Product recommendation method and device, electronic equipment, medium and program product

Abstract

The invention provides a product recommendation method and device, electronic equipment, a medium and a program product, relates to application of a large model in the field of financial science and technology, and can be applied to the technical field of big data, the technical field of artificial intelligence and the field of financial science and technology. The method comprises the following steps: acquiring cross-domain behavior data and external environment data; based on the cross-domain behavior data, performing unified semantic coding and semantic tendency analysis by utilizing the first large model to obtain a high-dimensional semantic vector and a behavior tendency vector, and fusing the high-dimensional semantic vector and the behavior tendency vector by utilizing time perception attention to obtain a dynamic user state vector; performing natural language understanding and information extraction processing by using the second large model based on the external environment data to obtain an environment feature vector; and generating candidate product features and corresponding dynamic weight distribution based on the dynamic user state vector and the environment feature vector, and determining a product recommendation result based on the candidate product features and the dynamic weight distribution.

Description

Technical Field

[0001] This application relates to the application of large models in the field of financial technology, and to the fields of big data technology and artificial intelligence technology. More specifically, it relates to a product recommendation method, apparatus, device, medium, and program product. Background Technology

[0002] Existing financial product recommendations are mostly based on collaborative filtering, rule matching, or time-series prediction methods, relying primarily on static analysis of structured data such as transactions and holdings. This makes it difficult to reflect the dynamic changes in user needs and the market environment. Their recommendation logic lacks the ability to integrate and adaptively update multi-source behavioral data, failing to adjust recommendation strategies in a timely manner according to market fluctuations or changes in user risk preferences. Furthermore, the system's decision-making process is largely a "black box" of static mapping, lacking causal reasoning and interpretability, resulting in rigid and unreliable recommendation results that fail to meet the demands of the digital finance era for personalized, flexible, and trustworthy intelligent recommendations. Summary of the Invention

[0003] In view of the above problems, this application provides product recommendation methods, apparatus, devices, media and program products.

[0004] According to a first aspect of this application, a product recommendation method is provided, the method comprising: acquiring user-authorized cross-domain behavior data and external environment data; based on the cross-domain behavior data, performing unified semantic encoding and semantic tendency analysis using a first major model to obtain a high-dimensional semantic vector and a behavior tendency vector; fusing the high-dimensional semantic vector and the behavior tendency vector using time-aware attention to obtain a dynamic user state vector; based on the external environment data, performing natural language understanding and information extraction processing using a second major model to obtain an environmental feature vector; and generating candidate product features and corresponding dynamic weight distributions based on the dynamic user state vector and the environmental feature vectors, and determining a product recommendation result based on the candidate product features and the dynamic weight distributions.

[0005] According to an embodiment of this application, the step of using a first large-scale language model to perform unified semantic encoding and semantic tendency analysis to obtain high-dimensional semantic vectors and behavioral tendency vectors includes: acquiring textual and non-textual data from the cross-domain behavioral data; inputting the textual data into the first large-scale language model to obtain a first semantic feature vector; inputting the non-textual data into the first large-scale language model after descriptive processing to obtain a second semantic feature vector; fusing the first and second semantic feature vectors in a unified semantic space to obtain the high-dimensional semantic vector; and performing semantic tendency analysis on the textual data based on the semantic tendency analysis task head of the first large-scale language model to generate the behavioral tendency vector.

[0006] According to an embodiment of this application, the method further includes: using the first large model to perform interest intensity quantization and time-aware encoding on the cross-domain behavior data to obtain a domain interest intensity vector and a time-decaying interest vector; and using time-aware attention to perform weighted fusion of the high-dimensional semantic vector, the behavior tendency vector, the domain interest intensity vector, and the time-decaying interest vector to obtain a dynamic user state vector.

[0007] According to an embodiment of this application, the step of using the first large model to quantify the interest intensity and time-aware encoding of the cross-domain behavior data includes: using the first large model to perform semantic analysis on the interaction content of different domains in the cross-domain behavior data, calculating the user's interest intensity score in each domain based on the semantic analysis results, so as to generate the corresponding domain interest intensity vector; and using the timestamp information of the cross-domain behavior data to perform time-series feature modeling on historical behavior records using a time decay function or a time embedding encoding mechanism to obtain the time decay interest vector.

[0008] According to an embodiment of this application, the step of obtaining an environmental feature vector by performing natural language understanding and information extraction processing using a second major model based on the external environment data includes: performing natural language understanding using the second major model based on the external environment data to identify environmental semantic elements and corresponding relationships; performing signal decoupling on the environmental semantic elements and corresponding relationships to obtain a first environmental sub-vector and a second environmental sub-vector, wherein the first environmental sub-vector reflects global environmental changes and the second environmental sub-vector reflects local environmental changes related to the target product category; and using the second major model to numerically encode the first environmental sub-vector and the second environmental sub-vector respectively, and combining them in a unified semantic space to generate the environmental feature vector.

[0009] According to an embodiment of this application, generating candidate product features and corresponding dynamic weight distributions based on the dynamic user state vector and the environmental feature vector includes: extracting candidate product features from a product database based on the dynamic user state vector; performing matching calculations on the candidate product features to obtain corresponding domain-based weights; and using a dual-signal nonlinear correction function to correct the domain-based weights to obtain the dynamic weight distribution.

[0010] According to an embodiment of this application, the step of correcting the domain basic weights using a dual-signal nonlinear correction function includes: globally correcting the domain basic weights using the first environment sub-vector; and locally correcting the domain basic weights of the target product category using the second environment sub-vector.

[0011] According to an embodiment of this application, determining the product recommendation result based on the candidate product features and the dynamic weight distribution includes: using the dynamic weight distribution to weight and adjust the candidate product features to obtain a comprehensive dynamic matching score for each candidate product; determining a recommendation risk safety threshold based on the first environment sub-vector; and selecting candidate products that meet the following conditions as product recommendation results: the comprehensive dynamic matching score is higher than a preset recommendation standard threshold, and the product risk level is lower than the recommendation risk safety threshold.

[0012] A second aspect of this application provides a product recommendation device, comprising: a data acquisition module for acquiring user-authorized cross-domain behavior data and external environment data; a dynamic user state vector acquisition module for performing unified semantic encoding and semantic tendency analysis based on the cross-domain behavior data using a first major model to obtain a high-dimensional semantic vector and a behavior tendency vector, and fusing the high-dimensional semantic vector and the behavior tendency vector using time-aware attention to obtain a dynamic user state vector; an environmental feature vector acquisition module for performing natural language understanding and information extraction processing based on the external environment data using a second major model to obtain an environmental feature vector; and a product recommendation module for generating candidate product features and corresponding dynamic weight distributions based on the dynamic user state vector and the environmental feature vectors, and determining product recommendation results based on the candidate product features and the dynamic weight distributions.

