflects the evolution of the data processing flow at the semantic level but also includes changes in resource consumption, performance indicators, and compliance constraints, ensuring that the state vector can dynamically depict the running situation of the entire pipeline. In each transition, the system embeds the newly generated intermediate result features into the semantic space, fuses them with the previous context representation, and forms a new context vector through weighted aggregation . At the same time, the pipeline history trajectory is updated to include the sequence description of the currently executed components, enabling subsequent candidate selection under more complete context information.

[0159] ,

[0160] The system gradually approaches the optimal solution of the target task through the above recursive state transition.

[0161] S5.4 Global evaluation and pipeline optimization

[0162] During the gradual construction of the pipeline, the invention introduces a global evaluation mechanism to dynamically score the candidate pipeline in multiple dimensions. This score covers accuracy, generalization, resource utilization, reasoning latency, explainability, and compliance, among other indicators. Specifically, the system simulates the performance of the pipeline on new data through a fast proxy model and corrects the bias based on heuristic correction methods; at the resource level, it estimates the computational complexity, memory consumption, and storage pressure of each candidate pipeline to ensure that it meets the requirements of the constraint set ; in terms of compliance and explainability, the system verifies whether each step of the pipeline has traceability through an explainability rule base and logic checker, avoiding "black box" operations. If the global score is below the threshold, the system reverts to the previous state and reselects the candidate action until a performance-converged and stable pipeline sequence is obtained.

[0163] S 5.5 Global evaluation and pipeline optimization

[0164] When the system completes the iteration and generates the final pipeline, the pipeline is submitted to the user or subsequent module as the output result of the inference. The output forms include executable component sequences, key hyperparameter configurations, resource consumption evaluations, and performance prediction indicators, ensuring that the results can be directly deployed and can also provide the basis for subsequent manual review and modification. At the same time, the system also establishes a feedback mechanism to compare the inference output with the verification result after actual execution. If a prediction deviation is found, the deviation information is returned to the policy network and the value function network, so that the model parameters are gradually corrected in future inferences.

[0165] ,

[0166] This closed-loop optimization not only ensures the reliability and repeatability of the output results, but also enables the invention to have the ability of self-iterative evolution.

[0167] Experimental verification:

[0168] To fully verify the universality and robustness of the system proposed in the invention, multiple public benchmark datasets and real business data are selected for experiments in this embodiment. The datasets used cover common classification and clustering datasets in machine learning, and can examine the performance of the system in classification, prediction, recommendation, and representation learning tasks from different dimensions.

[0169] The experimental environment is Ubuntu 22.04, equipped with NVIDIA RTX 4090 GPU and 256GB memory, and model training is implemented using PyTorch. In terms of hyperparameter settings, the batch size is 64, the learning rate is set to 0.001, the optimizer is Adam, and the weight decay coefficient is 1e-5. In the attention mechanism, the number of attention heads is set to 8, and the embedding dimension is set to 128. The step size of the diffusion model is set to 6, and the noise level is dynamically sampled in the range of [0.0001, 0.2]. The early stopping strategy (patience=10) is used during training to avoid overfitting

[0170] In terms of classification tasks, the UCI Adult and Kaggle Credit datasets are selected. The former contains more than 48,000 samples, with features including education, occupation, marriage, and other multi-dimensional information, and the target is a binary classification of income, used to verify the effectiveness of the system in processing medium-sized structured data. The latter contains more than 300,000 samples and is aimed at credit card fraud identification, with extremely unbalanced classes, suitable for testing the robustness of the system under sparse and unbalanced conditions.

[0171] In terms of clustering tasks, typical datasets from UCI ML Repository are adopted, including Iris and Wine datasets. The Iris dataset is a classic multi-dimensional flower dataset, containing 150 samples, and the target is unsupervised clustering and three class labels correspond; the Wine dataset contains 178 samples and 13 chemical attribute features, commonly used for clustering and classification hybrid tasks in multi-dimensional feature space, which can effectively test the adaptability of the system in small-scale high-dimensional data.

