A large model driven AI e-commerce intelligent agent personalized interaction service method

Abstract

This application discloses a large-model-driven AI e-commerce agent personalized interaction service method, including the following steps: S1: Collect real-time interaction data between users and the AI ​​e-commerce agent; S2: Perform hierarchical isolation and storage management of long-term inherent consumption preferences and short-term temporary consumption needs, storing long-term inherent consumption preferences in a global persistent preference library to achieve cross-session, cross-cycle retention and iterative updates; S3: Perform fine-grained semantic analysis on the user's real-time interactive dialogue content; S4: Combine the user's historical interaction characteristics and real-time decision-making behavior to determine the user's corresponding decision-making trait type; S5: Integrate four types of personalized factors: long-term inherent consumption preferences, short-term temporary consumption needs, implicit preference characteristics, and interaction strategy parameters corresponding to decision-making traits; S6: Input the corrected optimal personalized interaction parameter combination into the large model.

Description

Technical Field

[0001] This application relates to the fields of natural language processing and intelligent interaction technology, and in particular to a method for providing personalized interactive services for AI e-commerce intelligent agents driven by a large model. Background Technology

[0002] With the rapid development of the e-commerce industry, interactive services such as online shopping guides, pre-sales consultations, and demand responses have become core components of e-commerce platform operations. Traditional human customer service suffers from high labor costs, limited service timeliness, and difficulty in standardizing service, making it unsuitable for business scenarios with massive numbers of users simultaneously inquiring. In recent years, large language models with powerful semantic understanding, contextual dialogue, and text generation capabilities have been widely applied. AI e-commerce agents built based on these large models are gradually replacing traditional human customer service, becoming the mainstream human-computer interaction service carrier in the e-commerce field. To further improve service quality, the industry generally regards personalized interaction as the core optimization direction for AI e-commerce agents, that is, generating differentiated response content based on user consumption characteristics to achieve personalized shopping guides and consultation services.

[0003] Current AI-powered e-commerce personalized interaction solutions primarily rely on large models to analyze dialogues and generate responses. Their personalization logic typically builds a unified user preference database based on explicit user behavior data such as clicks, favorites, and add-to-cart actions, and then outputs interactive information based on the user's current dialogue content. However, existing technologies still have several inherent technical shortcomings in practical applications.

[0004] During e-commerce interactions, users simultaneously exhibit both long-term, stable, and inherent consumption preferences, as well as temporary purchasing needs arising from single conversations. Current technologies cannot effectively isolate and manage these two types of preferences. When users initiate interactions across conversations or time periods, these temporary needs are retained and misjudged as long-term preferences, leading to preference drift. This results in contradictions in the shopping guide logic and responses across multiple rounds of interaction by the agent, significantly reducing the consistency and stability of long-term personalized interactions. Furthermore, the lack of conversation boundary recognition and temporary preference lifecycle management mechanisms in current technologies further exacerbates the problems of mixed preference data and distorted interaction effects.

[0005] During the product selection process, users often have implicit consumer demands that they don't actively express, such as price concerns, quality worries, questions about parameter compatibility, and indecisiveness. This information is hidden in the tone, sentence structure, and semantic tendencies of the conversation. Current technologies lack the ability to deeply analyze the fine-grained semantics of dialogue, and cannot automatically identify users' implicit decision-making concerns. They can only achieve passive question-and-answer responses, causing personalized services to remain superficial and failing to accurately address users' real decision-making pain points, resulting in insufficient service refinement.

[0006] Existing e-commerce intelligent agents generally adopt a fixed and uniform interaction paradigm to provide services to all users. The interaction rhythm, guidance intensity, and dialogue style are all preset standardized templates, without dynamic adaptation to the individual decision-making characteristics of users. Different users have different decision-making habits. Some users make efficient decisions and prefer concise responses, while others make cautious decisions and require step-by-step guidance and information comparison. Existing solutions cannot distinguish user decision-making characteristics and match corresponding interaction strategies, ultimately resulting in a homogenized interaction experience for all users and failing to achieve truly in-depth personalized services.

[0007] To address this, we propose a large-model-driven method for personalized interactive services in AI e-commerce. Summary of the Invention

[0008] This application aims to at least partially solve one of the technical problems in the aforementioned technologies.

[0009] To achieve the above objectives, the first aspect of this application proposes a method for providing personalized interaction services to AI e-commerce intelligent agents driven by a large model, comprising the following steps:

[0010] S1: Collect real-time interaction data between users and AI e-commerce agents, and use a large-scale model-specific semantic discrimination model adapted to e-commerce scenarios to perform binary hierarchical recognition of user interaction preferences, distinguish and obtain users' long-term inherent consumption preferences and short-term temporary consumption needs.