[0013] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0014] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0015] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0016] According to embodiments of this application, by introducing a dual-model collaborative computing mechanism into the recommendation process, user behavior semantic modeling and market semantics are completed in parallel by the first and second large models, respectively. Time-aware attention is used to achieve efficient fusion of multi-source features, thereby reducing redundant computation and lag issues caused by isolated features and batch updates. Simultaneously, the time-aware attention mechanism and dynamic weight distribution generation algorithm enable the model to achieve rapid adaptive updates at the feature level, thus improving the computational efficiency and response speed of the recommendation engine. Furthermore, by fusing dynamic user state vectors and environmental feature vectors, bidirectional perception of user interest evolution and market fluctuations is achieved, enabling automatic adjustment of recommendation strategies during drastic market changes, generating results that balance personalization and risk safety. Moreover, the recommendation logic is based on interpretable semantic matching and weight distribution, allowing users to trace the reasons for recommendations, improving the transparency, trustworthiness, and user experience of the recommendation results. Attached Figure Description

[0017] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 The illustrations depict application scenarios of product recommendation methods, apparatus, devices, media, and program products according to embodiments of this application.

[0019] Figure 2 A flowchart illustrating a product recommendation method according to an embodiment of this application is shown schematically;

[0020] Figure 3 A schematic diagram illustrating the structure of a product recommendation device according to an embodiment of this application is shown; and

[0021] Figure 4 A block diagram schematically illustrates an electronic device suitable for implementing a product recommendation method according to an embodiment of this application. Detailed Implementation

[0022] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0024] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0025] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0026] With the rapid development of fintech, robo-advisors and financial product recommendation systems have become important tools for investors in asset allocation and risk management. However, against the backdrop of the deep integration of digitalization and behavioral finance, existing recommendation technologies are gradually revealing some limitations.

[0027] Currently, common financial product recommendation methods typically include collaborative filtering and content analysis-based methods, rule-matching-based expert systems, and time-series forecasting-based market analysis models. Collaborative filtering and content analysis-based methods analyze internal data such as users' transaction records, portfolio structures, or browsing behavior, combining algorithms like clustering and matrix factorization to recommend products purchased by similar groups to target users. Rule-matching systems rely on pre-defined financial logic or risk grading standards, generating recommendations based on the matching relationship between user tags and product tags. Time-series forecasting-based market analysis models introduce time series forecasting models to predict market trends or product volatility, assisting in recommending products with higher expected future returns. While these methods have achieved some success in their early stages, they still have structural shortcomings when facing complex and volatile market environments.

[0028] First, existing systems have limitations in terms of data dimensions, resulting in user profiles exhibiting a distinctly static nature. Most systems rely solely on structured data from within the institution, such as trading behavior, portfolio information, or risk assessment results, lacking utilization of external behavioral data that can reflect comprehensive user characteristics. For example, user information focus across different platforms, changes in consumption patterns, and financial information browsing trends—this multi-source data can more accurately reflect the evolution of investment preferences. However, due to a lack of cross-domain data fusion capabilities, existing systems struggle to construct comprehensive profiles that reflect users' dynamic needs.

[0029] Secondly, existing recommendation systems lack the ability to perceive and adapt to user behavior during market changes. Financial market fluctuations often trigger periodic adjustments in investors' risk preferences, but traditional systems typically make recommendations based on fixed rules or historical preferences, failing to update strategies according to changes in user behavior or real-time market characteristics. When the market experiences sharp fluctuations, this static decision-making approach may lead to recommendations that are out of sync with the user's current risk tolerance, thereby reducing the effectiveness and reliability of the recommendations.

[0030] Furthermore, the reasoning mechanisms of existing systems generally lack flexibility and interpretability. Whether it's collaborative filtering, rule matching, or predictive models, their internal reasoning processes are mostly static mappings, lacking causal reasoning and decision explanation capabilities, and unable to clearly explain "why this product is more suitable for a specific user in the current market situation." This "black box" recommendation mechanism struggles to maintain decision consistency and transparency in dynamic market environments, and also limits the system's adaptive optimization capabilities.

[0031] Based on this, embodiments of this application provide a product recommendation method, the method comprising: acquiring user-authorized cross-domain behavior data and external environment data; based on the cross-domain behavior data, performing unified semantic encoding and semantic tendency analysis using a first large model to obtain high-dimensional semantic vectors and behavior tendency vectors, fusing the high-dimensional semantic vectors and behavior tendency vectors using time-aware attention to obtain dynamic user state vectors; based on the external environment data, performing natural language understanding and information extraction processing using a second large model to obtain environmental feature vectors; and generating candidate product features and corresponding dynamic weight distributions based on the dynamic user state vectors and environmental feature vectors, and determining product recommendation results based on the candidate product features and the dynamic weight distributions. According to embodiments of this application, by introducing a dual-model collaborative computing mechanism in the recommendation process, user behavior semantic modeling and market semantics are completed in parallel by the first large model and the second large model, respectively, and time-aware attention is used to achieve efficient fusion of multi-source features, thereby reducing the problems of repetitive calculation and lag caused by isolated features and batch updates; at the same time, the time-aware attention mechanism and the dynamic weight distribution generation algorithm enable the model to achieve rapid adaptive updates at the feature level, thereby improving the computational efficiency and response speed of the recommendation engine. Furthermore, by fusing dynamic user state vectors and environmental feature vectors, the system achieves bidirectional perception of user interest evolution and market fluctuations. This enables it to automatically adjust recommendation strategies during periods of drastic market change, generating results that balance personalization and risk safety. Moreover, the recommendation logic is built on interpretable semantic matching and weight distribution, allowing users to trace the reasons for recommendations, thus enhancing the transparency, trustworthiness, and user experience of the recommendation results.

[0032] It should be noted that the product recommendation methods, apparatus, devices, media, and program products defined in this application relate to the application of large-scale models in the fintech field. They can be used in the fields of big data technology, artificial intelligence technology, and fintech, and can also be used in various other fields besides big data technology, artificial intelligence technology, and fintech. The application fields of the product recommendation methods, apparatus, devices, media, and program products provided in the embodiments of this application are not limited.

[0033] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0034] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all offer users corresponding operation entry points, allowing them to choose to agree to or reject the automated decision results; if the user chooses to reject, the process proceeds to expert decision-making. Here, "expert decision-making" refers to the decision-making activities of personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0035] Figure 1 The illustrations depict application scenarios of product recommendation methods, apparatus, devices, media, and program products according to embodiments of this application.

[0036] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0037] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0038] In the embodiments of this application, the first terminal device 101 can be an example of the first device, and the second terminal device 102 and / or the third terminal device 103 can be an example of at least one second device. The first device and the second device can communicate collaboratively through an internal client mechanism to implement the data distribution and rendering logic described in the product recommendation method.