[0172] Table 1 Dataset Information

[0173]

[0174] In addition, in order to further improve the comprehensiveness of the verification, 18 multi-class public comprehensive benchmark datasets are introduced (see Table 1 for dataset information). The sample size of these datasets ranges from thousands to hundreds of thousands, the number of features ranges from 6 to 81, the number of classes covers binary classification, multi-classification and clustering tasks, and some datasets also contain missing values (such as house and uscensus), further enhancing the complexity of the verification conditions. In addition, the difference in output dimension also provides a test basis for the generalization performance of the system in multi-task scenarios.

[0175] In terms of experimental parameter settings, the batch size is set to 64, the learning rate is set to 0.001, the optimizer is Adam, and the early stopping strategy is used in the training process to prevent overfitting. For the attention mechanism module, the number of attention heads is set to 8, and the embedding dimension is set to 128. The step size in the diffusion process is set to 6, and the noise level is dynamically sampled in the range of [0.0001, 0.2] to enhance the robustness of the model to missing data and noisy data.

[0176] In classification tasks, the goal is to maximize prediction accuracy and F1-score; in clustering and recommendation tasks, the focus is on the adaptability and generalization ability of the system in different class distributions and high-dimensional spaces. Due to the existence of extremely unbalanced class distribution in some datasets (such as Kaggle Credit), this experiment additionally uses a combination of undersampling and weighted loss function in the training process to avoid the loss of minority class information.

[0177] Table 2 Experimental Results

[0178]

[0179] The comparative methods include DEF, RS, DP-Fix, DP-Flex, DL, HAI-AI, SAGA and the CtxPipe system proposed in this paper. The experimental evaluation indexes cover test accuracy (Accuracy), macro-averaged precision (Precision), recall (Recall), F1-score, running time (Running Time), etc., and in addition, the significance test (t-test and sign test) is carried out on the extended data set to ensure that the results have statistical reliability.

[0180] Table 2 shows the average results of different methods on the DiffPrep dataset and the extended dataset. From the table, it can be seen that the CtxPipe proposed in this paper achieves the best performance in most indicators, the test accuracy reaches 0.806, which is 2.8% and 3.3% higher than the suboptimal methods DP-Flex (0.784) and DP-Fix (0.780) respectively; the highest values are also achieved in precision, recall and F1-score, which fully verify the effectiveness and robustness of the method in the multi-task and multi-data environment. At the same time, in terms of running efficiency, CtxPipe maintains an acceptable level, which is much better than the method with large computational overhead such as DP-Flex. The significance test results (p<0.05) further show that the improvement of the method in this paper has statistical significance.

[0181] A computer readable storage medium, wherein a plurality of instructions are stored, the instructions are suitable for being loaded and executed by a processor of a terminal device, and the instructions are suitable for implementing the context-aware semantic-aware data preparation pipeline recommendation method.

[0182] A terminal device, comprising a processor and a computer readable storage medium, the processor is used to implement instructions, and the computer readable storage medium is used to store a plurality of instructions, the instructions are suitable for being loaded and executed by the processor, and the instructions are suitable for implementing the context-aware semantic-aware data preparation pipeline recommendation method.

[0183] The above are preferred embodiments of the present application, which do not limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A context- and semantically aware data preparation pipeline recommendation method, characterized in that, include: Obtain the automated pipeline build dataset; Perform data preprocessing on the acquired automated pipeline construction dataset; For the preprocessed dataset, a dual-view feature fusion based on a self-attention mechanism is performed; For the preprocessed dataset, contextual semantic information is extracted based on a large language model; Based on feature fusion and information extraction, pipeline recommendation results are obtained through a deep reinforcement learning network; The pipeline recommendation results are evaluated and optimized. Output pipe; The aforementioned contextual semantic information extraction based on a large language model includes introducing a contextual semantic information extraction mechanism based on a large language model. This mechanism involves automatically mining multi-source contextual information from the dataset. Specifically, by converting the original data table into CSV format text with headers and data samples, the mechanism utilizes the large language model's ability to handle extremely long contexts and its token-level attention mechanism to capture key features such as field naming, value distribution patterns, and synonym categorization. This is represented as follows: ， in This represents the context vector generated by the large language model. This represents the embedding of inference functions into a large language model, followed by a context-based decision-making process that simulates the process by which experts narrow down the range of operational choices using domain knowledge. This achieves dynamic pruning and intelligent filtering of the candidate component space, and is represented as follows: ， Among them, when candidate components In the given context information conditional probability Greater than the preset threshold At that time, the component is selected into the final component set. By introducing contextual embedding features into each operational decision, the system automatically focuses on the component combinations most likely to improve data processing performance within the action space. Simultaneously, it proposes an optimized objective loss function, considering the impact of recommendation accuracy and contextual information to optimize the decision-making process in pipeline construction, expressed as: ， in, It is a recommended piping component. These are the components that are actually used. The regularization coefficient controls the context information. Its role in the optimization process.

