Construction method and device for user data modeling

Abstract

The invention provides a construction method and device for user data modeling, and relates to the technical field of artificial intelligence. The construction method for user data modeling comprises the following steps: generating user behavior snapshot information according to user original information; generating model input data according to the user behavior snapshot information; training a pre-trained data generation model according to a pre-constructed supervision fine tuning data set to obtain a user data modeling reasoning model; the model input data and a predefined modeling reasoning cue word are input into the user data modeling reasoning model, a user data modeling reasoning result is obtained, and the modeling reasoning cue word comprises model role positioning information, a reasoning behavior rule and a task execution logic sequence. Based on the world knowledge and context reasoning ability of the model, end-to-end joint reasoning of multi-dimensional modeling is achieved, the overall process is low in manual dependence degree, cross-dimensional reasoning can be achieved only through one model, and expansibility is good.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a method and apparatus for constructing user data modeling. Background Technology

[0002] Currently, B (Business) category user data modeling tags are core data assets for B2B (Business-to-Business) e-commerce platforms, supporting user segmentation management, precision marketing, and intelligent traffic distribution. However, the existing user data modeling system faces some deep-seated challenges: Modeling user data of type B relies on diverse data sources, including business registration information (images / text), transaction records (structured text), inquiry dialogue logs (multi-turn natural language), image search images (visual data), and search query sequences (short text). Traditional methods require expert experience for feature engineering design, which makes it difficult to effectively process unstructured data (such as images and dialogues), leading to fragmentation of information across different modalities, semantic loss, and ultimately resulting in a fragmented and low-coverage user data modeling system.

[0003] Besides the reliance on manual labor due to difficulties in data integration, existing user data modeling systems are also limited by manually pre-defined tagging systems. These systems mostly use static rules or shallow models to generate tags, failing to capture the subtle differences among B-type users in terms of industry segmentation, business models, and procurement strategies. Furthermore, rule updates lag behind market changes, easily creating "information cocoons," resulting in crude, incomplete, and untimely tags that are difficult to support refined management.

[0004] Furthermore, each modeling dimension typically requires independent modeling (such as procurement preference models, supplier relationship models, etc.), leading to an explosion in the number of models and an exponential increase in training and deployment costs. In addition, the lack of knowledge-sharing mechanisms between models makes it difficult to achieve cross-dimensional knowledge transfer and collaborative reasoning.

[0005] In summary, existing technologies suffer from high dependence on manual labor and poor scalability. Summary of the Invention

[0006] Based on this, this application provides a method and apparatus for constructing user data modeling, which enables the construction of user data modeling with low dependence on manual intervention and good scalability.

[0007] According to one aspect of this application, a method for constructing user data modeling is proposed, comprising: generating user behavior snapshot information based on original user information, wherein the original user information includes the user's original information within the target platform and pre-selected additional information; generating model input data based on the user behavior snapshot information; training a pre-trained data generation model based on a pre-constructed supervised fine-tuning dataset to obtain a user data modeling inference model, wherein the supervised fine-tuning dataset includes multiple input samples labeled with modeling labels, the input samples include user behavior samples and their contextual background samples, and the modeling labels include modeling inference target factors; inputting the model input data and predefined modeling inference prompts into the user data modeling inference model to obtain user data modeling inference results, wherein the modeling inference prompts include model role positioning information, inference behavior rules, and task execution logic order.

[0008] According to some embodiments, user behavior snapshot information is generated based on the user's original information, including: extracting the user's original behavior logs on the target platform, and extracting relevant content of preset indicators from the original behavior logs as preset indicator information; slicing the preset indicator information according to the flow cycle related to the preset indicators to obtain multi-granularity time slices; extracting the first preset number of data from the multi-granularity time slices to form single-time-window aggregation information; extracting information from the original behavior logs whose repetition period is greater than a preset period threshold to obtain long-cycle behavior information; arranging the long-cycle behavior information in chronological order as time-series modeling information; merging data with multiple facts and a correlation degree that meets preset correlation conditions in the original behavior logs according to preset association relationships to obtain multi-fact aggregation information; extracting and fusing additional information pre-selected by the user as dimensional correlation enhancement information, wherein the pre-selected additional information includes identity information, asset information, and / or preset behavior information; and obtaining user behavior snapshot information based on single-time-window aggregation information, time-series modeling information, multi-fact aggregation information, and dimensional correlation enhancement information.

[0009] According to some embodiments, generating model input data based on user behavior snapshot information includes: compressing and merging the user behavior snapshot information to construct an original wide data table; grouping the original wide data table according to preset dimensions, wherein the preset dimensions include basic information, behavior information, transaction information and / or additional information; constructing a user context description template based on the grouping results; and filling in information based on the user context description template according to the user behavior snapshot information to obtain model input data.

[0010] According to some embodiments, a user data modeling inference model is obtained by training a pre-trained data generation model based on a pre-built supervised fine-tuning dataset, including: constructing multiple base learners based on the pre-trained data generation model, wherein the multiple base learners correspond one-to-one with multiple pre-built thought chain inference links; training the multiple base learners independently according to the corresponding thought chain inference links based on the supervised fine-tuning dataset; and using the trained multiple base learners as the user data modeling inference model.

[0011] According to some embodiments, inputting model input data and predefined modeling and reasoning prompts into a user data modeling and reasoning model to obtain user data modeling and reasoning results includes: inputting model input data and predefined modeling and reasoning prompts into multiple base learners trained in the user data modeling and reasoning model to obtain multiple reasoning results; and performing voting aggregation based on the multiple reasoning results to obtain the user data modeling and reasoning results.