[0011] S2: Implement layered and isolated storage management for long-term inherent consumption preferences and short-term temporary consumption needs. Long-term inherent consumption preferences are stored in a global persistent preference library to achieve cross-session and cross-period retention and iterative updates. Short-term temporary consumption needs are stored in a time-limited buffer bound to the current interaction session. Through a preset session boundary determination mechanism, the entire lifecycle of short-term temporary consumption needs is managed.

[0012] S3: Perform fine-grained semantic analysis on the real-time interactive dialogue content of users, explore the implicit consumption concerns and potential decision-making demands that users do not explicitly express, construct the corresponding implicit preference features, and use the implicit preference features as the correction factor for personalized interaction.

[0013] S4: Combine user's historical interaction characteristics with real-time decision-making behavior to determine the user's corresponding decision-making trait type, and match adaptive interaction strategy parameters that are suitable for that trait type. The interaction strategy parameters include at least the interaction rhythm, guidance intensity, and speech style parameters.

[0014] S5: Integrates four types of personalized factors: long-term inherent consumption preferences, short-term temporary consumption needs, implicit preference characteristics, and interaction strategy parameters corresponding to decision-making traits. It calculates the deviation of the current conversation preference from the long-term inherent consumption preference, implements progressive preference drift correction based on the preference deviation, and outputs the optimal combination of personalized interaction parameters.

[0015] S6: Input the corrected optimal combination of personalized interaction parameters into the large model to drive the AI ​​e-commerce agent to generate personalized interactive service content that adapts to user consumption preferences, matches user decision-making habits, and covers user potential consumption needs, thus completing a complete personalized interactive service closed loop.

[0016] In addition, the large model-driven AI e-commerce intelligent agent personalized interaction service method proposed in this application may also have the following additional technical features:

[0017] As a further description of the above technical solution:

[0018] In step S1, a large-scale semantic discrimination model adapted to e-commerce scenarios completes preference binary classification based on e-commerce scenario-specific semantic rules.

[0019] Long-term, established consumption preferences are stable and normalized consumption characteristics of users, while short-term, temporary consumption needs are temporary and non-fixed consumption demands specific to a single interactive session.

[0020] As a further description of the above technical solution:

[0021] In step S2, the session boundary determination mechanism uses the interaction time interval, the semantic relevance of the previous and subsequent interactions, and the continuity of user intent as the core determination dimensions to distinguish between new interactive sessions and continued interactive sessions.

[0022] If the interaction is determined to be a brand new interaction session, the time buffer data corresponding to the previous interaction session will be automatically cleared; if the interaction is determined to be a continued interaction session, the current time buffer data will be retained and continuously updated in combination with real-time interaction content.

[0023] As a further description of the above technical solution:

[0024] The implementation method for progressive preference drift correction in step S5 is as follows:

[0025] A fixed preference deviation threshold is preset. When the preference deviation calculated in real time exceeds the threshold, the weight of long-term inherent consumption preferences is gradually increased while the weight of short-term temporary consumption needs is simultaneously decreased until the preference deviation falls back to the threshold range, effectively avoiding sudden deviations in personalized interactive content.

[0026] As a further description of the above technical solution:

[0027] In step S3, implicit consumer concerns include users' concerns about price sensitivity, product quality, parameter compatibility, and dilemmas in making cost-effective decisions. By analyzing the sentence structure, semantic tendencies, and expression states of user interaction texts, automated mining and identification can be achieved without explicit claims.

[0028] As a further description of the above technical solution:

[0029] In step S4, the user decision-making trait types include decisive and quick selection, repeated price comparison, cautious and conservative, and passive consultation. Each decision-making trait type is configured with an independent and standardized set of interaction strategy parameters.

[0030] As a further description of the above technical solution:

[0031] In step S5, the four types of personalized factors adopt a dynamic weight fusion mechanism, which can dynamically adjust the weight ratio of each dimension factor according to the current session scenario and the user's real-time interaction status, and the total weight of each factor remains at a fixed threshold to ensure the stability and scenario adaptability of the personalized interaction parameter output.

[0032] Advantages of this invention:

[0033] According to the large model-driven AI e-commerce intelligent agent personalized interaction service method of this application, the present invention manages the life cycle of short-term temporary consumption needs by performing binary hierarchical identification of user interaction preferences and adopting hierarchical isolation storage combined with a session boundary judgment mechanism. At the same time, it is equipped with a progressive preference drift correction strategy, which effectively avoids the problems of preference data aliasing, preference drift in cross-session interaction and logical contradictions in response content in traditional solutions, and effectively ensures the consistency and operational stability of long-term personalized interaction services of AI e-commerce intelligent agents.

