A large language model-based recommendation system user behavior sequence feature enhancement method

Abstract

The application discloses a recommendation system user behavior sequence feature enhancement method based on a large language model, comprising the following steps: acquiring a recommendation dataset and constructing a first text according to the feature structure analysis of the use and conventional fields of the recommendation dataset in a recommendation system, and inputting the large language model to generate a dataset description; constructing a second text according to a target task and a sequence feature type, and inputting the large language model to generate feature semantics together with the dataset description; constructing a third text based on the feature semantics, and inputting the large language model to generate feature codes; embedding the feature codes into a prediction model training environment to realize the automatic generation and loading of new features in training data; and performing a new feature experiment in a prediction model framework to evaluate a new round of sequence features. The application has stronger expressiveness, context perception and task adaptability in the recommendation system and the personalized system facing behavior data modeling, and can effectively solve the technical bottlenecks of existing methods in structure, expression and generalization.

Description

Technical Field

[0001] This invention belongs to the field of feature engineering technology, specifically relating to a method for enhancing user behavior sequence features in a recommender system based on a large language model (LLM). Background Technology

[0002] In various AI applications that rely on behavioral data (data-driven) such as recommender systems, personalized ranking, user modeling, classification tasks, financial risk control, and medical diagnosis, feature engineering has always been considered one of the key factors affecting model performance. Feature enhancement, as an important component of feature engineering, aims to generate higher-quality features with greater discriminative and expressive power based on the original features, thereby improving the performance of downstream models.

[0003] Existing feature enhancement techniques primarily focus on processing static, discrete features (such as gender, city, and product category) in structured data, including methods such as feature combination, feature cross-validation, multinomial expansion, and statistical feature extraction. Currently, several feature enhancement-related technical solutions have been proposed, with representative solutions including automated feature generation frameworks such as AutoFeat, Deep FeatureSynthesis, and OpenFE. These solutions, categorized by method type, mainly include: 1. Artificial construction feature scheme: In traditional machine learning methods, feature enhancement primarily relies on the knowledge and experience of domain experts. For example, in financial risk control, statistical features such as "average spending over the past 7 days" and "average monthly repayment rate" are constructed based on historical transaction data; in recommendation systems, features such as "number of user interactions with products" and "time of the most recent click" are commonly used as enhancement features. While these methods offer advantages such as strong interpretability and ease of implementation, they suffer from poor scalability and weak generalization ability, and are heavily reliant on human experience, making them difficult to adapt to large-scale, dynamically changing data scenarios.

[0004] 2. Automated feature combination and cross-scheme: As feature dimensions increase, many methods have been proposed to automatically combine features. For example, constructing multinomial combinations of original features, and using models such as DeepFM combined with factorization machines (FM) and neural networks to automatically mine low-order and high-order feature interactions.

[0005] While such methods expand the feature space through automatic combination, they are prone to introducing redundant features, and when the feature semantics are unclear, the combination is unlikely to produce meaningful enhanced information.

[0006] 3. Feature propagation and enhancement scheme based on graph structure: For structured data with user-item relationships, entity relationships, etc., some methods use graph neural networks (GNNs) and other techniques to enhance features.

[0007] 4. Representation augmentation schemes based on large models / pre-trained models: With the widespread application of pre-trained large models (such as BERT, GPT, CLIP, etc.), researchers are increasingly using these models to extract and enhance features from unstructured information such as text and images. For example, BERT is used to generate context vectors for product descriptions as supplementary product features; CLIP is used to extract image semantics as product or user preference features. Although these methods can capture complex patterns at the semantic level, they still face challenges such as how to effectively integrate them with structured features and how to adapt them to specific downstream tasks.

[0008] 5. Scheme based on Automatic Feature Generation (AutoFE): Some automated feature engineering frameworks (such as AutoFeat and OpenFE) attempt to automatically mine potential augmented features through heuristic search, evolutionary algorithms, and reinforcement learning. While these methods can systematically extract high-quality features from raw data, alleviating reliance on human experience, they are typically computationally expensive, have unclear optimization objectives, and struggle to quickly adapt to specific scenario requirements, especially exhibiting unstable performance in small sample sizes or cold-start situations.