[0012] According to some embodiments, the supervised fine-tuning dataset is pre-constructed through the following steps: generating standardized semantic descriptions for pre-constructed modeling and inference target factors, and defining model training prompts based on the standardized semantic descriptions, wherein the standardized semantic descriptions include dimension names, definition descriptions, judgment criteria, typical positive and negative examples, and / or inference rule templates; acquiring user behavior samples and their contextual background samples from multiple users to construct input samples; inputting the input samples and model training prompts into a pre-trained data processing model to obtain preliminary annotation results; performing multiple independent annotations on the preliminary annotation results based on preset expert experience information; comparing the results of multiple independent annotations and determining modeling labels based on the comparison results; and constructing the supervised fine-tuning dataset based on the modeling labels, the inference process of the model inference preliminary annotation results, and the input samples.

[0013] According to some embodiments, the method further includes: constructing a preference dataset, wherein the preference dataset includes multiple sets of preference pairs, the preference pairs including model input, preference model output, and non-preference model output; and using a direct preference optimization (DPO) strategy to iteratively optimize the user data modeling inference model based on the preference dataset.

[0014] According to some embodiments, the method further includes: visualizing and / or analyzing the results of user data modeling and inference; Personalized searches are performed based on the results of user data modeling and inference; and / or user raw information is processed based on the results of user data modeling and inference.

[0015] According to some embodiments, the visualization and / or user analysis of user data modeling and inference results include: visualizing user data modeling and inference results based on a preset structured style; generating display images based on user data modeling and inference results using a data generation model; and / or displaying user data modeling and inference results to users based on a preset real-time dialogue model.

[0016] According to some embodiments, personalized search is performed based on the results of user data modeling and inference, including: forming a user data modeling dataset based on the results of user data modeling and inference; matching in the user data modeling dataset according to the search intent; and obtaining relevant user behavior snapshot information based on the matching results.

[0017] According to some embodiments, personalized search is performed based on user data modeling and inference results, including: based on search intent, enhanced retrieval is performed based on user original information to obtain information related to the search intent from user data modeling and inference results corresponding to the user original information.

[0018] According to some embodiments, the method further includes: constructing a user data modeling backend interface based on the user data modeling inference model, so that external platforms can call the user data modeling inference model through the user data modeling backend interface to generate user data modeling inference results; and / or converting the user data modeling inference results into structured language to generate characteristic population descriptions, so that external platforms can configure according to the characteristic population descriptions.

[0019] According to one aspect of this application, a user data modeling construction apparatus includes: a behavior information module for generating user behavior snapshot information based on original user information, wherein the original user information includes the user's original information within a target platform and pre-selected additional information; an input data module for generating model input data based on the user behavior snapshot information; a model training module for training a pre-trained data generation model based on a pre-constructed supervised fine-tuning dataset to obtain a user data modeling inference model, wherein the supervised fine-tuning dataset includes multiple input samples labeled with modeling tags, the input samples include user behavior samples and their contextual background samples, and the modeling tags include modeling inference target factors; and a model inference module for inputting the model input data and predefined modeling inference prompts into the user data modeling inference model to obtain user data modeling inference results, wherein the modeling inference prompts include model role positioning information, inference behavior rules, and task execution logic order.

[0020] According to one aspect of this application, an electronic device is provided, comprising: one or more processors; a storage device for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the method as described above.

[0021] According to one aspect of this application, a computer-readable medium is provided that stores a computer program or instructions thereon, which, when executed by a processor, implement the method as described above.

[0022] Through the embodiments provided in this application, multi-source heterogeneous data stored by users in multiple sources are integrated to generate user behavior snapshot information of various modalities, thereby forming model input data and providing effective data support for model inference. During training, the output of the data generation model is optimized through a pre-built supervised fine-tuning dataset to obtain a user data modeling and inference model. During inference, the model input data and predefined modeling and inference prompts are input into the user data modeling and inference model. Based on the model's world knowledge and contextual reasoning ability, end-to-end joint inference of multi-dimensional modeling is realized. The overall process has low dependence on manual intervention, and only one model is needed to realize cross-dimensional inference, with good scalability. Attached Figure Description

[0023] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application.

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings, without exceeding the scope of protection claimed by this application.

[0025] Figure 1 A flowchart illustrating the method for constructing user data modeling as provided in the embodiments of this application; Figure 2 A flowchart for generating user behavior snapshot information based on original user information is provided in this application embodiment; Figure 3 A flowchart for generating model input data based on user behavior snapshot information provided in this application embodiment; Figure 4 A flowchart of the supervised fine-tuning dataset construction steps provided in the embodiments of this application; Figure 5 A block diagram of a user data modeling construction apparatus provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0028] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0029] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0030] It should be understood that although the terms first, second, third, etc., may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of this application. As used herein, the term "and / or" includes all combinations of any one and more of the associated listed items.

[0031] For specific implementation details, please refer to the following examples.

[0032] Figure 1 A flowchart illustrating the method for constructing user data modeling as provided in this application embodiment. Figure 1 As shown, the method includes steps S110-S140.

[0033] In step S110, user behavior snapshot information is generated based on the user's original information, wherein the user's original information includes the user's original information within the target platform and pre-selected additional information.

[0034] Traditional data warehouses primarily rely on structured and dimensional modeling to serve BI (Business Intelligence) analysis and rule engines, but they struggle to meet the demands of data generation models for rich context, semantic coherence, and behavioral serialization. This application utilizes a data modeling method of "semantic compression + context reconstruction" to transform raw user information into data that the data generation model can understand, forming a "user behavior snapshot."

[0035] It should be noted that the transformation process includes the collection, compression, reconstruction, and aggregation of user data from multiple sources. The user behavior snapshot information generated by the transformation is essentially a comprehensive user feature repository oriented towards a data generation model.

[0036] User behavior snapshot information can be described in natural language or in other forms (such as encoded sequences), and this application does not impose any restrictions on this.

[0037] User raw information comes from multiple sources, including raw information from the target platform and pre-selected supplementary information. The target platform is the primary source of user raw information and the application scenario for building user data models. In other words, user raw information also includes pre-selected supplementary information. This pre-selected supplementary information includes other information that can be legally obtained from outside the target platform.