[0039] In some embodiments, the first device and at least one second device may be different display modules, windows or screens on the same computing terminal (such as a host), or multiple physical devices that work together through a network, such as different client instances deployed on a desktop computer, tablet terminal or mobile device respectively.

[0040] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smart mobile terminals, tablet computers, laptop computers, and desktop computers.

[0041] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0042] It should be noted that the product recommendation method provided in this application embodiment can generally be executed by server 105. Correspondingly, the product recommendation device provided in this application embodiment can generally be located in server 105. The product recommendation method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the product recommendation device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0043] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0044] The following will be based on Figure 1 The described scene, through Figure 2 The product recommendation method of the disclosed embodiments is described in detail.

[0045] Figure 2 A flowchart illustrating a product recommendation method according to an embodiment of this application is shown.

[0046] like Figure 2 As shown, the product recommendation method 200 of this embodiment includes operations S210 to S240.

[0047] In operation S210, obtain user-authorized cross-domain behavior data and external environment data.

[0048] In the embodiments of this application, before performing data collection, the system first confirms the scope of data access and use through a user authorization mechanism. Users can explicitly choose whether to authorize the system to access specific types of data sources on the application interface or terminal device. After receiving the authorization request, the system can generate a unique user authorization credential and record the authorization scope, authorization time, and accessible data types in the permission management module for subsequent compliance auditing and access control.

[0049] For example, the user authorization process can be implemented through multi-level permission verification. Before accessing user data for the first time, the system will display an authorization statement and a description of the data's purpose to the user. The user must confirm authorization through biometrics (such as fingerprint or facial recognition) or two-factor authentication to prevent unauthorized access to the data. The system will only read user data within the authorization period. When the authorization expires or the user actively revokes the authorization, the system will immediately stop the collection and processing of relevant data and will securely store or destroy the stored user data to ensure user privacy and security.

[0050] In the embodiments of this application, the collection of cross-domain behavioral data may include both structured and unstructured data. Structured data, such as numerical transaction records, categorical asset distributions, and tagged consumption records, can be directly extracted through data interfaces; unstructured data, such as users' textual expressions on external platforms, comments on information content, or search terms, can be extracted and cleaned through text parsing and natural language processing modules. The system can automatically remove noisy information or low-confidence samples through preset data filtering strategies, such as deleting low-frequency interactions, abnormal consumption records, or content irrelevant to financial decisions, to ensure the quality and representativeness of the input data.

[0051] In some embodiments, to improve data integrity and timeliness, the system may employ an asynchronous acquisition and real-time update mechanism. Specifically, some static data (such as long-term behavior records and historical evaluation results) can be acquired through batch synchronization; while dynamic data (such as new text expressions, the latest information reading records, or instant query behaviors) can be continuously accessed through real-time data streams or message queues. Upon receiving new data, the system can automatically trigger data update events, enabling the model to automatically refresh feature states based on recent user behavior and environmental changes, ensuring the timeliness of the recommendation logic.

[0052] In the embodiments of this application, external environmental data may come from multiple publicly available information sources, such as regular information platforms, research-based analytical materials, industry updates, and publicly released index-type reference data. The system can perform preliminary analysis on this external environmental information, including identifying common structures, keyword paragraphs, and thematic relevance of the text content, thereby generating unstructured input that can be used for further processing. It should be noted that this analysis process focuses on the formalization and content expression of the text information itself, without involving judgment of the nature of events, thus ensuring the objectivity of the processing.

[0053] In some embodiments, the system can combine cloud and local computing resources to perform data collection and preprocessing in a distributed architecture. Cross-domain behavioral data can be encrypted by a local security proxy node and then uploaded to the cloud for unified formatting and feature annotation, thereby achieving multi-source information fusion while ensuring data privacy. To meet the needs of different application scenarios, the system can also customize the collection dimensions according to the user's authorized scope, such as extracting only financial behavior-related data, or expanding to auxiliary dimensions such as lifestyle and consumption patterns to support more refined recommendation strategies.

[0054] Preferably, the system can be configured with a data quality assessment module to detect the completeness, consistency, and timeliness of cross-domain behavioral data and external environment data. For example, if the system detects that the number of cross-domain data samples from users is insufficient or the time span is too short, it can trigger a data completion process to call historical caches or public datasets to fill in the missing parts. If external environment data experiences delays or abnormal fluctuations, the system can correct the input results through redundant source comparison and confidence weighting, thereby ensuring the stability and accuracy of the data entering the model.

[0055] In operation S220, based on the cross-domain behavior data, a unified semantic encoding and semantic tendency analysis are performed using the first major model to obtain a high-dimensional semantic vector and a behavior tendency vector. The high-dimensional semantic vector and the behavior tendency vector are then fused using time-aware attention to obtain a dynamic user state vector.

[0056] In the embodiments of this application, based on cross-domain behavior data authorized by the user, a first major model is used to perform unified semantic encoding and semantic tendency analysis on behavior information from different sources, so as to obtain high-dimensional semantic vectors and behavior tendency vectors that can comprehensively represent user characteristics.

[0057] In some embodiments, the system can leverage the semantic orientation analysis capabilities of the first major model to identify and quantify user behavior data. Specifically, the system can analyze the semantic direction of text expressions published by users on external platforms to identify the orientation features of their content expression in the semantic space. For example, the system can use a finely tuned semantic classification task head to perform semantic orientation classification, semantic strength estimation, or content orientation confidence calculation on the input text, thereby generating a corresponding vectorized representation. The aforementioned vectorized results can serve as behavioral orientation vectors to describe the directional features of different text content in the semantic dimension, enabling the model to perform more refined structured processing and dynamic feature fusion of cross-domain behavioral data in subsequent stages.

[0058] In the embodiments of this application, after obtaining high-dimensional semantic vectors and behavioral tendency vectors, the system can further utilize a time-aware attention mechanism to fuse multi-source semantic features to generate a dynamic user state vector that reflects the changing trends of user behavior. Specifically, the system identifies the update frequency and content differences of behavior records within different time windows through timestamp information, assigning higher weights to recently occurring behaviors or semantic expressions and applying decay factors to earlier behavior records to reflect the influence of different temporal information on the overall behavior pattern analysis. The system can employ a Transformer-based temporal embedding structure, an exponential decay function, or a sliding window weighting method to implement the above dynamic weight allocation, so that the generated user state vector reflects both long-term stable behavior patterns and recent changes in semantic features, thereby more accurately characterizing the temporal characteristics of behavioral data.

[0059] In some embodiments, to further enhance the expressive power of dynamic user state vectors, the system can adopt a multi-task learning framework, incorporating unified semantic encoding, semantic tendency analysis, and time series modeling into the multi-task structure of the first main model for joint training. For example, the backbone network can be responsible for generating unified semantic representations, while different task heads perform auxiliary tasks such as topic recognition, text pattern classification, and related content clustering, enabling the model to simultaneously learn behavioral content features and behavioral pattern change features during joint optimization, thereby improving the discriminability and stability of dynamic user state vectors.