2. The context-semantic-aware data preparation pipeline recommendation method according to claim 1, characterized in that, The process of preprocessing the acquired automated pipeline construction dataset includes: identifying data fields closely related to the target labels from various heterogeneous data sources according to the task definition; performing category distribution statistics by explicitly identifying label columns after initial feature and label screening; simultaneously capturing contextual information of tabular data during the extraction phase and achieving precise labeling of data fields through a column name-semantic correspondence mechanism; and finally, introducing a consistency and standardization mechanism to uniformly map label codes from different sources to a consistent category set by aligning the labels. To ensure the uniqueness of the classification semantics, uniform numerical operations are performed on the features. Numerical fields are normalized or binning strategies, while discrete fields are transformed into dictionary indices or embedding vectors to ensure comparability of different features in the same space. For missing data, imputation or removal methods are used to ensure the integrity of the feature matrix. After this process, the final classification task representation is obtained. ， in and These are the standardized features and labels, This is contextual metadata that is bound to the data.

3. The context-semantic-aware data preparation pipeline recommendation method according to claim 2, characterized in that, The self-attention-based dual-view feature fusion includes dataset statistical view feature extraction and pipeline history component feature extraction. For the dataset statistical view feature extraction, a detailed statistical feature extraction is first performed on the input dataset. For cases where the numerical feature distribution has a large span, a combination of fractional and exponential parts is used to represent all numerical features. by Encoding is performed in the following manner, where The fractional part representing the feature. Indicates its exponential part, both are expressed through The original features are jointly represented; for feature extraction of pipeline history components, a sequence modeling neural network is introduced as the core encoding module, which extracts features from all executed data processing components in historical order. It is transformed into a dense vector sequence through the embedding layer. ,in For the number of components, For the embedding dimension, then the dense vector sequence The input is fed into a multi-layered stacked Transformer layer, which models the global dependencies between all components in the sequence in parallel, as follows: ， Then, a weighted pooling approach is used to extract global features from the sequence to represent the historical context information of the current pipeline.

4. The context-semantic-aware data preparation pipeline recommendation method according to claim 3, characterized in that, The self-attention-based dual-view feature fusion also includes introducing a cross-view attention-based feature fusion method to dynamically capture the interaction between the data distribution view and the pipeline structure view in an end-to-end trainable manner. Here, the statistical features of the dataset are assumed to be... Historical global characteristics of pipelines Using these as input sources for different modalities, and by using the statistical features of the dataset as query vectors, As the key, the attention weights of statistical features under different pipeline historical representations are automatically calculated and expressed as follows: ， in, , , As a trainable reference, the global representation of the pipeline components is obtained after aggregation through an attention mechanism. , represented as .

5. The context-semantic-aware data preparation pipeline recommendation method according to claim 4, characterized in that, The pipeline recommendation result, obtained through a deep reinforcement learning network based on feature fusion and information extraction, includes embedding all information from the pipeline construction process as a state. In the decision-making process, through in-depth The network predicts all possible actions. The value, achieved by using the Bellman equation to balance the expected cumulative rewards over multiple future steps, is a dynamic equilibrium between short-term and long-term returns, expressed as: ， in, Indicates the state given Take action below And according to the current parameters The strategy that can achieve the long-term cumulative expected value Indicates the execution of an action Then transition to the next state , Represents the discount factor, and then based on the strategy Choose the component that maximizes the reward. ,Right now Given the current state s as input, output the priority or probability of each possible action. The value calculation follows the Bellman equation, which evaluates the selected component in the current step using the formula. The expected value of the cumulative reward in the future is expressed as: ， in Indicates the selected component in the current step. The instant reward value obtained at that time As a discount factor for rewards, if A value close to 1 indicates that the AI ​​is highly attentive to the benefits of each future step, taking long-term effects into primary consideration; if... If the value is less than the set value, the agent will focus more on the immediate reward obtained from the current action and pay less attention to future rewards.