[0038] Furthermore, the pre-selected additional information may also include non-platform data, such as the user's original compliance information in business registration, public opinion, and industry databases.

[0039] In some embodiments, non-platform data also includes relevant organizational data, relevant organizational assets, spatiotemporal environmental data (such as regional regulations, seasonal solar terms, trend hotspots, etc.).

[0040] It should be emphasized that the original user information covers multiple modalities such as user asset information, user spatiotemporal behavior trajectory, user interaction logs, user transaction links, and user image content, and is multi-source heterogeneous data.

[0041] User behavior snapshots constructed using original user information can provide effective data support for model inference.

[0042] In step S120, model input data is generated based on user behavior snapshot information.

[0043] Based on user behavior snapshot information, feature aggregation and format processing are performed as input data for the model.

[0044] In step S130, the pre-trained data generation model is trained based on the pre-constructed supervised fine-tuning dataset to obtain a user data modeling and inference model. The supervised fine-tuning dataset includes multiple input samples labeled with modeling labels. The input samples include user behavior samples and their contextual background samples. The modeling labels include modeling and inference target factors.

[0045] To ensure the accuracy of the model's inference in a specific domain (i.e., the domain corresponding to the target platform), a high-quality supervised fine-tuning (SFT) dataset was constructed. The data generation model was fine-tuned and trained using the supervised fine-tuning dataset to obtain a user data modeling inference model that can accurately infer user data modeling in a specific domain.

[0046] The supervised fine-tuning dataset includes multiple high-quality input samples. Each input sample includes: model input: user behavior context (i.e., user behavior samples and their context background samples) and model output: structured modeling results (i.e., modeling labels).

[0047] Furthermore, according to the example embodiment, the model output in each input sample also includes: a natural language reasoning process (including supporting / opposing evidence).

[0048] To standardize the inference process of user data modeling, the supervised fine-tuning dataset needs to be constructed according to a defined standard inference paradigm. Specifically, the model output for each input sample is obtained based on its model input, using predefined modeling inference factors and their corresponding modeling inference rules. In other words, the modeling label for each input sample includes its corresponding modeling inference factors.

[0049] The input sample can be constructed by technicians rewriting existing data, or it can be reconstructed by collecting data.

[0050] Understandably, the construction of the supervised fine-tuning dataset itself involves building a multi-dimensional evaluation system for the input samples. According to the example implementation, this multi-dimensional evaluation system includes assessments of data integrity, behavioral representativeness, and label accuracy.

[0051] Furthermore, to accelerate the construction of supervised fine-tuning datasets and thus improve model training efficiency, an automated and human-coordinated input sample screening pipeline can be established. Specifically, by combining manual annotation with an online traffic verification feedback mechanism, high-quality input samples that are highly aligned with "user behavior—data modeling—potential needs" can be continuously collected and accumulated, forming a high signal-to-noise ratio training data closed loop.

[0052] In terms of model training, a pre-trained deep learning-based data generation model is used as the foundation. In the technical solution provided in this application, the data generation model can be a deep learning model with a relatively large number of model parameters. However, this application does not limit the number of model parameters supported by the deep learning model used, aiming to meet actual needs. The deep learning model involved in this application can be an artificial intelligence-based language model (LM) or a multimodal model (MM).

[0053] Based on the selected base model, domain-adaptive fine-tuning is performed using a supervised fine-tuning dataset. This application does not restrict the specific techniques used for fine-tuning training.

[0054] According to the example implementation, LoRA (Low-Rank Adaptation) technology is used to achieve efficient parameter fine-tuning, which reduces computational overhead while retaining the generalization ability of the base model.

[0055] In step S140, the model input data and predefined modeling and reasoning prompts are input into the user data modeling and reasoning model to obtain the user data modeling and reasoning result. The modeling and reasoning prompts include model role positioning information, reasoning behavior rules, and task execution logic order.

[0056] This application utilizes the macroeconomic knowledge, industry common sense, business logic, and causal reasoning capabilities inherent in the data generation model to design a prompt-driven multi-task joint reasoning framework, achieving reasoning alignment from user "behavioral snapshots" to "data modeling".

[0057] Specifically, modeling and reasoning prompts that conform to the modeling tasks of this application are designed. In the specific implementation process, a three-element structured modeling method of "role-rule-execution order" is adopted to clarify the role positioning of the model in the reasoning task (i.e., model role positioning information), the behavioral rules of the reasoning model (i.e., reasoning behavior rules), and the logical order of task execution, so as to ensure that the model output has interpretability and consistency.

[0058] According to the example implementation, the model role positioning information is "modeling analyst", and the behavior rule of the inference model is "to make inferences based on data facts".

[0059] It should be noted that the task execution logic order refers to the specific workflow of the model in executing tasks.

[0060] According to the example embodiment, the task execution logic order can be based on the CoT (Chain-of-Thought) reasoning mechanism. It should be explained that the CoT reasoning mechanism decomposes the user data modeling inference task into three progressive sub-steps: "data basis → reasoning rationale → tag generation." This embodiment of the application solves the problem of "black box reasoning" or leapfrogging judgments that easily occur when data generation models generate user data modeling tags, leading to uncontrollable and unreliable output results, by combining structured prompt engineering with a chain-based reasoning mechanism. This improves the model's reasoning ability and output stability in complex scenarios.

[0061] In the specific reasoning process, guided by modeling and reasoning prompts, the model first extracts key factual evidence from observable user behavior data in the model input data. Then, it combines industry knowledge and contextual information to perform causal reasoning, ultimately generating structured user tags as the result of user data modeling and reasoning. This application embodiment integrates multi-source heterogeneous data stored by users in multiple sources to generate user behavior snapshot information in various modalities, thereby forming model input data and providing effective data support for model inference. During training, the output of the data generation model is optimized through a pre-built supervised fine-tuning dataset to obtain a user data modeling and inference model. During inference, the model input data and predefined modeling and inference prompts are input into the user data modeling and inference model. Based on the model's world knowledge and contextual reasoning ability, end-to-end joint inference of multi-dimensional modeling is realized. The overall process has low dependence on manual intervention, and only one model is needed to achieve cross-dimensional inference, with good scalability.