[0060] In some embodiments, the system can adaptively control the recommendation triggering mechanism based on the update magnitude and stability of the dynamic user state vector. When a significant change in the feature pattern of the user state vector over time is detected (e.g., a new topic direction appears in the user's browsing or interaction content recently), the system can automatically trigger a new recommendation calculation process to dynamically adjust the candidate product set and weight distribution, ensuring that the recommendation results are consistent with the user's recent behavioral characteristics. This mechanism ensures the timeliness and responsiveness of the recommendation process, enabling the system to more accurately adapt to changing user behavior trends.

[0061] Preferably, the system can introduce a cross-modal adaptive alignment mechanism during the fusion process to comprehensively utilize complementary information from different types of data, such as textual behavior, numerical features, and interaction patterns. Specifically, textual data is encoded into semantic vectors using a language model, while structured data such as transaction records and numerical indicators are mapped into vector representations via a multi-layer perceptual network. Subsequently, the system employs a cross-modal attention structure during the fusion phase to model the correlation between features from different modalities, thereby capturing the interaction relationships between semantic expressions, behavioral records, and temporal features, and improving the expressive power of dynamic user state vectors in multi-source data fusion scenarios.

[0062] In operation S230, based on the external environment data, the second major model is used to perform natural language understanding and information extraction processing to obtain environmental feature vectors.

[0063] In the embodiments of this application, a second major model can be used to perform natural language understanding and information extraction processing on various types of publicly available text related to the target product, based on external environmental information data. Specifically, the system can perform sub-tasks such as word segmentation, syntactic analysis, named entity recognition, keyword extraction, and topic clustering on unstructured text from information platforms, analysis materials, and public index descriptions. It can identify representative topic fragments, key phrases, and their contextual dependencies from the original text, and represent the above information as computable text features through vectorized encoding, thereby forming a structured environmental text feature description for subsequent recommendation calculations.

[0064] In some embodiments, the system can perform multi-level semantic analysis and decoupling processing through a second major model to hierarchically represent environmental text features. Specifically, the model can distinguish between general content features in the external environment information that do not specifically point to any product category and content features related to the target product category, based on preset rules or training-obtained classification labels. It then constructs a first environmental sub-vector to represent general environmental information and a second environmental sub-vector to represent information related to the target product category. These two sub-vectors are combined in a unified semantic space to form an environmental feature vector, enabling the system to simultaneously consider overall environmental background information and text features related to the target product category in subsequent calculations.

[0065] In some embodiments, to improve the responsiveness of real-time market information, the system can combine a data streaming architecture with an incremental update mechanism to dynamically refresh the environmental feature vector. Specifically, the system can continuously receive external environmental information data streams from multiple information sources in the background. When it detects new or updated text related to the target product category (such as new public announcements, analytical materials, or news reports), it triggers the second major model to perform rapid re-parsing, semantic re-parsing, and feature updates on the relevant text. The newly generated text features can be written to or merged into the environmental feature vector in real time, enabling the recommendation system to reflect the latest changes in external environmental information in a timely manner without changing the existing overall architecture, thereby supporting timely updates to the recommendation strategy.

[0066] In some embodiments, the system can also perform weighted optimization on the feature dimensions in the environmental feature vector to highlight the semantic contributions under different time scales and topic categories. Specifically, the model can introduce temporal attention or temporal weighting mechanisms to assign relatively high weights to recently updated external environmental text features and relatively stable weights to long-term persistent topic information. This allows the generated environmental feature vector to reflect changes in external information over a short period of time while retaining the ability to describe long-term stable topics or background information, enabling subsequent recommendation calculations to maintain the stability of the overall feature representation while utilizing the latest information.

[0067] In some embodiments, the system can combine the second major model with a domain knowledge graph to enhance the accuracy and interpretability of environmental information extraction and representation. By introducing a knowledge graph structure, the system can align domain entities and their relationships (such as associations between industry categories, organization names, etc.) appearing in the text with predefined concept nodes, and aggregate and encode relevant text fragments accordingly. For concepts or relationships that appear repeatedly in the external environment and are strongly associated with the target product category, the model can assign them higher feature weights in the environmental feature vector to explicitly reflect the existence of these concept dimensions in the subsequent recommendation process, thereby improving the interpretability of the recommendation result generation process without inferring specific economic effects or market trends.

[0068] In some embodiments, to ensure data security and compliance, the system may introduce tiered caching and anonymization mechanisms during the collection and analysis of external environmental information. The second model processes only anonymized and standardized text data, and filters irrelevant or potentially biased information through a built-in content moderation module to ensure the security and standardization of the generated environmental feature vectors.

[0069] In operation S240, based on the dynamic user state vector and the environmental feature vector, candidate product features and corresponding dynamic weight distributions are generated, and product recommendation results are determined based on the candidate product features and the dynamic weight distributions.

[0070] In the embodiments of this application, candidate product features and corresponding dynamic weight distributions can be generated based on the aforementioned dynamic user state vector and environmental feature vector, thereby obtaining product recommendation results. For example, candidate product information containing multi-dimensional features can first be extracted from the product database, such as historical performance indicators, volatility-related indicators, liquidity indicators, term attributes, risk identification indicators, etc., and these structured features are input into the recommendation calculation unit. Subsequently, the system combines the dynamic user state vector and environmental feature vector as context input to generate a semantic fusion representation that characterizes the joint relationship between user-side feature information and external environment information. This fusion representation is used to calculate the matching degree between user features and the features of each candidate product, and a dynamic weight distribution is generated through a weight allocation mechanism to determine the relative priority of each candidate product in the recommendation process.

[0071] In some embodiments, reinforcement learning or policy modeling can be used to utilize large models or other policy networks as inference units, adaptively adjusting the recommendation strategy based on the joint representation of dynamic user state vectors and environmental feature vectors. During training, the model can use historical interaction feedback information (e.g., records of browsing, clicking, holding, or not adopting behaviors) as optimization criteria, gradually obtaining a more reasonable weight distribution through policy updates. In this way, the system can automatically learn the balance between user feature dimensions and environmental feature dimensions under different combinations of external environmental features. For example, when a certain type of volatility-related feature in the environmental feature vector is at a high level, the model can correspondingly reduce the weight of volatility-sensitive product features; when the environmental features are relatively stable and the user has a high level of attention to revenue-generating products in their historical records, the model can appropriately increase the weight of features related to revenue performance.