[0034] This invention relies on fine-grained dialogue semantic parsing to automatically mine users' implicit consumption concerns and potential decision-making needs. It breaks through the limitations of existing technologies that only rely on users' explicit expressions to respond, enriches the dimensions of personalized features, improves the depth of demand identification and the accuracy of interactive responses, can proactively predict and respond to users' potential needs, and enhances the level of service refinement.

[0035] This invention can also identify decision-making trait types based on users' historical interaction characteristics and real-time decision-making behavior, and match corresponding adaptive interaction strategy parameters such as interaction rhythm, guidance intensity, and speech style. It abandons the traditional fixed and unified interaction paradigm, completely improves the drawbacks of homogenized interaction experience, and achieves precise adaptation of interaction strategy and individual user characteristics.

[0036] This invention requires no modification to the large model base or the original AI e-commerce agent's main structure, making it easy to deploy, versatile, and compatible. The engineering implementation cost is controllable. Overall, this invention comprehensively optimizes the operation of the large model-driven AI e-commerce agent from multiple dimensions such as preference management, demand mining, and interaction adaptation, significantly improving the overall quality of personalized interactive services and user experience.

[0037] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0038] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0039] Figure 1 This is a schematic diagram of the architecture of a large model-driven AI e-commerce intelligent agent personalized interaction service method according to an embodiment of this application.

[0040] Figure 2 This is a flowchart illustrating a method for providing personalized interactive services for an AI e-commerce intelligent agent driven by a large model, according to an embodiment of this application.

[0041] Figure 3 This is a schematic diagram illustrating the preference hierarchical storage, session determination, and preference drift correction logic of a large model-driven AI e-commerce intelligent agent personalized interaction service method according to an embodiment of this application. Detailed Implementation

[0042] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0043] The following description, in conjunction with the accompanying drawings, illustrates a large-model-driven AI e-commerce intelligent agent personalized interaction service method according to Embodiment 1 of this application. Figure 1-3 As shown:

[0044] The invention adopts a cloud-based distributed deployment model, with supporting hardware divided into three categories: application service nodes, data storage nodes, and large model inference nodes. Each node has independent functions and interconnects with the network to meet the operational needs of high-concurrency e-commerce interaction scenarios. The application service nodes use industrial-grade cloud servers, which mainly carry out business logic operations such as interactive data collection, preference recognition, session determination, feature fusion, and parameter scheduling. The data storage nodes are divided into persistent storage servers and high-speed cache servers, which are used for the permanent storage and iteration of long-term inherent consumption preferences. The high-speed cache servers use high-performance memory-based servers, which are specifically used to carry out session-level time-sensitive caching for short-term temporary consumption needs. The large model inference nodes are equipped with professional AI inference servers, which are responsible for core inference tasks such as large model semantic parsing and text generation. Data is transmitted in real time between the nodes through a gigabit intranet, ensuring the stability of data transmission under massive concurrent user interactions.

[0045] The operating system uniformly adopts Linux Enterprise Edition to ensure the stability and security of server operation. At the data storage level, the global persistent preference database uses the relational database MySQL, which relies on the data table structure to realize the structured storage, query and update of users' long-term preferences. The session-bound time-limited buffer uses the distributed in-memory database Redis, which leverages its high-performance read and write, key-value pair management and automatic expiration cleanup features to realize lifecycle management of short-term temporary consumption needs, which is fully matched to the use case of session-level temporary data. The middleware deploys a message queue component for queuing and processing interactive data in high-concurrency scenarios to avoid data congestion. The large language model uses a commercial large model with general semantic understanding, multi-turn dialogue and text generation capabilities, and only completes lightweight Prompt adaptation and scenario rule binding for e-commerce dialogue scenarios.

[0046] During the above preparation process, it is necessary to initialize and configure the entire process rule base, judgment threshold, weight range, and parameter template, as follows:

[0047] The threshold for continuous interaction time in a single session is set at 30 minutes. If the time interval between two interactions by the same user exceeds 30 minutes, it is directly determined as a brand new interaction session. If the time interval is less than or equal to 30 minutes, it is determined as a continued interaction session. The semantic relevance is calculated using the cosine similarity algorithm, and the threshold for semantic relevance is set at 0.6. The session status is determined by combining the three dimensions of interaction time interval, semantic relevance, and user intent continuity.