[0062] According to some embodiments, refer to Figure 2 In step S110, user behavior snapshot information is generated based on the user's original information, which can be achieved through steps S210-S280.

[0063] In step S210, the user's original behavior logs on the target platform are extracted, and relevant content of preset indicators is extracted from the original behavior logs as preset indicator information.

[0064] Raw behavior logs are the on-site behavior data of the target platform, including PC (computer) search / image search / factory search behavior, collection / add to cart / transaction behavior, etc., as well as APP (mobile) communication behavior, sales / financial behavior, etc.

[0065] According to the example implementation, the preset indicators include user purchase amount, number of purchase orders, purchase category, and new product categories launched in the store.

[0066] Extract the corresponding content of the preset metrics from the target platform's original behavior logs and use it as the preset metric information.

[0067] In step S220, the preset index information is sliced ​​according to the flow cycle associated with the preset index to obtain multi-granularity time slices.

[0068] Based on preset indicators and combined with the different business cycle (i.e. turnover cycle) of the corresponding users, multi-granularity time slices are constructed.

[0069] According to the example embodiment, the granularity of the multi-granularity time slice is, for example, the past 30 days, the past 6 months, the past 1 year, etc.

[0070] In step S230, the first preset number of data points are extracted from the multi-granularity time slices to form a single time window aggregated information.

[0071] Based on the constructed multi-granularity time slices, the top (highest) data of the text data is retained and merged into statistical data, which is recorded as single-time-window aggregated information.

[0072] It should be emphasized that the preset number of items to be retained can be set according to the actual situation, and this application does not impose any restrictions on this.

[0073] In step S240, information with a repetition period greater than a preset period threshold is extracted from the original behavior log to obtain long-cycle behavior information.

[0074] Based on the original behavior logs, extract users' long-term behavior.

[0075] In this embodiment, a preset period threshold is used to distinguish long-cycle behaviors; that is, behaviors with a repetition period greater than the preset period threshold are considered long-cycle behaviors, and the corresponding information is extracted and recorded as long-cycle behavior information. The preset period threshold can be set according to the specific circumstances of each platform, and this application does not impose any restrictions on it.

[0076] It needs to be explained that long-term behaviors include user purchasing behaviors such as buying, communicating, searching, and selling, as well as search and management behaviors.

[0077] In step S250, the long-cycle behavior information is arranged in chronological order and used as time-series modeling information.

[0078] Long-term user behavior is preserved in chronological order as a sequence, serving as temporal modeling information. This temporal modeling information can be used to infer user behavior cycles and determine decaying behaviors.

[0079] According to the example embodiment, long-cycle behavioral information includes information related to new product launch behavior, procurement behavior, etc. Based on this, the construction rules for time-series modeling information can be defined according to the actual situation, which will not be elaborated here.

[0080] In step S260, data from the original behavior log that contain multiple facts and whose correlation meets the preset correlation conditions are merged according to the preset correlation relationship, and are used as multi-fact aggregated information.

[0081] This step essentially aims to merge data that is multifactual but highly correlated. "Multifactual" refers to data consisting of multiple different variables or indicators (e.g., "top purchase amount," "primary category," "price segment preference"). "Highly correlated" means that there is a stable and measurable relationship between the changing trends of the multiple variables or indicators.

[0082] In practice, the category tree system set by the target platform can be used as the basis for determining the degree of association (relevance). Based on this, the preset association conditions can be the mapping relationship between the category tree system and the association relationship; or they can be the threshold of the set quantitative indicators, that is, using indicators to quantify the relevance (such as the correlation coefficient), and setting thresholds constitutes the preset association conditions. This application does not impose too many restrictions on this.

[0083] By compressing and merging multiple facts that are highly relevant, the understanding of the data generation model can be improved.

[0084] During the merging process, one of the multiple variables or indicators is selected for quantification.

[0085] The embodiments of this application replace the traditional method of manually setting thresholds, and the model can learn high, medium and low thresholds from the data itself.

[0086] In step S270, the user-selected additional information is extracted and fused as dimensional association enhancement information. The pre-selected additional information includes identity information, asset information and / or preset behavior information.

[0087] According to the example embodiment, the preset behavioral information of the pre-selected additional information includes: product category, store ranking, fan modeling, etc. Asset information includes store name, etc.

[0088] Furthermore, the pre-selected additional information includes pre-defined behavioral information such as cross-border high-confidence behavior.

[0089] Extract and merge the relevant data from the corresponding platforms and / or regions.

[0090] Furthermore, in some embodiments, in addition to the additional information pre-selected by the user, the data to be merged may also include the identity information filled in by the user during the identity verification process.

[0091] This step is used to integrate additional user information and complete the context.

[0092] In step S280, user behavior snapshot information is obtained based on single-time-window aggregation information, time-series modeling information, multi-fact aggregation information, and dimensional correlation enhancement information.

[0093] Single-time-window aggregated information, time-series modeling information, multi-fact aggregated information, and dimension-related enhanced information together constitute a user behavior snapshot, denoted as user behavior snapshot information.

[0094] The generated user behavior snapshot information is readable rich text, which can meet the needs of the data generation model.

[0095] According to some embodiments, in step S120, referring to Figure 3 Based on the user behavior snapshot information, model input data is generated, which can be achieved through steps S310-S340.

[0096] In step S310, the user behavior snapshot information is compressed and merged to construct the original wide data table.

[0097] This application does not impose restrictions on the specific implementation algorithm of the compression and merging process; the algorithm can be selected according to the actual situation.