[0072] In some embodiments, the system can calculate a weighted similarity score based on the semantic matching results of the dynamic user state vector and the candidate product feature vector to quantify the degree of fit between each candidate product and the current feature context. The model can use weighted cosine similarity, attention matching function, or vector dot product-based scoring function to calculate a comprehensive dynamic matching score for all candidate products. Subsequently, based on preset recommendation standard thresholds and security control parameters, candidate products with a comprehensive dynamic matching score not lower than the preset standard and meeting internal risk control or compliance rules can be selected as the final recommendation list. To improve the transparency of the recommendation process, the system can also generate explanatory text containing the basis for the recommendation based on relevant features and calculation processes. For example, without involving individual emotions or psychological judgments, it can summarize the correspondence between a certain recommendation and some of the user's historical behavioral characteristics, some product attributes, and some external environmental information.

[0073] In some embodiments, to further enhance the stability and diversity of recommendations, the system may introduce product clustering and multi-objective optimization mechanisms. When calculating recommendation results, the model not only considers the matching degree between users and individual products, but also analyzes the correlation between candidate products to ensure that the final recommendation list is balanced in terms of risk level, industry distribution, and revenue structure.

[0074] In embodiments of this application, the system may also introduce an interpretability generation module in the recommendation output stage. This module, based on a generative language model, summarizes and organizes key decision factors in the recommendation results (such as specific topic-related features in some external environmental information, key semantic dimensions in dynamic user state vectors, and changes in contribution due to time weight adjustments), and generates user-oriented natural language explanatory text.

[0075] According to embodiments of this application, a user authorization mechanism enables compliant collection of multi-source data, ensuring the reliability, security, and controllability of data sources during the recommendation process. A first-level model is used to perform unified semantic encoding and semantic tendency analysis on cross-domain behavioral data, combined with time-aware attention to generate dynamic user state vectors, transforming user profiles from static label representations to continuously evolving dynamic semantic representations. A second-level model performs natural language understanding and information extraction on external environmental data, decoupling complex macroeconomic conditions, industry sentiment, and market events into structured environmental feature vectors, thus enabling the recommendation engine to have real-time perception capabilities of market fluctuations and external risks. Based on this, the dynamic user state vectors and environmental feature vectors are fused and calculated to generate candidate product features and dynamic weight distributions, and personalized, adaptive recommendation decisions are achieved through weighted matching and other methods. Through this multi-model collaborative mechanism, the system can achieve efficient, real-time, and interpretable recommendation reasoning in multimodal, multi-source heterogeneous data environments, not only improving the relevance and timeliness of recommendation results but also realizing intelligent risk adjustment and preference matching under complex market conditions.

[0076] The product recommendation method of this application will be specifically described below by way of preferred embodiments.

[0077] In the embodiments of this application, during the process of performing unified semantic encoding and semantic tendency analysis, the system can utilize a first major model to perform deep semantic understanding and unified feature representation on cross-domain behavioral data to obtain high-dimensional semantic vectors and behavioral tendency vectors. Specifically, textual and non-textual data can be extracted first from user-authorized data sources.

[0078] Subsequently, textual data can be directly input into the first large model, where the model's multi-layer semantic representation capabilities extract a first semantic feature vector reflecting the text's semantic content, thematic information, and expression direction. For non-textual data, descriptive processing can be used to transform numerical or categorical information into natural language descriptions that can be processed by large language models. For example, "Transaction amount: 10,000 yuan, product type: money market fund" can be transformed into "The user recently made a transaction related to a money market product worth approximately 10,000 yuan," ensuring consistency in the representation of different data modalities at the semantic level. After the transformed descriptive text is input into the same language model, a second semantic feature vector is obtained. This vector is used to characterize the semantic association features behind structured behavior. Furthermore, the system can fuse the first and second semantic feature vectors in a unified semantic space, obtaining a high-dimensional semantic vector that comprehensively represents the characteristics of cross-domain behavioral data through multi-layer attention aggregation or feature concatenation.

[0079] In the embodiments of this application, to further extract semantic tendency features reflected in behavioral data, the system can utilize the semantic tendency analysis task head integrated in the first major model to perform semantic direction recognition and intensity estimation on the input text data. This task head can be fine-tuned for tasks such as semantic tendency classification and content attitude judgment, outputting vectorized results representing the semantic tendency direction of the text (e.g., preset categories such as positive, neutral, or conservative) and the degree of semantic intensity. The system transforms these results into behavioral tendency vectors, which reflect the expression direction and focus of text content under different themes or contexts. This enables the model to identify changes in content distribution and attention across various thematic dimensions from the behavioral data, thereby better personalizing the matching of user-related thematic features in subsequent recommendation calculations.

[0080] In some embodiments, the system can further optimize the consistency between semantic encoding and tendency analysis through joint training. Specifically, the semantic encoding network and the semantic tendency analysis task head can share the underlying language representation layer and be jointly optimized through a multi-task loss function. This enables the model to improve its ability to distinguish different semantic tendency categories and content patterns while extracting semantic features, thereby enhancing the correlation and consistency between high-dimensional semantic vectors and behavioral tendency vectors.

[0081] Through the above process, the system can achieve semantic unification and tendency recognition based on multi-source heterogeneous behavioral data, transforming the originally scattered and structurally diverse data into continuous, high-dimensional, and computable semantic vector representations. This upgrades user profiles from static label-based modeling to a dynamic embedded structure that reflects the changes in semantic features and content attention over time, thereby improving the accuracy, expressiveness, and dynamic response capability of user behavior modeling.

[0082] Preferably, the interest intensity quantification and time-aware encoding of the first model can also be used to characterize the change patterns of user-related behavioral features over time.

[0083] In the embodiments of this application, when the system performs interest intensity quantification and time-aware encoding based on the first major model, it can first use the first major model to perform semantic analysis on the interactive content involved in cross-domain behavior data, and use the contextual understanding capability of the pre-trained language model to identify the domain topics and related semantic features corresponding to each behavior. During the analysis process, the model can extract key domain entities and their associated sentence fragments and contextual features, thereby establishing a correspondence between user behavior and semantic topics.

[0084] Furthermore, user attention intensity scores in various domains can be calculated based on the semantic analysis results described above. This calculation process can comprehensively consider multiple factors such as the frequency of behavior occurrence, interaction duration, content relevance, and semantic matching confidence. For example, if a user's browsing and interaction records under a certain topic are significantly more recent than those under other topics, the system can assign a higher weight to that topic when calculating domain attention intensity. The first major model can use a multilayer perceptual network or attention aggregation structure to weightedly integrate these features to generate a domain interest intensity vector that reflects the user's level of attention and participation in different domains.