[0009] Therefore, it is evident that the aforementioned existing solutions primarily target single-point, static feature enhancement tasks, lacking the ability to model dynamic sequence features in behavioral data, and exhibiting the following prominent problems: 1. This approach only enhances discrete features, neglecting the structural and semantic representation of temporal behavioral features. Current methods generally employ manual construction methods to process user behavior sequences, such as building statistical features like "recently clicked product IDs" or "number of views in the last 7 days." This approach has limited ability to express dynamic user interests and contextual changes, and lacks versatility and scalability.

[0010] 2. There is a lack of methods for generating sequence features and a lack of a systematic modeling framework. Unlike static features, sequence features are variable in length, time-dependent, and context-sensitive. Currently, there are no mature methods to support automatic, task-related feature generation for sequence data, resulting in existing feature engineering processes being crude in their handling of sequence data and failing to fully realize their modeling value.

[0011] 3. Enhanced features lack personalization and semantic fusion capabilities, making them unsuitable for complex recommendation scenarios. Existing enhancement methods lack the ability to model individual user differences and contextual information, and cannot dynamically generate personalized sequence features with semantic differences based on specific recommendation tasks (such as click-through rate prediction, ranking, and multi-objective optimization).

[0012] 4. The potential of large models has not yet been effectively utilized in structured sequence feature generation. Although large language models (such as GPT and T5) have demonstrated powerful generative capabilities in fields such as natural language processing, existing research has not fully explored their applications in behavioral data modeling and structured sequence feature generation. The related technologies are still immature and suffer from problems such as poor controllability, ambiguous generated structures, and difficulty in adapting to model training. Summary of the Invention

[0013] To address the problems existing in the prior art, this invention provides a method for enhancing user behavior sequence features in recommendation systems based on a large language model. This method has stronger expressive power, context awareness, and task adaptability in recommendation systems and personalized systems that model behavioral data, and can effectively solve the technical bottlenecks of existing methods in terms of structure, expression, and generalization.

[0014] To solve the above-mentioned technical problems and achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: A method for enhancing user behavior sequence features in a recommender system based on a large language model includes the following steps: Step 1) Obtain the recommendation dataset; Step 2) Based on the purpose and common fields of the recommended dataset in the recommendation system, the feature structure of the recommended dataset is parsed, and the first text (Prompt 1) is automatically constructed using the first custom code generation template. Then, the first text (Prompt 1) is input into the Large Language Model (LLM) to generate a natural language description related to the recommended dataset, i.e., a dataset description, which is used to provide a semantic basis for subsequent feature generation. Step 3) Based on the target task (such as click-through rate prediction, ranking optimization, etc.) and the sequence feature type, the second text (Prompt 2) is automatically constructed using the second custom code generation template. Then, the dataset description and the second text (Prompt 2) are input into the Large Language Model (LLM) to generate an enhanced text feature description for the user behavior sequence (such as recent user behavior patterns, cross-time period interest changes, etc.), i.e. feature semantics. Step 4) Based on the newly added feature semantics, the third text (Prompt 3) is automatically constructed using the third custom code generation template. Then, the third text (Prompt 3) is input into the Large Language Model (LLM) to generate the source code that implements the feature extraction, i.e., the feature code. Step 5) Embed the confirmed feature code into the prediction model training environment of the recommendation dataset, so as to combine it with the original data processing and feature engineering to realize the automatic generation and loading of new features in the training data; Step 6) In the existing prediction model framework of the recommendation system, conduct experiments on new features, input the new prediction results (detection index data) such as accuracy, recall, and AUC obtained by the prediction model into the cache, and evaluate and generate a new round of sequence features based on this.

[0015] Furthermore, the large language model (LLM) can be replaced with other models that have natural language understanding and generation capabilities, such as multimodal large models (combining text and structured data inputs) or dedicated sequence modeling neural networks (such as Transformer, BERT, T5, etc.), to achieve semantic understanding of data fields and generation of sequence features.

[0016] Furthermore, the first text (Prompt 1), the second text (Prompt 2), and the third text (Prompt 3) are all generated using dynamic Prompt generation technology, which automatically optimizes the Prompt content based on the characteristics of the dataset, such as by using meta-learning or rule-based Prompt adjustment mechanisms, thereby improving the relevance and executability of the model-generated results.

[0017] Furthermore, the stated uses and common fields include user identifier, item identifier, behavior type, timestamp, and contextual attributes.