[0098] For example, compress and merge according to a custom format: perform composite splicing.

[0099] To save on labor costs, this step can be implemented using a deep learning model, and this application does not impose any restrictions on it.

[0100] In step S320, the original data wide table is grouped according to preset dimensions, wherein the preset dimensions include basic information, behavioral information, transaction information and / or additional information.

[0101] Based on the original wide table after compression and merging, it is grouped according to dimensions such as basic information, behavioral information, transaction information, and additional information.

[0102] This application does not restrict the specific information for each dimension, which can be determined according to the actual situation.

[0103] In step S330, a user context description template is constructed based on the grouping results.

[0104] The constructed user context description template serves as a unified format specification for the input of the data generation model.

[0105] According to the example embodiment, the user context description template is in a standardized JSON format.

[0106] In step S340, based on the user context description template, information is populated according to the user behavior snapshot information to obtain model input data.

[0107] This population step is essentially an aggregation of user feature information based on user behavior snapshots. After entity recognition, intent recognition, sequence mining, behavior clustering, association rule mining, heterogeneous data alignment, and feature cross-enhancement, the user behavior snapshots are aggregated to obtain basic user information, user behavior information, and user transaction information. This forms a comprehensive user feature repository for the data generation model, transforming multi-source heterogeneous data into structured natural language descriptions usable by the model, containing all user information in a minimal token. This solves the problem of difficulty in task modeling caused by the complexity of understanding user business models, transforming complex business models into a task structure that the data generation model can understand and reason about.

[0108] The generated data is input into the data generation model for subsequent inference.

[0109] According to some embodiments, in step S130, the pre-trained data generation model is trained based on the pre-built supervised fine-tuning dataset to obtain the user data modeling inference model, which can be specifically implemented through steps S131-S133.

[0110] In step S131, multiple base learners are constructed based on the pre-trained data generation model, wherein each base learner corresponds one-to-one with a pre-constructed multiple thought chain inference link.

[0111] To further improve the stability and robustness of the reasoning results, this application introduces a majority voting mechanism to construct a multi-path reasoning chain, which corresponds one-to-one with multiple base learners.

[0112] In step S132, based on the supervised fine-tuning dataset, multiple base learners are trained independently according to their respective thought chain inference links.

[0113] In step S133, the trained base learners are used as user data modeling inference models.

[0114] Multiple base learners, once trained, form multiple independent inference paths in parallel.

[0115] The user data modeling and reasoning model constructed in this way can effectively reduce the randomness bias caused by a single reasoning path and enhance the consistency and credibility of the model's decisions.

[0116] According to some embodiments, in step S140, the model input data and predefined modeling and reasoning prompts are input into the user data modeling and reasoning model to obtain the user data modeling and reasoning result, which can be specifically achieved through steps S141-S142.

[0117] In step S141, the model input data and predefined modeling and reasoning prompts are input into multiple base learners trained in the user data modeling and reasoning model to obtain multiple reasoning results.

[0118] The model input data and modeling inference prompts are fed into multiple trained base learners. Multiple inference results are obtained through multiple parallel and independent inference paths formed by the multiple trained base learners.

[0119] In step S142, voting aggregation is performed based on multiple inference results to obtain the user data modeling inference result.

[0120] The user data modeling inference results are obtained by aggregating the tag outputs corresponding to multiple inference results through voting.

[0121] This application does not impose specific restrictions on the specific voting aggregation algorithm; the algorithm can be selected according to the actual situation.

[0122] This application's embodiments introduce a majority vote chain to enhance inference, solving the problem that inconsistencies or contradictions are easily generated by single thought chain inference due to the sparseness and high noise of B-type user behavior data.

[0123] According to some embodiments, refer to Figure 4 The supervised fine-tuning of the dataset is achieved through steps S410-S460.

[0124] In step S410, a standardized semantic description is generated for the pre-constructed modeling and reasoning target factors, and model training prompts are defined based on the standardized semantic description. The standardized semantic description includes dimension name, definition description, judgment basis, typical positive and negative examples and / or reasoning rule template.

[0125] Based on the established multidimensional modeling and reasoning target factors, standardized semantic descriptions are written for each factor, forming modeling and reasoning rules for the corresponding target factors, which serve as a standard reasoning paradigm. The standardized semantic descriptions include: dimension names, definitions, judgment criteria, typical positive and negative examples, and reasoning rule templates.

[0126] According to the example implementation, the inference rule template is as follows: If condition A is met, then the business model belongs to 'B'.

[0127] The model is trained with prompt words based on standardized semantic descriptions, which serves as a prompting engineering process to construct sample data.

[0128] It should be noted that the target factors for modeling and inference are constructed under data-driven conditions, in conjunction with the experience of experts from multiple departments of the target platform, including management, risk control, product, and technology departments.

[0129] According to the example implementation, driven by data, the platform combines the experience of experts from multiple departments to conduct demand alignment and expert review meetings, sort out the decision-making dependencies of B-type users in core scenarios such as user classification, precision marketing, service recommendation, risk warning, and growth strategy, integrate the objective behavioral characteristics of B-type users with the actual business scenario needs, and obtain the modeling and inference target factors after multiple rounds of iteration.

[0130] It is understandable that the modeling inference target factor is not only the output structure of model training, but also the link between the model's capabilities and value, playing multiple key roles: 1) Guiding multimodal data fusion and feature engineering. Specifically, each target factor corresponds to a set of cross-modal input signals, which in turn drive the organization of the data warehouse, realizing "starting with the end in mind" feature construction. Instead of relying on manual feature design based on human experience, it automatically mines feature combinations with strong relevance and high interpretability around the target factor.

[0131] 2) To ensure that the model output inference results are structured, interpretable, and visualized.