[0085] In another embodiment, to reflect the changing patterns of user-related behavioral characteristics over time, the system encodes the timestamp information in cross-domain behavioral data. Specifically, a time decay function or a time embedding encoding mechanism can be applied to historical behavioral records, so that behaviors closer to the current moment have a greater influence weight in the model than earlier behaviors. For example, the system can use an exponential decay function to achieve time weighting, or introduce time embedding vectors in the Transformer architecture to involve time information in attention calculations, thereby capturing the dependencies of behavioral events across different time periods. In this way, the system can generate time-decayed interest vectors that characterize the temporal change patterns of behavioral features, describing the relative relationship between recent behavioral features and long-term stable behavioral patterns.

[0086] In some embodiments, the system can also combine the interaction characteristics of domain interest intensity vectors and time-decaying interest vectors to model the differences in the distribution of user attention to topics within a specific time window. The model can compare the current attention intensity distribution with historical statistical distributions to analyze the direction and magnitude of changes in attention focus across different domains, thereby identifying recently emerging high-frequency topics or newly enhanced attention directions. For example, when the system detects that a user has repeatedly accessed content related to a new topic in a recent period, but that topic has a low proportion in long-term statistics, the system can increase the weight of that topic in the domain interest intensity vector through a time-weighted mechanism to ensure that dynamic user representation can reflect changes in this attention direction in a timely manner.

[0087] Through the above process, the system uses the first major model to complete the modeling transformation from cross-domain behavioral data to interest intensity quantification and time-aware encoding, so that user interest features have both semantic interpretability and temporal dynamics.

[0088] In the embodiments of this application, the system can input high-dimensional semantic vectors, behavioral tendency vectors, domain interest intensity vectors, and time decay interest vectors into a time-aware attention network for feature fusion to form a dynamic user state vector that can comprehensively reflect different temporal behavioral characteristics.

[0089] Optionally, the system can also introduce a multi-head interaction mechanism in the time-aware attention network, and achieve multi-layer fusion of cross-temporal features in a unified semantic space by computing the semantic relevance and interest intensity association under different time windows in parallel.

[0090] Through the above process, the system utilizes the first major model to achieve quantitative modeling of interest intensity-related features and dynamic fusion of time-aware features. This transforms user-related behavioral features from traditional static label-based expressions into semantic temporal structures that can be updated with input data, thereby improving the continuity and fine-grained expressive ability of user-related feature modeling and providing a more robust feature foundation for subsequent strategy decisions and product recommendations.

[0091] In the embodiments of this application, when the system performs external environment data processing based on the second major model, it can perform semantic parsing and structured extraction on data sources from multiple sources to generate environmental feature vectors that reflect the changing trends of the external environment. Specifically, the system can first receive relevant unstructured text data and input it into the pre-trained or fine-tuned second major model. The second major model has topic-level natural language understanding capabilities, enabling it to perform hierarchical semantic parsing of text content and identify topic-based semantic elements and contextual relationships in the text.

[0092] After identifying the environmental semantic elements and their relationships, the system can further utilize a second major model to perform grouped representation of semantic features, distinguishing and structurally representing information at different levels described in the text. Specifically, the system can employ contextual clustering and semantic pattern recognition methods to categorize information reflecting the overall environmental background into the first category of environmental features, used to characterize general environmental information applicable to most product categories; simultaneously, semantic features highly relevant to certain product categories or industry themes are categorized into the second category of environmental features, used to reflect information fragments related to specific categories.

[0093] Furthermore, the aforementioned signals can be numerically encoded to enable a computable quantized representation in the semantic space. Specifically, the second major model can use embedding vectorization to map the first and second environment sub-vectors into continuous vector forms, and generate corresponding confidence weights based on the context. To achieve a unified representation across texts and data sources, the model maps both types of signals into a unified semantic space during the output stage to form a structured environment feature vector.

[0094] Furthermore, the system can numerically encode the aforementioned semantic features of different categories to achieve a computable quantized representation in a unified semantic space. Specifically, the second major model uses a vectorized embedding method to map the first and second environment sub-vectors into continuous vector forms, and assigns corresponding confidence information to these vectors based on contextual semantics. To achieve consistent representation of data from different sources, the system can map the two types of environment features to a unified semantic space during the vector output stage, thereby constructing a structured environment feature vector.

[0095] In some embodiments, the system can dynamically update the environmental feature vector using time series modeling techniques. The second major model performs incremental semantic analysis on external environmental text over continuous time periods, identifying trends in semantic elements' frequency of description, thematic consistency, or relevance. For example, when a particular theme frequently appears in multiple data sources, the system can appropriately increase the relative weight of that feature in the environmental feature vector, thereby reflecting the activity level of that theme in external information without making risk judgments or value assessments.

[0096] Through the above process, the system utilizes the second major model to extract and group structural semantic features from multi-source external environment text, enabling the environmental feature vector to simultaneously represent general background information and textual semantic information related to specific product categories. This supports the recommendation system in achieving dynamic and adaptive feature fusion calculation under different external environment conditions, improving the model's adaptability and scalability in complex data environments.

[0097] It should be noted that although the first and second large models in the embodiments of this application both belong to the application scope of large language models, they have a clear division of labor and complementary relationship in terms of functional positioning, input features, training objectives and output format.

[0098] The first major model primarily focuses on user-side information modeling tasks, aiming to perform semantic understanding and user feature abstraction on user-authorized cross-domain behavioral data. This model emphasizes individualized semantic modeling and contextualized intent recognition. By uniformly semantically encoding and analyzing the user's textual and non-textual behaviors, it generates high-dimensional semantic and behavioral tendency vectors reflecting user interests, risk preferences, and behavioral trends. To enhance temporal sensitivity, the first major model can further incorporate a time-aware attention mechanism or a temporal embedding module to weightedly fuse short-term fluctuations and long-term preferences, thereby forming a dynamically updatable user state vector. Therefore, the output of the first major model focuses on expressing "who the user is, what they are currently thinking, and where their recent behavior is heading," and its results form the semantic basis for subsequent recommendation decisions.

[0099] The second major model primarily targets tasks related to understanding the external market environment and extracting information. It is mainly used for semantic parsing and structural transformation of external environment data. The core objective of this model is to understand the external semantic background and the evolution of risk factors. Through tasks such as topic-level natural language understanding, named entity recognition, and signal decoupling, it identifies and extracts the first and second environmental sub-vectors that influence the trend of the external environment, thereby generating quantitative environmental feature vectors.

[0100] Preferably, the first and second models can also be related. This relationship manifests in the fact that the first and second models can interact and dynamically fuse features within a unified semantic space. Specifically, the dynamic user state vector generated by the first model and the environmental feature vector generated by the second model can be used as joint inputs, driving the generation and optimization of recommendation strategies through semantic concatenation and weighted matching mechanisms. This dual-model collaborative structure enables the recommendation system to establish a mapping between user semantic and market semantic information, thus maintaining the accuracy and stability of personalized recommendations even under rapidly changing external environments.