[0018] Furthermore, the first text (Prompt 1) constructed from the first custom code generation template includes at least the phrases "task topic," "input content," and "task requirements"; wherein, The “task theme” section includes sections on “content described,” “method of description,” and “purpose of description”; The meaning of the "content of description" paragraph is: generate a description of the input recommendation dataset and its various fields; The meaning of the "method of description" paragraph is: the description is presented in a detailed and natural language manner; The meaning of the “purpose of description” paragraph is that description should serve as a semantic basis for feature engineering in the recommender system; The “Input Content” field contains the dataset name, a list of domain names, and their example values; The “Task Requirements” section includes sections such as “Overall Description Requirements for the Dataset”, “Description Requirements for Each Field of the Dataset”, “Description Statement Requirements”, “Example Value Requirements”, and “Applicability Requirements”. The meaning of the “overall description requirements for the dataset” paragraph is: describe the overall purpose or domain of the input recommended dataset based on its name and field content; The phrase “requirements for describing each field of the dataset” means that for each field of the input recommendation dataset, describe its meaning, its potential role in model building, and the type of information it contains. The meaning of the “Description Statement Requirements” section is: descriptions should use clear, formal, and complete statements; The meaning of the “example value requirement” paragraph is: avoid repeating precise example values ​​unless it is necessary to explain their meaning; The meaning of the “applicability requirement” paragraph is: to make the description universal, so that it applies to a dataset rather than a single record.

[0019] Furthermore, the second text (Prompt 2) constructed from the second custom code generation template includes at least the following segments: "Dataset Description," "Prediction Target," "Input Field Content," "Input Field Format," "Baseline Performance of the Default Model," "Previously Generated Features and Their Evaluation Results," "Generation Requirements," and "Output Content." The dataset description is filled in the “Dataset Description” section; The “prediction target” section contains the object to be predicted by the recommendation system model; The “Input Field Content” section contains the name, sample, and brief description of the dataset fields; The “Input Field Format” section specifies the format for data input; The "Baseline performance of the default model" section contains existing baseline detection values; The phrase “previously generated features and their evaluation results” contains the rounds of previously generated sequence features and the corresponding detection results. If the detection result is determined to be a valid sequence feature, then the corresponding detection result is input; if the detection result is determined not to be a valid sequence feature, then the judgment of non-sequence feature is input. The “Generation Requirements” section specifies the generation directions to be avoided and the generation directions to be attempted. The generation directions to be avoided include: avoiding copying sequence feature names and descriptions generated in previous rounds, and avoiding relying on patterns that performed poorly in previous rounds; The proposed directions for generation include: exploring new combinatorial features driven by behavior, time, or diversity, and exploring creative combinations of existing columns. The “output content” segment contains the feature name, feature description, and feature type.

[0020] Furthermore, the third text (Prompt 3) constructed from the third custom code generation template shall at least include segments such as "description of newly added sequence features", "prediction target", and "input field format"; wherein, The “Description of New Sequence Features” section contains the “Output Content” from the previous round; The “prediction target” section contains the object to be predicted by the recommendation system model; The “Input Field Format” section specifies the format for data input.

[0021] Furthermore, if the semantic features generated by the Large Language Model (LLM) conform to the sequence feature definition required to construct the third text (Prompt 3), they are used for the construction of the third text (Prompt 3) in the next step. If they do not conform to the sequence feature definition required to construct the third text (Prompt 3), the construction is considered to have failed, and the text is returned to the cache for parsing and filtering. The semantic features that conform to the task objectives and semantic constraints are retained and recorded in the local feature enhancement document for subsequent construction of the third text (Prompt 3).

[0022] A computer device includes a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface communicate with each other via the communication bus. The memory stores at least one executable instruction, which causes the processor to perform the operation corresponding to the above-mentioned method for enhancing user behavior sequence features in a recommendation system based on a large language model.

[0023] A computer storage medium storing at least one executable instruction that causes a processor to perform the operation corresponding to the above-described method for enhancing user behavior sequence features in a recommendation system based on a large language model.