[0048] Three levels of deviation thresholds are preset: the basic safety threshold is 0.1, the first-level warning threshold is 0.2, and the second-level forced correction threshold is 0.3. The progressive weight adjustment step size is fixed at 0.05. The deviation is recalculated after each weight adjustment until the deviation falls back to the basic safety threshold range.

[0049] The total weight of the four personalized factors in this invention is fixed at 1. The initial basic weight allocation is as follows: long-term inherent consumption preference weight 0.4, short-term temporary consumption demand weight 0.3, implicit preference feature weight 0.15, and decision-making trait interaction parameter weight 0.15. During operation, the proportion of each factor is dynamically adjusted only within a preset range. The weight value range is set to 0~0.6 to avoid the imbalance of interaction logic caused by excessive weight of a single factor.

[0050] For four scenarios—price sensitivity, product quality, parameter compatibility, and cost-effectiveness decision-making—corresponding text feature words, sentence structure rules, and semantic tendency tags are preset to establish a standardized matching rule set for automated recognition in the fine-grained semantic parsing process.

[0051] The system categorizes decision-making traits into four types: decisive and quick selection, repeated price comparison, cautious and conservative, and passive consultation. For each trait, it assigns a unique parameter template for interaction rhythm, guidance intensity, and communication style. Interaction rhythm parameters include the frequency of questions and answers per round, the interval between responses, and the maximum number of characters in a single reply. Guidance intensity parameters include the number of proactive questions, the number of product information pushes, and the proportion of comparison content displayed. Communication style parameters differentiate between four paradigms: concise, detailed, question-and-answer, and guiding. All parameter templates are standardized and stored, and can be directly accessed based on the assessment results.

[0052] Clearly define the criteria and feature labeling system for long-term inherent consumption preferences and short-term temporary consumption needs, provide a unified basis for semantic judgment of large models, and ensure the consistency of preference classification results.

[0053] This invention executes the interactive data acquisition unit, preference hierarchical identification and storage unit, implicit concern mining unit, decision trait discrimination and strategy matching unit, and multi-factor fusion and interactive generation unit in a sequential manner, while setting up a data write-back link to send the effective feature data generated in each round of interaction back to the corresponding storage area.

[0054] The specific steps of the large-model-driven AI e-commerce intelligent agent personalized interaction service method in Embodiment 1 of this application are as follows:

[0055] S1 collects real-time interactive data and completes binary hierarchical preference recognition:

[0056] When users initiate text-based inquiries, product requests, or purchase communication on the e-commerce front-end page, the interaction data collection unit captures all raw interaction data in real time. The collected data includes six core categories: user's unique account ID, current session's globally unique ID, interaction initiation timestamp, user's input text content, current device identifier, and brief identifier of the user's historical sessions. After collection, the raw data is preprocessed to remove invalid empty characters, garbled characters, special symbols, and other interfering content, retaining clean interaction text and basic identifier information. It is then packaged into a standard interaction data package according to a preset data format and pushed to the large-scale e-commerce-specific semantic discrimination model.

[0057] In this embodiment, the large-scale e-commerce-specific semantic discrimination model is built upon the native semantic understanding capabilities of the large-scale model and combines e-commerce scenario-specific semantic rules to carry out preference classification. The model loads a pre-configured preference binary classification rule library. First, it performs word segmentation, syntactic parsing, and intent extraction on the user interaction text. Then, it matches the parsing results with the rule library to strictly distinguish between long-term inherent consumption preferences and short-term temporary consumption needs. Long-term inherent consumption preferences are defined as stable and normalized consumption attribute characteristics of users, specifically covering features such as users' fixed purchase categories, long-term acceptable price ranges, consistently preferred product styles, fixed brand series, and commonly used product specifications and sizes, which have continuous and stable characteristics. Short-term temporary consumption needs are defined as temporary and non-fixed consumption demands specific to a single interaction session, specifically including demands that only apply to the current single session, such as purchasing on behalf of relatives and friends, temporary purchases during holidays, combining orders for discounts, temporary attempts at new products, and emergency purchases.

[0058] After the model completes parsing and matching, it outputs two types of results: first, a preference classification conclusion, clarifying whether the current interaction request belongs to a long-term inherent consumption preference or a short-term temporary consumption need; second, a set of structured feature tags corresponding to the preference, such as category tags, price tags, style tags, and demand scenario tags. The model's recognition confidence threshold is set to 0.7. If the recognition confidence is higher than 0.7, the classification result is directly output; if the confidence is lower than 0.7, a second semantic review is triggered, combining the context and historical interaction content for a comprehensive judgment, ultimately ensuring the accuracy and reliability of the preference classification result. After classification is completed, the preference classification conclusion, feature tag set, and basic identification data are pushed to the next execution stage.