6. The context-semantic-aware data preparation pipeline recommendation method according to claim 5, characterized in that, The pipeline recommendation results obtained through feature fusion and information extraction via a deep reinforcement learning network also include training the deep reinforcement learning network by first randomly selecting datasets from a large-scale training dataset pool containing different domains and diverse data distributions. Get the current data The current statistical characteristics of the dataset and the existing historical information of the pipeline constitute the state. Select Action Automatically generate an adapted data preparation pipeline and evaluate its effectiveness using performance metrics from downstream models. Evaluation results... As a reward signal Feedback is given to the agent to optimize the decision-making strategy, and is represented as: ， Through continuous training on heterogeneous cross-domain data, the ability to train on samples is achieved. Minimize the loss, expressed as: ， and with learning rate Update This ensures that the automated pipeline recommendation system can consistently output high-quality pipeline combinations across diverse data environments, and adapts to new task data during the inference phase. The optimal pipeline is represented as: ， in Represents the set of all candidate pipelines The pipeline configuration that maximizes expected cumulative rewards; It represents the mathematical expectation of a stochastic process and measures the average performance.

7. The context-semantic-aware data preparation pipeline recommendation method according to claim 6, characterized in that, The evaluation and optimization of pipeline recommendation results includes taking into account new task data. The optional tags for the supervision task are: Input inference is performed; multi-dimensional labels are generated using field type recognition and consistency verification mechanisms, including numerical, categorical, text, time series, and geographic information or identifiers; representative sample rows are selected using stratified sampling methods; value range sketching is performed on data features; and statistical views are extracted from the overall data. Finally, similar tasks are found in the task library based on an embedding retrieval mechanism. If the similarity exceeds a set threshold, the corresponding pipeline prefix is ​​extracted as the initial pipeline history. and and The initial state vector for reasoning is formed by fusion. : ， After the initial state vector is generated, in the component library Candidate reduction is carried out under constraints based on contextual semantic representation. Calculate conditional prior probabilities Then, a comprehensive score is calculated for each candidate component: ， Invoke the pre-training strategy and its corresponding action value function Under a gentle exploration mechanism, the optimal component sequence is selected step by step and the pipeline is built.

8. The context-semantic-aware data preparation pipeline recommendation method according to claim 7, characterized in that, The evaluation and optimization of the pipeline recommendation results also includes updating the current environment state in real time based on the state transition mechanism of reinforcement learning after the candidate component is selected and executed. In each transition, the newly generated intermediate features are embedded into the semantic space, fused with the previous context representation, and a new context vector is formed through weighted aggregation. Meanwhile, the historical trajectory of the pipeline The sequence description is updated to include the currently executed components, allowing subsequent candidate selections to be made with more complete contextual information. Through the aforementioned recursive state transitions, the optimal solution for the target task is gradually approximated. During pipeline construction, a global evaluation mechanism is introduced to dynamically score candidate pipelines across multiple dimensions. When the iteration is complete and the final pipeline is generated, the pipeline is used as the output of the inference. Based on a feedback mechanism, the inference output is compared with the verification results after actual execution. If a prediction deviation is found, this deviation information is fed back to the policy network and the value function network, thereby gradually correcting the model parameters in future inferences. This is expressed as: ， in This represents the difference between actual performance and predicted performance, used to measure the discrepancy between the predictive model and the actual results. Indicates model parameters, The learning rate controls the step size for each parameter update. This is used to measure performance prediction error.

9. A context-semantic-aware data preparation pipeline recommendation system, executing the context-semantic-aware data preparation pipeline recommendation method as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire the dataset built by the automated pipeline; The preprocessing module is configured to preprocess the acquired automated pipeline construction dataset; The feature fusion module is configured to perform dual-view feature fusion based on a self-attention mechanism on the preprocessed dataset. The information extraction module is configured to extract contextual semantic information based on a large language model from the preprocessed dataset. The pipeline recommendation module is configured to obtain pipeline recommendation results through a deep reinforcement learning network based on feature fusion and information extraction. The optimization module is configured to evaluate and optimize the pipeline recommendation results; The output module is configured as an output pipe.

Images