[0024] The beneficial effects of this invention are as follows: This invention proposes a method for generating and enhancing sequence features based on a large model. This method establishes an automatic sequence feature generation mechanism for time-series behavioral data, thereby breaking through the dependence of traditional feature enhancement methods on static discrete features. It enhances the dynamic modeling capability of user interests, improves the relevance and adaptability of generated features to specific task objectives (such as CTR prediction), achieves the fusion generation of structured behavioral information and unstructured semantic information (such as text, images, and contextual cues), and enhances the expressive power of features. This improves the generalization performance and robustness of the model in challenging scenarios such as complex behavior prediction, cold start, and small sample sizes.

[0025] In summary, this invention provides a sequence feature enhancement technology solution with stronger expressiveness, context awareness, and task adaptability for recommendation and personalization systems oriented towards behavioral data modeling. It can effectively solve the technical bottlenecks of existing methods in terms of structure, expression, and generalization.

[0026] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0027] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the user behavior sequence feature enhancement method for recommender systems based on a large language model, as described in this invention. Figure 2 This is a template diagram of the first text generated using the first custom code template of the present invention; Figure 3 This is a template diagram of the second text constructed using the second custom code generation template of the present invention; Figure 4 This is a template diagram of the third text generated using a third custom code template according to the present invention. Detailed Implementation

[0028] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings to provide a clearer understanding of the invention's purpose, features, and advantages. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the invention's technical solution.

[0029] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known apparatuses, structures, and techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0030] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0031] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0032] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0033] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0034] See Figure 1 As shown, this invention provides a method for enhancing user behavior sequence features in a recommender system based on a large language model, comprising the following steps: Step 1) Obtain the AOMI recommendation dataset.

[0035] Step 2) Dataset usage and field analysis.

[0036] First, based on the purpose of the Aumi recommendation dataset in the recommendation system and its common fields (including user identifier, item identifier, behavior type, timestamp, and contextual attributes), the feature structure of the Aumi recommendation dataset is analyzed. The first text (Prompt 1) is constructed using a static first custom code generation template. Then, the first text (Prompt 1) is input into the Large Language Model (LLM) to generate a natural language description related to the recommendation dataset, i.e., the dataset description, which provides a semantic basis for subsequent feature generation.

[0037] The first text (Prompt 1) constructed from the first custom code generation template contains at least the following segments: "Task Topic," "Input Content," and "Task Requirements"; among which, The “task topic” section includes sections on “content described,” “method of description,” and “purpose of description”; The meaning of the "content of description" paragraph is: generate a description of the input recommendation dataset and its various fields; The meaning of the "method of description" paragraph is: the description is presented in a detailed and natural language manner; The meaning of the “purpose of description” paragraph is that description should serve as a semantic basis for feature engineering in the recommender system; The “Input Content” field contains the dataset name, a list of domain names, and their example values; The “Task Requirements” section includes sections such as “Overall Description Requirements for the Dataset”, “Description Requirements for Each Field of the Dataset”, “Description Statement Requirements”, “Example Value Requirements”, and “Applicability Requirements”. The meaning of the “overall description requirements for the dataset” paragraph is: describe the overall purpose or domain of the input recommended dataset based on its name and field content; The phrase “requirements for describing each field of the dataset” means that for each field of the input recommendation dataset, describe its meaning, its potential role in model building, and the type of information it contains. The meaning of the “Description Statement Requirements” section is: descriptions should use clear, formal, and complete statements; The meaning of the “example value requirement” paragraph is: avoid repeating precise example values ​​unless it is necessary to explain their meaning; The meaning of the “applicability requirement” paragraph is: to make the description universal, so that it applies to a dataset rather than a single record.

[0038] See Figure 2 As shown, Figure 2 The template corresponding to the first text (Prompt 1) is shown, where {dataset_name} is the dataset name, {field_name} is the dataset field name, and {field_ds} is the dataset field description and introduction.

[0039] Step 3) Feature semantic generation.

[0040] Based on the target task (such as click-through rate prediction, ranking optimization, etc.) and the sequence feature type, a second text (Prompt 2) is constructed using a static second custom code generation template. Then, the dataset description and the second text (Prompt 2) are input together into the Large Language Model (LLM) described in step 2 to generate an enhanced text feature description for the user behavior sequence (such as recent user behavior patterns, changes in interests across time periods, etc.), i.e., feature semantics.