[0101] In some embodiments, during the process of generating candidate product features and dynamic weight distributions based on dynamic user state vectors and environmental feature vectors, the system can first retrieve relevant candidate product features from a product database based on the interest topics, risk preferences, and behavioral tendencies contained in the dynamic user state vectors. The product database may include structured attribute information of different types of financial products, such as return volatility, risk level, industry affiliation, asset class, and historical performance. By performing semantic similarity calculations on the dynamic user state vectors and product feature vectors in a unified semantic space, the system can filter out a set of candidate product features most relevant to the user's current state, forming a preliminary product matching candidate pool.

[0102] Furthermore, the system can perform matching calculations on the aforementioned candidate product feature set to obtain the domain-based weight for each candidate product. Specifically, the system can employ matching algorithms based on vector dot product, cosine similarity, or weighted attention mechanisms to measure the semantic consistency between the candidate product feature vector and the user state vector. In this process, the semantic overlap, risk feature matching degree, and temporal relevance between the user's focus topics and product domain labels can be emphasized. For example, when the user's state vector shows a high preference for stable returns, the domain-based weight of bond or money market products can be increased in the matching calculation; conversely, when the user recently focuses on emerging industries and exhibits a positive investment tendency, the weight of growth or thematic products can be increased accordingly. Through this calculation, the system obtains the importance ranking of various candidate products within the current semantic and behavioral context.

[0103] Preferably, a dual-signal nonlinear correction function can be used to adjust the aforementioned domain-based weights to achieve adaptive adjustment of recommendation weights as external environmental information changes. Specifically, the system performs hierarchical correction on the weight distribution of candidate products based on two distinct components—the first environmental sub-vector and the second environmental sub-vector—generated by the aforementioned second major model. On the one hand, the first environmental sub-vector is used for global weight adjustment of the entire candidate set to reflect the influence of general environmental information in the overall recommendation process. On the other hand, the second environmental sub-vector is used for local adjustment of basic weights related to specific themes or product categories, thereby highlighting product categories that better match the current environmental text features while maintaining overall structural stability.

[0104] In some embodiments, the system can employ a smoothing constraint mechanism to control the rate of weight change based on the dynamic amplitude of the dual-signal correction, in order to prevent the recommendation strategy from deviating drastically due to short-term market fluctuations. This smoothing constraint mechanism can be implemented based on a time-weighted average or exponential smoothing model, ensuring that the dynamic weight distribution is updated smoothly over continuous time periods, thereby guaranteeing the interpretability and stability of the recommendation results.

[0105] Through the above process, the system achieves multi-layer fusion of user-side semantic features, external environment information features, and candidate product feature weights, thereby generating a dynamic weight distribution that can adaptively adjust as input features change. This dynamic weight distribution comprehensively considers the temporal changes of user-related behavioral features, the semantic structure changes of external environment information, and the correlation of each candidate product feature in different contexts, enabling the recommendation strategy to be automatically and continuously updated based on the coupling relationship between features.

[0106] In the embodiments of this application, when the system adjusts the domain-based weights based on a dual-signal nonlinear correction function, it can use a first environment sub-vector and a second environment sub-vector to perform global and local corrections on the weight distribution of candidate products, respectively, so that the recommendation weights can be adaptively updated according to changes in the content of external environment information. Specifically, the system first obtains a first environment sub-vector, which is used to represent general semantic content from external environment information, such as topic information that appears in texts from multiple sources, has a wide range of descriptions, and has holistic characteristics. When the semantic features corresponding to the first environment sub-vector appear more frequently or have higher expression intensity in a short period of time, the dual-signal nonlinear correction function can apply a global scaling to the domain-based weights, so that the overall recommendation weights are appropriately adjusted on a unified dimension to maintain the sensitivity and stability of the recommendation process to changes in external general semantics.

[0107] Meanwhile, the system can utilize the second environment sub-vector within the environment feature vector to locally adjust the weights of specific product categories. The second environment sub-vector represents semantic fragment features strongly correlated with a particular product or industry theme in the external environment information; for example, frequently occurring descriptive content in publicly available text related to a specific theme or category. When such semantic features are more prominent in recent external text, the correction function can perform non-linear enhancement on the corresponding domain's weight dimension, for example, through local activation or amplification mechanisms to ensure the category's weight aligns with its semantic relevance in the text. Conversely, when the frequency of a particular theme decreases in external text or related content diminishes, the system can automatically reduce its relative proportion in the recommendation weights, achieving localized dimensional suppression adjustment.

[0108] In some embodiments, to achieve coordinated control of global and local corrections, the system can introduce weight adjustment coefficients into the dual-signal nonlinear correction function to dynamically control the influence ratio of the two environment sub-vectors on the result. For example, when the general semantic features of external information undergo significant content changes in a short period, the system can enhance the global correction effect of the first environment sub-vector to improve the adaptability of the recommendation weights to changes in overall environment features; conversely, when the external text has higher content density or enhanced semantic consistency under a certain topic, the system can increase the local correction weight of the second environment sub-vector to more accurately represent the content features related to that topic. This function can be implemented through a multilayer perceptron or attention mechanism to make the correction process trainable and interpretable.

[0109] Through the above process, the system implements a two-layer semantic correction mechanism in the calculation of recommendation weights. This mechanism responds to general semantic changes in external environmental information on a global level, and reflects semantic changes in content related to specific topics on a local level. Global correction enhances the system's overall stability when external environmental content changes rapidly, while local correction improves the system's ability to identify subdivided topic features. The synergistic effect of these two mechanisms allows the generated dynamic weight distribution to maintain flexibility and consistency in a changing information environment, thereby improving the adaptability and overall practical value of the recommendation strategy.

[0110] In the embodiments of this application, during the process of determining product recommendation results based on candidate product features and dynamic weight distribution, the system can utilize the dynamic weight distribution to weight and adjust the candidate product features to obtain a comprehensive dynamic matching score for each candidate product. Specifically, the user's dynamic state vector, the external environment's environmental feature vector, and the candidate product's structured feature vector can be fused and calculated in a unified semantic space, and weights can be assigned to different feature dimensions according to the dynamic weight distribution. Through methods such as weighted cosine similarity, attention matching function, or embedding vector dot product, the system can comprehensively measure the degree of matching between the user's current focus topic, changes in external environment text features, and candidate product attributes, thereby generating a comprehensive score for each candidate product that is dynamically updated over time.

[0111] Furthermore, to ensure that the recommendations are generated within a controllable risk range, the system determines a recommendation risk safety threshold based on the first environment sub-vector obtained from the second largest model. This threshold quantifies the overall semantic variation presented in the current external environment data. For example, when the first environment sub-vector exhibits high consistency or significant semantic variation across multiple time segments, the system can lower the recommendation risk safety threshold, making the recommendations more biased towards robust products. Conversely, when the first environment sub-vector presents relatively stable content with minimal variation, the system can appropriately increase the recommendation risk safety threshold to support a wider range of products.