[0132] 3) Enhance cross-scenario reuse and strategy linkage capabilities. Specifically, a standardized factor system can serve as a common language for different modules, enabling one-time modeling and multiple calls, significantly reducing the cost of repetitive development.

[0133] 4) Supports dynamic evolution and continuous optimization. Specifically, the target factors themselves are semantically clear, making it easy to receive feedback from people through a natural language interface; forming a closed-loop evolutionary path of "input → factor adjustment → model iteration → effect feedback".

[0134] In step S420, user behavior samples and their contextual background samples from multiple users are obtained to construct input samples.

[0135] In step S430, the input samples and model training prompts are input into the pre-trained data processing model to obtain preliminary annotation results.

[0136] The data processing model was selected as the strong base model, and preliminary annotation results were generated based on the prompting engineering.

[0137] The data processing model used here may be the same as or different from the base model of the user data modeling inference model; this application does not impose any restrictions on this.

[0138] In step S440, the preliminary annotation results are annotated independently multiple times based on preset expert experience information.

[0139] A labeling team composed of management, data analysis, and other technical personnel conducted multiple independent labeling operations based on their expert experience and in accordance with standard reasoning paradigms.

[0140] In multiple independent annotations, each annotation can be done by different people to eliminate subjectivity.

[0141] In step S450, the results of multiple independent annotations are compared, and the modeling labels are determined based on the comparison results.

[0142] After multiple independent annotations and comparisons of differences, disputed samples are submitted to an expert panel for adjudication, thereby determining the modeling labels corresponding to the input samples.

[0143] In step S460, a supervised fine-tuning dataset is constructed based on the modeling labels, the reasoning process of the preliminary annotation results of the model reasoning, and the input samples.

[0144] Each sample in the supervised fine-tuning dataset contains: model input: input sample, i.e., user behavior context (textualized encapsulation), model output; structured modeling results (i.e. modeling labels) + natural language reasoning process (including supporting / opposing evidence).

[0145] This application constructs a high-quality sample screening pipeline, which involves manual annotation and traffic verification to collect high-quality user behavior, data modeling, and high-quality inference samples for potential needs. The model output is optimized on the basis of the data generation model, which greatly improves the accuracy of user business model and demand preference inference in business model inference tasks.

[0146] According to some embodiments, the method further includes steps S150-S160.

[0147] In step S150, a preference dataset is constructed, wherein the preference dataset includes multiple sets of preference pairs, and the preference pairs include model input, preference model output, and non-preference model output.

[0148] We introduce the Direct Preference Optimization (DPO) strategy, which uses human evaluation feedback to construct preference pair data, including model input, preference model output, and non-preference model output.

[0149] The preference pairs can be selected during model training, or they can be chosen from supervised fine-tuning datasets or other datasets, or they can be constructed directly; this application does not impose any restrictions on this.

[0150] In step S160, the Direct Preference Optimization (DPO) strategy is used to iteratively optimize the user data modeling inference model based on the preference dataset.

[0151] By adopting a joint optimization strategy of parameter fine-tuning (e.g., LoRA) + DPO, the alignment of the model output with expert judgment is further optimized, significantly improving the consistency between the user's business model understanding and demand preference inference results and human evaluation, and solving the problems of domain knowledge deficiency and preference understanding bias that still exist in the vertical scenario of B-type user data modeling.

[0152] According to some embodiments, the method further includes one or more of steps S170, S180, and S190.

[0153] In step S170, the user data modeling and reasoning results are visualized and / or analyzed.

[0154] Visualize the results of user data modeling and inference to improve data readability, enhance collaborative modeling applications, and increase annotation efficiency.

[0155] Based on the results of user data modeling and reasoning, user research reports and management suggestions can be further analyzed according to user needs.

[0156] According to an example embodiment, a user's needs may include, "What service needs and preferences do buyers with user data modeling A have?" In step S180, a personalized search is performed based on the modeling and reasoning results of the user data.

[0157] Based on the results of user data modeling and reasoning, further personalized searches are conducted according to the user's needs.

[0158] Personalized search includes general business reasoning and feature modeling reasoning.

[0159] In some embodiments, user data modeling can be queried based on search intent. That is, this embodiment is no longer limited to searching only user IDs (identifiers) and supports querying user data modeling with open-ended questions.

[0160] According to the example implementation, the search intent includes: "Search for users who want to open a women's clothing store with customization needs".

[0161] In step S190, the user's original information is processed based on the modeling and reasoning results of the user data.

[0162] Based on the results of user data modeling and reasoning, the original user information is processed (e.g., intelligent summarization) to make the original information more readable.

[0163] This application's embodiments improve data readability and enhance collaborative efficiency while expanding future application scenarios for user data modeling, becoming a set of data infrastructure capabilities for the target platform. This enables the construction of an interpretable and interactive user data modeling visualization system, achieving transparent presentation of the reasoning process and results.

[0164] According to some embodiments, in step S170, the user data modeling and reasoning results are visualized and / or analyzed, which can be achieved through one or more of steps S171, S172, and S173.

[0165] In step S171, the results of user data modeling and reasoning are visualized based on a preset structured style.

[0166] The preset structured styles include structured styles such as cards and components, which will not be listed exhaustively in this application.

[0167] The user data modeling and reasoning results are transformed into structured states such as cards and components, and the modeled text data is then displayed.

[0168] In step S172, a display image is generated based on the data generation model and the modeling and reasoning results of the user data.

[0169] The data generation model's ability to generate images is utilized to visually depict user data modeling and generate display images to showcase the results of user data modeling and reasoning.

[0170] In step S173, the user data modeling and reasoning results are displayed to the user based on a preset real-time dialogue model.

[0171] By productizing user data modeling tasks through two-way interaction using real-time dialogue models (such as chatbots), the results of user data modeling and reasoning are presented to users in a two-way interactive format, breaking the original framework of one-way interaction in data dashboards.