[0041] The second text (Prompt 2) constructed from the second custom code generation template includes at least the following segments: "Dataset Description," "Prediction Target," "Input Field Content," "Input Field Format," "Baseline Performance of the Default Model," "Previously Generated Features and Their Evaluation Results," "Generation Requirements," and "Output Content." The dataset description is filled in the “Dataset Description” section; The “prediction target” section contains the object to be predicted by the recommendation system model; The “Input Field Content” section contains the name, sample, and brief description of the dataset fields; The “Input Field Format” section specifies the format for data input; The "Baseline performance of the default model" section contains existing baseline detection values; The phrase “previously generated features and their evaluation results” contains the rounds of previously generated sequence features and the corresponding detection results. If the detection result is determined to be a valid sequence feature, then the corresponding detection result is input; if the detection result is determined not to be a valid sequence feature, then the judgment of non-sequence feature is input. The “Generation Requirements” section specifies the generation directions to be avoided and the generation directions to be attempted. The generation directions to be avoided include: avoiding copying sequence feature names and descriptions generated in previous rounds, and avoiding relying on patterns that performed poorly in previous rounds; The proposed directions for generation include: exploring new combinatorial features driven by behavior, time, or diversity, and exploring creative combinations of existing columns. The “output content” segment contains the feature name, feature description, and feature type.

[0042] See Figure 3 As shown, Figure 3 The template for the second text (Prompt2) is shown, where {dataset_ds} is the dataset description, {target_var} is the model prediction object, {df_columns} is the name, sampling and description of the dataset field, {base_line_value} is the existing baseline detection value, {i+1} is the i-th generation round of sequence features, {cache} is the judgment of the detection results generated in the previous rounds. If it is a valid sequence feature, the corresponding detection result is input; otherwise, the judgment of non-sequence features is input, and {list(existing_features)} is the name and description of the sequence features generated in the previous rounds.

[0043] Step 4) Collect new sequence feature descriptions.

[0044] If the semantic features generated by the Large Language Model (LLM) conform to the sequence feature definition required to construct the third text (Prompt 3), they are used for the construction of the third text (Prompt 3) in the next step. If they do not conform to the sequence feature definition required to construct the third text (Prompt 3), the construction is considered to have failed, and the text is returned to the cache for parsing and filtering. The semantic features that conform to the task objectives and semantic constraints are retained and recorded in the local feature enhancement document for subsequent construction of the third text (Prompt 3).

[0045] Step 5) Feature code generation.

[0046] Based on the newly added semantic features, a third text (Prompt 3) is constructed using a static third custom code generation template. Then, the third text (Prompt 3) is input into the Large Language Model (LLM) in step 2 to generate the source code that implements the feature extraction, i.e., the feature code.

[0047] The third text (Prompt 3) constructed by the third custom code generation module includes at least the following segments: "Description of New Sequence Features", "Prediction Target", and "Input Field Format"; wherein, The “Description of New Sequence Features” section contains the “Output Content” from the previous round; The “prediction target” section contains the object to be predicted by the recommendation system model; The “Input Field Format” section specifies the format for data input.

[0048] See Figure 4 As shown, Figure 4 The template for the third text (Prompt3) is shown, where {feature_desc} is a description of the newly added sequence features, {target_var} is the prediction target of the prediction model, and {df_columns} is a brief description of the existing fields in the dataset.

[0049] Step 6) Feature code integration.

[0050] The confirmed feature codes are embedded into the prediction model training environment of the Aomi recommendation dataset, combining them with the original data processing and feature engineering to achieve the automatic generation and loading of new features in the training data.

[0051] Step 7) Recommendation system model experiment and performance feedback.

[0052] Within the existing predictive model framework of the Aomi recommendation system, we conduct new feature experiments (ablation experiments). The new prediction metrics (detection index data) such as accuracy, recall, and AUC obtained from the predictive model are input into the cache. Based on this, we evaluate and generate a new round of sequence features.

[0053] Alternatives to model types: In addition to using large language models (LLM) to generate sequence features, other models with natural language understanding and generation capabilities can be used, such as multimodal large models (combining text and structured data input) and dedicated sequence modeling neural networks (such as Transformer, BERT, T5, etc.) to achieve semantic understanding and feature generation of data fields.

[0054] Alternatives to Prompt generation: In addition to using static custom code generation templates to generate Prompts, dynamic Prompt generation techniques can also be employed. These techniques automatically optimize Prompt content based on dataset features, such as by using meta-learning or rule-based Prompt adjustment mechanisms, thereby improving the relevance and executability of the model-generated results.