[0059] S2 Tiered Isolation Storage Management, Session Boundary Determination, and Lifecycle Control:

[0060] First, a hierarchical isolation storage operation is performed. For the identified long-term inherent consumption preferences, the system retrieves the MySQL global persistent preference database, using the user's unique account ID as the primary key, to query the user's stored historical long-term preference tag set. The newly identified long-term preference features are then fused and iterated with the historical features, removing invalid features and adding new stable features. After the data update is completed, it is rewritten to the persistent preference database. The global persistent preference database adopts a permanent storage mechanism, and the data will not be deleted when the session ends or the interaction is interrupted. It supports global access across sessions, dates, and devices, serving as the basis for the user's personalized interaction throughout the entire lifecycle.

[0061] For the identified short-term temporary consumption needs, the system calls the Redis high-speed time-sensitive buffer, uses the current session's globally unique ID as the key name, and stores the feature tag set corresponding to the short-term needs, session creation time, and last interaction time as key values ​​to achieve a forced binding between short-term needs and sessions. Short-term temporary consumption needs are only effective within the current session period and do not participate in cross-session data inheritance.

[0062] The system then initiates a session boundary determination mechanism, which strictly adheres to three preset core dimensions: interaction time interval, semantic relevance of previous and subsequent interactions, and continuity of user intent. The system extracts the timestamp of the user's last valid interaction and calculates the time interval between the current and previous interactions; simultaneously, it retrieves the interaction text and intent vectors from historical sessions and calculates the semantic relevance of the content using a cosine similarity algorithm; finally, it combines the direction of the requests from the two interactions and the selected product category to determine the continuity of intent. The session state is then determined by comprehensively considering these three indicators.

[0063] If the interaction time interval is greater than 30 minutes, or the semantic relevance is less than 0.6, it is directly determined as a brand new interaction session. The system automatically retrieves and clears the Redis time-limited buffer data corresponding to the previous historical session to avoid interference from past temporary needs to the current interaction, and only retains the short-term temporary consumption needs of the current session.

[0064] If the interaction time interval is less than or equal to 30 minutes, and the semantic relevance is greater than or equal to 0.6 and the intent remains continuous, it is determined to be a continuation of the interaction session. The system retains all the data in the current Redis time-limited buffer and merges and updates the newly generated short-term demand feature tags with the original tag set to realize the dynamic iteration of temporary demands within the same session.

[0065] After the session status determination and data processing are completed, the system integrates the long-term preference tag set, the current session short-term preference tag set, the session status identifier, and the basic session information into a preference storage data packet and pushes it to the next stage.

[0066] S3 fine-grained semantic parsing and mining of implicit consumer concerns:

[0067] The system extracts the original interactive dialogue content of the user in this round and conducts full-dimensional fine-grained semantic analysis. The analysis dimensions include five categories: text sentence structure, lexical features, tone tendency, semantic logic, and expression state. Relying on the deep semantic analysis capabilities of the large model, it deconstructs the potential tendencies behind each sentence of the user. During the analysis process, the system calls the pre-configured implicit consumption concern identification rule base and compares the obtained sentence features, semantic tendencies, and expression states with the feature words and matching rules in the rule base one by one to automatically identify various implicit consumption concerns of the user. In this embodiment, four typical implicit demands are identified: price sensitivity concerns, product quality concerns, parameter adaptation concerns, and cost-effectiveness decision-making dilemmas.

[0068] Among them, price sensitivity concerns correspond to users' potential worries about product pricing, promotional activities, and discount levels; product quality concerns correspond to users' potential questions about materials, workmanship, after-sales service, and stability of use; parameter compatibility concerns correspond to users' uncertainty about the matching of specifications, sizes, models, and usage scenarios; and cost-effectiveness decision-making dilemmas correspond to users' psychological state of weighing and comparing multiple products and having difficulty making a quick decision. The entire mining process does not require users to actively label implicit demands. It is completely based on publicly available interactive text to complete automated identification. After identification, a standardized implicit preference feature vector is generated. The vector dimensions correspond one-to-one with the preset rule base. The value range of each dimension is 0 to 1. The higher the value, the stronger the corresponding implicit concern.

[0069] The system defines the generated implicit preference feature vector as a personalized interaction correction factor. This correction factor will subsequently participate in multi-dimensional feature fusion to adjust the agent's response focus, question-answering direction, and content detail. This step outputs the implicit preference feature vector, concern type identifier, and original semantic parsing results, and merges them with the preference data from the previous step to form a comprehensive user feature package containing explicit preferences and implicit features, which is then pushed to the decision trait discrimination step.