[0172] According to some embodiments, in step S180, a personalized search is performed based on the modeling and reasoning results of user data, which can be specifically implemented through steps S181-S183.

[0173] In step S181, a user data modeling dataset is formed based on the user data modeling inference results.

[0174] In step S182, a match is performed in the user data modeling dataset according to the search intent.

[0175] In step S183, relevant user behavior snapshot information is obtained based on the matching results.

[0176] Personalized search includes searching for people using natural language in a single sentence. In other words, artificial intelligence technology is used to identify the intent of search queries, and the results are matched against a user data model dataset to obtain corresponding snapshots of user behavior, enabling the retrieval of user groups based on underlying data modeling.

[0177] It should be explained that this step requires storing the inference results of the user data modeling to form a user data modeling dataset, which can then be used by users for personalized searches.

[0178] According to some embodiments, in step S180, a personalized search is performed based on the modeling and reasoning results of user data, which can be specifically implemented through step S184.

[0179] In step S184, based on the search intent, enhanced retrieval is performed according to the user's original information to obtain information related to the search intent from the user data modeling and reasoning results corresponding to the user's original information.

[0180] Personalized search includes RAG (Retrieval Augmentation Generation) custom inference.

[0181] In practice, the user's original information is used as RAG retrieval enhancement to obtain information related to the search intent.

[0182] According to the example implementation, the search intent includes: what membership benefits this user might need.

[0183] The embodiments of this application are no longer limited to offline inference modeling data, but support custom inference, meet the refined and personalized needs, and give more imagination and possibilities. According to some embodiments, the method further includes steps S101 and / or S102.

[0184] In step S101, a user data modeling backend interface is constructed based on the user data modeling inference model, so that external platforms can call the user data modeling inference model through the user data modeling backend interface to generate user data modeling inference results.

[0185] Build a user data modeling backend interface to connect external platforms with the model, and call the AI ​​user data modeling capability of the user data modeling inference results.

[0186] In step S102, the user data modeling and reasoning results are converted into structured language to generate a characteristic audience description, which is then configured by external platforms based on the characteristic audience description.

[0187] Based on the user data modeling and reasoning results, SQL (Structured Query Language) is automatically generated to produce characteristic audience descriptions. These characteristic audiences can then be imported into external platforms with one click, enabling rapid configuration of audience management strategies and improving efficiency.

[0188] Furthermore, in one specific embodiment, the user data modeling construction method provided in this application is applied to the target platform, and the model output is connected to downstream application scenarios such as search and push traffic side, merchant commercialization, and user product side of the target platform.

[0189] In this embodiment, an experiment was conducted on the matching strategy of B-type user data modeling on the search side of the APP and PC. By the end of the experiment, the performance of the key indicator GMV was observed, and both sides showed a positive increase. GMV stands for Gross Merchandise Volume, which represents the total transaction amount of goods.

[0190] In this embodiment, a product recall experiment was conducted in a search scenario, segmenting user groups based on B-type user data modeling. By the end of the experiment, increases were observed in direct-to-consumer GMV, direct-to-store GMV, UVL2O, and APP GMV. Here, UV stands for Unique Visitor, representing the number of different users visiting the website; L2O stands for Leads to Order.

[0191] In this embodiment, user groups were segmented based on B-type user data in the recommendation scenario, and a matching strategy was launched for experimentation. By the end of the experiment, GMV showed a positive increase over 14 days. The following describes an apparatus embodiment of this application, which can be used to perform the method embodiment of this application. For details not disclosed in the apparatus embodiment of this application, please refer to the method embodiment of this application.

[0192] Figure 5 A block diagram of a construction apparatus for user data modeling according to an exemplary embodiment is shown.

[0193] Figure 5 The apparatus shown can perform the aforementioned method for constructing user data modeling according to embodiments of this application.

[0194] like Figure 5 As shown, the apparatus for building user data modeling may include: See Figure 5Referring to the preceding description, the behavior information module 510 is used to generate user behavior snapshot information based on the user's original information, wherein the user's original information includes the user's original information within the target platform and pre-selected additional information.

[0195] The input data module 520 is used to generate model input data based on user behavior snapshot information.

[0196] The model training module 530 is used to train the pre-trained data generation model based on the pre-built supervised fine-tuning dataset to obtain the user data modeling and inference model. The supervised fine-tuning dataset includes multiple input samples labeled with modeling labels. The input samples include user behavior samples and their contextual background samples. The modeling labels include modeling and inference target factors.

[0197] The model reasoning module 540 is used to input model input data and predefined modeling reasoning prompts into the user data modeling reasoning model to obtain user data modeling reasoning results. The modeling reasoning prompts include model role positioning information, reasoning behavior rules, and task execution logic order.

[0198] The device performs functions similar to those described above; other functions are described in the preceding descriptions and will not be repeated here.

[0199] This application discloses an electronic device, including: a processor; and a memory storing a computer program, which, when executed by the processor, causes the processor to execute the above-described instruction generation method.

[0200] For example, refer to Figure 6 , Figure 6 The illustrated electronic device 600 includes a processor 601 and a memory 603. The processor 601 and the memory 603 are connected, for example, via a bus 602. Optionally, the electronic device 600 may also include a transceiver 604. It should be noted that in practical applications, the transceiver 604 is not limited to one type, and the structure of this electronic device 600 does not constitute a limitation on the embodiments of the present invention.

[0201] Processor 601 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in this disclosure. Processor 601 may also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0202] Bus 602 may include a pathway for transmitting information between the aforementioned components. Bus 602 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 602 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0203] The memory 603 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other storage medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0204] The memory 603 stores application code that executes the present invention, and its execution is controlled by the processor 601. The processor 601 executes the application code stored in the memory 603 to implement the content shown in the foregoing method embodiments.

[0205] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0206] This application discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, causes the processor to execute an instruction generation method.