[0055] Alternatives to code generation and integration methods: In addition to directly outputting executable code through a large language model, it can also be generated using an intermediate language (such as pseudocode or process script) and then converted into target language code by a code translator; new feature generation logic can be integrated into a low-code / no-code platform without the need for manual writing or adjustment of the underlying code.

[0056] Alternatives to the verification mechanism: In addition to ablation experiments in the Aomi model, performance can also be verified in other recommendation models, time series models, or multi-task learning models.

[0057] The verification method can also be replaced by an automated model evaluation pipeline, including online A / B testing, cross-validation, or simulated user behavior experiments.

[0058] Alternatives to input data processing methods: In addition to using field names and example values ​​directly, data profiling, statistical feature descriptions, or distribution information can be used as model inputs to generate sequence features, further enhancing adaptability when original samples are lacking.

[0059] In summary, this invention proposes a sequence feature generation and enhancement method based on Large Language Model (LLM). This method addresses the shortcomings of existing feature generation techniques, which mainly focus on discrete features and lack research on sequence feature generation. By constructing natural language descriptions for dataset field information, combining custom Prompt templates, and implementing a multi-round generation and verification process, it achieves automated extraction, description, and code generation of potential sequence features from the original data, and efficiently integrates and verifies them within existing model environments.

[0060] The main technical improvements of this invention include the following key points: 1. Semantic Data Processing: This invention describes structured data containing field names and example values ​​in natural language, enabling the model to understand the contextual meaning of the fields and providing a semantic foundation for subsequent sequence feature generation.

[0061] 2. Custom Prompt-driven multi-stage generation: This invention mainly designs two types of custom Prompt templates, which are used to generate natural language descriptions of sequence features and corresponding executable code, and form a multi-round iterative mechanism to improve the accuracy and usability of the generated results.

[0062] 3. Automatic embedding and verification of feature code: This invention directly embeds the generated sequence feature processing code into the target model environment (such as the Aumi model), and verifies the contribution of the newly added sequence features to the model performance through ablation experiments, thereby realizing an automated feature engineering closed loop.

[0063] In summary, through the above-mentioned key technologies, this invention can significantly expand the applicability of existing feature generation methods in sequence data processing and improve the feature representation ability and prediction performance of the model.

[0064] The present invention also provides a computer device, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform the operation corresponding to the above-described method for enhancing user behavior sequence features in a recommendation system based on a large language model.

[0065] The present invention also provides a computer storage medium storing at least one executable instruction, which causes a processor to perform the operation corresponding to the above-described method for enhancing user behavior sequence features in a recommendation system based on a large language model.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for enhancing user behavior sequence features in a recommender system based on a large language model, characterized in that, Includes the following steps: Step 1) Obtain the recommendation dataset; Step 2) Based on the purpose and common fields of the recommended dataset in the recommendation system, the feature structure of the recommended dataset is parsed, the first text is automatically constructed using the first custom code generation template, and then the first text is input into the large language model to generate a natural language description related to the recommended dataset, i.e., the dataset description, which is used to provide a semantic basis for subsequent feature generation. Step 3) Based on the target task and sequence feature type, the second text is automatically constructed using the second custom code generation template. Then, the dataset description and the second text are input into the large language model to generate an enhanced text feature description for the user behavior sequence, i.e., feature semantics. Step 4) Based on the newly added semantic features, a third text is automatically constructed using a third custom code generation template. Then, the third text is input into the large language model to generate the source code that implements the feature extraction, i.e., the feature code. Step 5) Embed the confirmed feature code into the prediction model training environment of the recommendation dataset, so as to combine it with the original data processing and feature engineering to realize the automatic generation and loading of new features in the training data; Step 6) In the existing prediction model framework of the recommendation system, conduct experiments on new features, input the prediction results obtained by the prediction model, including the new precision, recall, and AUC, into the cache, and evaluate and generate a new round of sequence features based on this.

2. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, Other models with natural language understanding and generation capabilities, including multimodal large models and dedicated sequence modeling neural networks, are used to replace the large language model to achieve semantic understanding and sequence feature generation of data fields.

3. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, The first text, the second text, and the third text are all generated using dynamic Prompt generation technology, which automatically optimizes the Prompt content based on the characteristics of the dataset and uses meta-learning or rule-based Prompt adjustment mechanisms to improve the relevance and executability of the model-generated results.

4. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, The purposes and common fields include user identifier, item identifier, behavior type, timestamp, and contextual attributes.

5. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, The first text constructed from the first custom code generation template contains at least the following segments: "task topic", "input content", and "task requirements"; wherein, The "task topic" section includes sections on "content described", "method of description", and "purpose of description"; The meaning of the "content of description" paragraph is: generate a description of the input recommendation dataset and its various fields; The meaning of the "method of description" paragraph is: the description is presented in a detailed and natural language manner; The meaning of the "purpose of description" paragraph is that description should serve as a semantic basis for feature engineering in the recommender system; The "Input Content" field contains the dataset name, a list of domain names, and their example values; The "Task Requirements" section includes sections for "Overall Dataset Description Requirements", "Description Requirements for Each Field of the Dataset", "Description Statement Requirements", "Example Value Requirements", and "Applicability Requirements". The meaning of the "overall description requirements for the dataset" paragraph is: based on the name and field content of the input recommended dataset, describe its overall purpose or domain; The phrase "requirements for describing each field of the dataset" means that for each field of the input recommendation dataset, describe its meaning, its potential role in model building, and the type of information it contains. The meaning of the "requirements for descriptive statements" section is: descriptions should use clear, formal, and complete statements; The meaning of the "example value requirement" paragraph is: avoid repeating precise example values ​​unless it is necessary to explain their meaning; The meaning of the "applicability requirement" paragraph is: to make the description universal, so that it applies to a dataset rather than a single record.

6. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, The second text constructed from the second custom code generation template shall include at least the following segments: "Dataset Description", "Prediction Target", "Input Field Content", "Input Field Format", "Baseline Performance of the Default Model", "Previously Generated Features and Their Evaluation Results", "Generation Requirements", and "Output Content"; wherein, The dataset description is filled in the "Dataset Description" section; The "prediction target" field contains the object to be predicted by the recommendation system model. The "Input Field Content" section contains the name, sampling, and brief description of the dataset fields; The "Input Field Format" section specifies the format for data input. The "Baseline performance of the default model" section contains existing baseline detection values; The phrase "previously generated features and their evaluation results" includes the rounds of previously generated sequence features and the corresponding detection results. If the detection result is determined to be a valid sequence feature, then the corresponding detection result is input; if the detection result is determined not to be a valid sequence feature, then the judgment of non-sequence feature is input. The "Generation Requirements" section specifies the generation directions to be avoided and the generation directions to be attempted. The generation directions to be avoided include: avoiding copying sequence feature names and descriptions generated in previous rounds, and avoiding relying on patterns that performed poorly in previous rounds; The proposed directions for generation include: exploring new combinatorial features driven by behavior, time, or diversity, and exploring creative combinations of existing columns. The "Output Content" field contains the feature name, feature description, and feature type.

7. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, The third text, constructed from the third custom code generation template, contains at least the following segments: "Description of New Sequence Features," "Prediction Target," and "Input Field Format." The "Description of New Sequence Features" section contains the "Output Content" from the previous round; The "prediction target" field contains the object to be predicted by the recommendation system model. The "Input Field Format" section specifies the format for data input.

8. The method for enhancing user behavior sequence features in a recommender system based on a large language model according to claim 1, characterized in that, If the semantic features generated by the large language model conform to the sequence feature definition required to construct the third text, they are used for the construction of the third text in the next step. If they do not conform to the sequence feature definition required to construct the third text, the construction is considered to have failed, and the text is returned to the cache for parsing and filtering. The semantic features that conform to the task objectives and semantic constraints are retained and recorded in the local feature enhancement document for subsequent construction of the third text.

9. A computer device, characterized in that, include: The system includes a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface communicate with each other via the communication bus. The memory stores at least one executable instruction, which causes the processor to perform the operation corresponding to the user behavior sequence feature enhancement method for recommendation systems based on large language models as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer-readable storage medium stores at least one executable instruction that causes a processor to perform the operation corresponding to the method for enhancing user behavior sequence features of a recommendation system based on a large language model as described in any one of claims 1-7.

Images