[0070] S4 User Decision Trait Identification and Adaptive Interaction Strategy Parameter Matching:

[0071] The system first extracts two types of feature data for decision-making trait determination: the first type is the user's historical interaction features, including statistical data such as the average interaction frequency of the user's historical sessions, the number of questions per round, the duration of historical purchase decisions, and the termination points of historical interactions; the second type is the real-time decision-making behavior features of the current round, including real-time data such as the question format, the number of questions, the focus of the request, and the interaction response speed. The two types of feature data are input into the decision-making trait discrimination rule base, and the comprehensive decision-making feature score is calculated according to the pre-designed scoring rules. Based on the score range, the user's decision-making trait type is divided into four categories: decisive and quick selection type, repeated price comparison type, cautious and conservative type, and passive consultation type. The judgment criteria and feature boundaries of the four types are clear and there is no overlap or confusion.

[0072] After classifying decision-making traits, the system retrieves the corresponding set of standardized interaction strategy parameters based on the classification results. These interaction strategy parameters include at least three categories: interaction rhythm, guidance intensity, and communication style. Each category of parameters has specific quantitative indicators.

[0073] For decisive and quick-selection users, use simplified interaction parameters, set the interaction rhythm to low question-and-answer frequency and long interaction response interval, and control the maximum number of characters in a single reply to within 200 characters; set the guidance intensity to low intensity, with no more than one proactive question per round and no more than two product pushes; use a concise and direct style of communication, focusing on outputting core information and reducing redundant descriptions.

[0074] For users who repeatedly compare prices, the interaction parameters are matched with comparison-oriented parameters, the interaction rhythm is set to a medium question-and-answer frequency, and the maximum number of characters in a single reply is increased to 500 characters; the guidance intensity is set to medium to high intensity, actively displaying comparisons of parameters and prices of multiple products; the wording style adopts a data-rich approach, focusing on listing differences to assist users in making comparison decisions.

[0075] For cautious and conservative users, the interaction parameters are set to a Q&A style, and the interaction rhythm is set to a low-frequency, slow-response mode. The guidance is mainly passive response, reducing active recommendation behavior. The language style is rigorous and detailed, focusing on explaining product qualifications, after-sales policies, and quality assurance.

[0076] For passively seeking advice users, we match guided interaction parameters and set the interaction rhythm to a high-frequency, light question-and-answer mode; the guidance intensity is set to high intensity, gradually and proactively guiding users to clarify their needs; the language style is friendly and guiding, gradually lowering the user's decision-making threshold.

[0077] After parameter matching is completed, the system encapsulates the decision trait type identifier and the complete set of interaction strategy parameters, merges them with the preceding comprehensive user feature package, and forms a complete multi-dimensional feature data package containing explicit preferences, implicit features, and decision parameters, which is then pushed to the multi-factor fusion and drift correction stage.

[0078] S5 Four-Dimensional Factor Dynamic Weight Fusion and Gradual Preference Drift Correction:

[0079] First, the system performs dynamic weight fusion of four-dimensional personalized factors. The four types of factors are long-term inherent consumption preferences, short-term temporary consumption needs, implicit preference characteristics, and interaction strategy parameters corresponding to decision-making traits. The fusion calculation strictly follows the rule that the total weight is fixed at 1. Based on the initial basic weight allocation scheme, the system dynamically fine-tunes the weight ratio of each factor in combination with the current session scenario and the user's real-time interaction status. The weight adjustment range is limited to a preset range to avoid abnormal weights of single factors. The fusion process adopts a weighted vector superposition algorithm, which calculates the feature vectors corresponding to the four types of factors according to their respective weights to generate the current session comprehensive preference vector. This vector represents the user's overall personalized characteristics in the current state.

[0080] The system then calculates the preference deviation, using the long-term inherent consumption preference vector stored in the global persistent preference library as the baseline vector and the current session comprehensive preference vector generated in the previous step as the comparison vector. The Euclidean distance formula is used to calculate the deviation value between the two sets of vectors, thereby quantifying the degree of deviation of the current preference relative to the user's long-term stable preference. After the calculation is completed, the deviation value is compared with a pre-set three-level threshold, and different processing logics are executed according to different scenarios:

[0081] When the deviation is ≤0.1 (basic safety threshold), it is determined that there is no obvious preference drift, no correction operation is required, and the current comprehensive preference vector and interaction strategy parameters are directly integrated into a preliminary personalized interaction parameter combination.