[0207] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0208] The above are only some embodiments of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for constructing user data modeling, characterized in that, include: Based on the user's original information, a snapshot of user behavior is generated, wherein the original user information includes the user's original information within the target platform and pre-selected additional information; Based on the user behavior snapshot information, generate model input data; Based on a pre-built supervised fine-tuning dataset, a pre-trained data generation model is trained to obtain a user data modeling and inference model. The supervised fine-tuning dataset includes multiple input samples labeled with modeling labels. The input samples include user behavior samples and their contextual background samples. The modeling labels include modeling and inference target factors. The model input data and predefined modeling and reasoning prompts are input into the user data modeling and reasoning model to obtain the user data modeling and reasoning result. The modeling and reasoning prompts include model role positioning information, reasoning behavior rules, and task execution logic order.

2. The method according to claim 1, characterized in that, Based on the user's original information, generate user behavior snapshot information, including: Extract the user's original behavior logs on the target platform, and extract relevant content of preset indicators from the original behavior logs as preset indicator information; Based on the turnover cycle associated with the preset indicator, the preset indicator information is sliced ​​to obtain multi-granularity time slices; Extract the first preset number of data points from the multi-granularity time slices to form a single time window aggregated information; From the original behavior log, information with a repetition period greater than a preset period threshold is extracted to obtain long-cycle behavior information; The long-cycle behavioral information is arranged in chronological order and used as time-series modeling information; Based on preset association relationships, data in the original behavior log that contain multiple facts and whose association degree meets preset association conditions are merged to form multi-fact aggregated information. Extract and fuse user-selected additional information as dimensional association enhancement information, wherein the pre-selected additional information includes identity information, asset information and / or preset behavior information; User behavior snapshot information is obtained based on the single-time-window aggregation information, the time-series modeling information, the multi-fact aggregation information, and the dimension association enhancement information.

3. The method according to claim 1, characterized in that, Based on the user behavior snapshot information, model input data is generated, including: The user behavior snapshot information is compressed and merged to construct the original wide data table; The original wide data table is grouped according to preset dimensions, wherein the preset dimensions include basic information, behavioral information, transaction information and / or additional information; Construct a user context description template based on the grouping results; Based on the user context description template, information is populated according to the user behavior snapshot information to obtain model input data.

4. The method according to claim 1, characterized in that, Based on a pre-built supervised fine-tuning dataset, a pre-trained data generation model is trained to obtain a user data modeling and inference model, including: Based on a pre-trained data generation model, multiple base learners are constructed, wherein each of the multiple base learners corresponds one-to-one with a pre-constructed multiple thought chain reasoning links; Based on the supervised fine-tuning dataset, the multiple base learners are trained independently according to their respective thought chain reasoning links; The trained base learners are used as the user data modeling and inference model.

5. The method according to claim 4, characterized in that, The model input data and predefined modeling and reasoning prompts are input into the user data modeling and reasoning model to obtain the user data modeling and reasoning results, including: The model input data and predefined modeling and reasoning prompts are input into the multiple base learners trained in the user data modeling and reasoning model to obtain multiple reasoning results; The user data modeling reasoning results are obtained by aggregating the multiple reasoning results through voting.

6. The method according to claim 1, characterized in that, The supervised fine-tuning dataset is pre-built through the following steps: For the pre-constructed modeling and reasoning target factors, a standardized semantic description is generated, and model training prompt words are defined based on the standardized semantic description. The standardized semantic description includes dimension name, definition description, judgment basis, typical positive and negative examples and / or reasoning rule template. Obtain user behavior samples and their contextual background samples from multiple users to construct input samples; The input samples and the model training prompts are input into a pre-trained data processing model to obtain preliminary annotation results; Based on preset expert experience information, the preliminary annotation results are annotated independently multiple times; Compare the results of the multiple independent annotations, and determine the modeling labels based on the comparison results; Based on the modeling labels, the reasoning process of the preliminary annotation results of the model inference, and the input samples, a supervised fine-tuning dataset is constructed.

7. The method according to claim 1, characterized in that, The method further includes: Construct a preference dataset, wherein the preference dataset includes multiple sets of preference pairs, and the preference pairs include model input, preference model output, and non-preference model output; The DPO strategy is optimized using direct preference optimization, and the user data modeling inference model is iteratively optimized based on the preference dataset.

8. The method according to claim 1, characterized in that, The method further includes: The user data modeling and inference results are visualized and / or analyzed. Personalized searches are performed based on the modeling and reasoning results of the user data; and / or Based on the modeling and reasoning results of the user data, the original user information is processed.

9. The method according to claim 8, characterized in that, Visualizing and / or analyzing the user data modeling and inference results, including: The user data modeling and reasoning results are visualized based on a preset structured style; Based on the data generation model, a display image is generated according to the modeling and inference results of the user data; and / or Based on a preset real-time dialogue model, the modeling and reasoning results of the user data are displayed to the user.

10. The method according to claim 8, characterized in that, Based on the modeling and reasoning results of the user data, personalized searches are performed, including: Based on the user data modeling and reasoning results, a user data modeling dataset is formed; Matching is performed in the user data modeling dataset based on the search intent; Based on the matching results, obtain relevant user behavior snapshot information.

11. The method according to claim 8, characterized in that, Based on the modeling and reasoning results of the user data, personalized searches are performed, including: Based on the search intent, enhanced retrieval is performed using the original user information to obtain information related to the search intent from the user data modeling and inference results corresponding to the original user information.

12. The method according to claim 1, characterized in that, The method further includes: Based on the aforementioned user data modeling and inference model, a user data modeling backend interface is constructed, allowing external platforms to call the user data modeling and inference model through the user data modeling backend interface to generate user data modeling and inference results; and / or The user data modeling and reasoning results are converted into a structured language to generate a characteristic audience description, which can then be configured by external platforms based on the characteristic audience description.

13. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-12.

14. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the method as described in any one of claims 1-12.

Images