[0082] When 0.1 < deviation ≤ 0.2 (first-level warning threshold), it is judged as a slight preference drift, and a gradual correction mechanism is initiated. The weight of long-term inherent consumption preferences is gradually increased in a fixed step of 0.05, while the weight of short-term temporary consumption demand is simultaneously decreased. After the weight adjustment is completed, the deviation is recalculated until the value falls back to the basic safety threshold range.

[0083] When 0.2 < deviation ≤ 0.3 (secondary forced correction threshold), it is judged as moderate preference drift. The frequency of weight adjustment is increased. The deviation is reviewed immediately after each weight adjustment. Gradual correction is continuously implemented to strictly control the drift range and avoid the interaction logic from deviating significantly from the user's long-term habits.

[0084] When the deviation is greater than 0.3, it is judged as a severe drift. While performing a gradual weight adjustment, an anomaly log is recorded to facilitate backend operation and maintenance verification.

[0085] The entire correction process adopts a gradual adjustment mode, rather than a one-time forced reset of preferences, which can effectively avoid sudden deviations in personalized interactive content and ensure the consistency of the interactive experience. Once the deviation falls back to a safe range, the system integrates all feature vectors, corrected weight parameters, and interaction strategy parameters to generate the final optimal combination of personalized interaction parameters, completing all operations in this stage, and then pushes the parameter combination to the interactive content generation stage.

[0086] S6 Large Model Calling and Personalized Interactive Service Content Generation:

[0087] The system encapsulates the optimal personalized interaction parameter combination, full-dimensional user feature tags, and basic conversation information output from the previous step, organizes the input content according to the pre-set call interface specifications of the large model, and loads e-commerce exclusive guidance prompts. It explicitly requires the large model to combine four types of information—long-term user preferences, current temporary needs, implicit consumption concerns, and decision-making traits—to generate response content, and strictly follows the matching interaction rhythm, guidance intensity, and conversation style parameters to carry out text creation.

[0088] After receiving input data, the large model combines its own text generation capabilities with all personalized constraints to generate personalized interactive service content that adapts to user consumption preferences, matches user decision-making habits, and proactively covers users' potential implicit needs. The content formats include text Q&A, product interpretation, purchase guidance, activity descriptions, etc., which are fully aligned with the business scenarios of e-commerce shopping guides. The intelligent agent pushes the generated interactive content to the front-end page to display to the user, completing a single human-computer interaction response.

[0089] After a single interaction is completed, the system performs a data write-back operation: newly generated valid long-term preference features from this round of interaction are written back to the MySQL global persistent preference database, completing the iterative update of the user's long-term profile. The short-term demand features from this round are retained in the Redis time-limited buffer corresponding to the current session for subsequent interactions within the same session. The interaction behavior data and decision trait judgment results from this round are simultaneously stored in the user behavior statistics database to accumulate samples for subsequent decision trait discrimination. At this point, a complete personalized interaction service process is finished. If the user continues to initiate new interaction requests, the system will repeat the above steps S1 to S6 to achieve continuous multi-round interactions. If the session times out or the user actively exits the session, the system automatically triggers the Redis time-limited buffer clearing operation to release temporary data resources.

[0090] In summary, the large-model-driven AI e-commerce intelligent agent personalized interaction service method proposed in this application relies on five core technologies: hierarchical preference recognition and isolated storage, session boundary control, implicit concern mining, adaptive matching of decision traits, multi-factor dynamic fusion, and progressive drift correction. It addresses the technical shortcomings of existing large-model e-commerce intelligent agents from multiple dimensions. In actual operation, this solution can effectively distinguish between users' long-term inherent preferences and short-term temporary needs, completely eliminating preference drift and logical inconsistencies in response content during cross-session interactions. It significantly improves the consistency and stability of long-term personalized interaction services through fine-grained... This solution uses semantic parsing to uncover users' implicit consumption concerns, breaking through the limitations of traditional solutions that rely solely on explicit demand responses. It enhances the depth of demand identification and the refinement of services, dynamically adjusting the interaction rhythm, guidance intensity, and communication style based on user decision-making characteristics. This breaks away from the fixed and uniform interaction paradigm, achieving a deeply personalized interactive experience tailored to each user. Furthermore, this solution adopts a lightweight, plug-in architecture design, requiring no modification to the large model base or the original intelligent agent's main structure. It has low hardware deployment threshold, strong software compatibility, and controllable operation and maintenance costs, and can be directly and scalably applied to various e-commerce AI intelligent agent scenarios, possessing extremely high engineering implementation value and market promotion value.

[0091] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0093] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for providing personalized interactive services to AI e-commerce intelligent agents driven by a large model, characterized in that, Includes the following steps: S1: Collect real-time interaction data between users and AI e-commerce agents, and use a large-scale model-specific semantic discrimination model adapted to e-commerce scenarios to perform binary hierarchical recognition of user interaction preferences, distinguish and obtain users' long-term inherent consumption preferences and short-term temporary consumption needs. S2: Implement layered and isolated storage management for long-term inherent consumption preferences and short-term temporary consumption needs. Long-term inherent consumption preferences are stored in a global persistent preference library to achieve cross-session and cross-period retention and iterative updates. Short-term temporary consumption needs are stored in a time-limited buffer bound to the current interaction session. Through a preset session boundary determination mechanism, the entire lifecycle of short-term temporary consumption needs is managed. S3: Perform fine-grained semantic analysis on the real-time interactive dialogue content of users, explore the implicit consumption concerns and potential decision-making demands that users do not explicitly express, construct the corresponding implicit preference features, and use the implicit preference features as the correction factor for personalized interaction. S4: Combine user's historical interaction characteristics with real-time decision-making behavior to determine the user's corresponding decision-making trait type, and match adaptive interaction strategy parameters that are suitable for that trait type. The interaction strategy parameters include at least the interaction rhythm, guidance intensity, and speech style parameters. S5: Integrates four types of personalized factors: long-term inherent consumption preferences, short-term temporary consumption needs, implicit preference characteristics, and interaction strategy parameters corresponding to decision-making traits. It calculates the deviation of the current conversation preference from the long-term inherent consumption preference, implements progressive preference drift correction based on the preference deviation, and outputs the optimal combination of personalized interaction parameters. S6: Input the corrected optimal combination of personalized interaction parameters into the large model to drive the AI ​​e-commerce agent to generate personalized interactive service content that adapts to user consumption preferences, matches user decision-making habits, and covers user potential consumption needs, thus completing a complete personalized interactive service closed loop.

2. The method for personalized interactive services of AI e-commerce intelligent agents driven by a large model according to claim 1, characterized in that, In step S1, a large-scale semantic discrimination model adapted to e-commerce scenarios completes preference binary classification based on e-commerce scenario-specific semantic rules. Long-term, established consumption preferences are stable and normalized consumption characteristics of users, while short-term, temporary consumption needs are temporary and non-fixed consumption demands specific to a single interactive session.

3. The method for personalized interactive services of AI e-commerce intelligent agents driven by a large model according to claim 1, characterized in that, In step S2, the session boundary determination mechanism uses the interaction time interval, the semantic relevance of the previous and subsequent interactions, and the continuity of user intent as the core determination dimensions to distinguish between new interactive sessions and continued interactive sessions. If it is determined to be a brand new interaction session, the time buffer data corresponding to the previous interaction session will be automatically cleared; If the interaction session is determined to be a continuation session, the current time-sensitive buffer data is retained and continuously updated in combination with real-time interaction content.

4. A method for providing personalized interactive services for AI e-commerce intelligent agents driven by a large model, as described in claim 1, is characterized in that... The implementation method for progressive preference drift correction in step S5 is as follows: A fixed preference deviation threshold is preset. When the preference deviation calculated in real time exceeds the threshold, the weight of long-term inherent consumption preferences is gradually increased while the weight of short-term temporary consumption needs is simultaneously decreased until the preference deviation falls back to the threshold range, effectively avoiding sudden deviations in personalized interactive content.

5. A method for providing personalized interactive services for AI e-commerce intelligent agents driven by a large model, as described in claim 1, characterized in that... In step S3, implicit consumer concerns include users' concerns about price sensitivity, product quality, parameter compatibility, and dilemmas in making cost-effective decisions. By analyzing the sentence structure, semantic tendencies, and expression states of user interaction texts, automated mining and identification can be achieved without explicit claims.

6. A method for providing personalized interactive services for AI e-commerce intelligent agents driven by a large model, as described in claim 1, is characterized in that... In step S4, the user decision-making trait types include decisive and quick selection, repeated price comparison, cautious and conservative, and passive consultation. Each decision-making trait type is configured with an independent and standardized set of interaction strategy parameters.

7. A method for providing personalized interactive services for AI e-commerce intelligent agents driven by a large model, as described in claim 1, is characterized in that... In step S5, the four types of personalized factors adopt a dynamic weight fusion mechanism, which can dynamically adjust the weight ratio of each dimension factor according to the current session scenario and the user's real-time interaction status, and the total weight of each factor remains at a fixed threshold to ensure the stability and scenario adaptability of the personalized interaction parameter output.

Images