[0112] In the embodiments of this application, after generating a comprehensive dynamic matching score and a risk safety threshold, the system can perform multi-dimensional screening of all candidate products. The screening logic may include two conditions: first, the comprehensive dynamic matching score of the candidate product must be higher than a preset recommendation standard threshold to ensure that the recommendation results are highly relevant to user needs; second, the risk level of the candidate product must be lower than the recommendation risk safety threshold calculated by the system to ensure the robustness of the recommendation and risk control capabilities.

[0113] In some embodiments, the system can further sort and interpret the filtered product results. Specifically, the system can prioritize candidate products according to their comprehensive dynamic matching scores from high to low, and call the language generation module to generate recommendation explanation text, providing a natural language explanation of the recommendation logic. For example, the system can output, "Due to the recent increase in the proportion of content related to the theme of stable returns in external environmental texts, and your recent behavioral characteristics showing a stable preference, we recommend that you pay attention to product xx." In this way, the system not only provides actionable product recommendation results, but also enhances the transparency of the recommendation process and user trust.

[0114] Through the above process, the system achieves two-dimensional dynamic constraints and optimization of the recommendation results by comprehensively calculating the dynamic matching score and risk safety threshold. This mechanism enables the recommendation results to simultaneously meet the requirements of personalized relevance and risk controllability, improving the accuracy, robustness, and interpretability of the recommendations.

[0115] According to embodiments of this application, a deep integration of user-side information understanding and external environment semantic understanding is achieved by constructing an intelligent product recommendation method based on collaborative reasoning using two major models. With user authorization, the system collects cross-domain behavioral data and external environment data. The first major model performs semantic encoding, behavioral tendency recognition, and time-aware modeling to generate a continuously updated dynamic user state vector, upgrading the user profile from a static, label-based representation to a semantic structure reflecting temporal changes. Simultaneously, the second major model performs hierarchical semantic understanding and element decoupling on text content from external information sources, extracting a first environment sub-vector representing the overall environmental change trend and a second environment sub-vector describing the fluctuations in content within specific sub-domains. This forms a quantifiable environmental feature vector, enabling real-time perception of environmental semantic changes. The system further constructs candidate product features based on the dynamic user state vector and environmental feature vector, generating a dynamic weight distribution that adjusts with changes in external data. A dual-signal nonlinear correction function coordinates the influence of user preferences and environmental trends, thereby calculating a comprehensive dynamic matching score and a screening threshold, ultimately selecting recommendation results that combine high matching degree and adaptability to environmental changes.

[0116] Corresponding to the above product recommendation method, embodiments of this application also provide a product recommendation device.

[0117] Figure 3 A schematic block diagram of a product recommendation device according to an embodiment of this application is shown.

[0118] like Figure 3 As shown, the product recommendation device 300 in this embodiment includes a data acquisition module 310, a dynamic user state vector acquisition module 320, an environmental feature vector acquisition module 330, and a product recommendation module 340.

[0119] The data acquisition module 310 can be used to acquire cross-domain behavioral data and external environment data authorized by the user. In one embodiment, the data acquisition module 310 can be used to perform the operation S210 described above, which will not be repeated here.

[0120] The dynamic user state vector acquisition module 320 can be used to perform unified semantic encoding and semantic tendency analysis based on the cross-domain behavior data using the first major model to obtain high-dimensional semantic vectors and behavioral tendency vectors. Time-aware attention is then used to fuse the high-dimensional semantic vectors and the behavioral tendency vectors to obtain a dynamic user state vector. In one embodiment, the dynamic user state vector acquisition module 320 can be used to execute the operation S220 described above, which will not be repeated here.

[0121] The environmental feature vector acquisition module 330 can be used to obtain environmental feature vectors based on the external environment data and by performing natural language understanding and information extraction processing using the second major model. In one embodiment, the environmental feature vector acquisition module 330 can be used to perform the operation S230 described above, which will not be repeated here.

[0122] The product recommendation module 340 can be used to generate candidate product features and corresponding dynamic weight distributions based on the dynamic user state vector and the environmental feature vector, and determine the product recommendation result based on the candidate product features and the dynamic weight distribution. In one embodiment, the product recommendation module 340 can be used to perform the operation S240 described above, which will not be repeated here.

[0123] According to an embodiment of this application, the dynamic user state vector acquisition module 320 can also be used to acquire textual data and non-textual data in the cross-domain behavior data; input the textual data into the first large-scale language model to obtain a first semantic feature vector; input the non-textual data into the first large-scale language model after descriptive processing to obtain a second semantic feature vector; perform feature fusion of the first semantic feature vector and the second semantic feature vector in a unified semantic space to obtain the high-dimensional semantic vector; and perform semantic tendency analysis on the textual data based on the semantic tendency analysis task head of the first large-scale language model to generate the behavior tendency vector.

[0124] According to an embodiment of this application, the dynamic user state vector acquisition module 320 can also be used to perform interest intensity quantization and time-aware encoding on the cross-domain behavior data using the first large model to obtain a domain interest intensity vector and a time-decaying interest vector; and to use time-aware attention to perform weighted fusion of the high-dimensional semantic vector, the behavior tendency vector, the domain interest intensity vector and the time-decaying interest vector to obtain a dynamic user state vector.

[0125] According to an embodiment of this application, the dynamic user state vector acquisition module 320 can also be used to perform semantic analysis on the interaction content of different domains in the cross-domain behavior data using the first large model, calculate the user's interest intensity score in each domain based on the semantic analysis results, and generate the corresponding domain interest intensity vector; and based on the timestamp information of the cross-domain behavior data, use a time decay function or a time embedding coding mechanism to perform time-series feature modeling on historical behavior records to obtain the time decay interest vector.

[0126] According to an embodiment of this application, the environmental feature vector acquisition module 330 can also be used to perform natural language understanding based on the external environment data using the second large model, identify environmental semantic elements and corresponding relationships; decouple the environmental semantic elements and corresponding relationships by signal to obtain a first environmental sub-vector and a second environmental sub-vector, wherein the first environmental sub-vector is used to reflect global environmental changes and the second environmental sub-vector is used to reflect local environmental changes related to the target product category; and use the second large model to numerically encode the first environmental sub-vector and the second environmental sub-vector respectively, and combine them in a unified semantic space to generate the environmental feature vector.

[0127] According to an embodiment of this application, the product recommendation module 340 can also be used to extract candidate product features from the product database based on the dynamic user state vector; perform matching calculations on the candidate product features to obtain corresponding domain basic weights; and use a dual-signal nonlinear correction function to correct the domain basic weights to obtain the dynamic weight distribution.

[0128] According to an embodiment of this application, the product recommendation module 340 can also be used to globally modify the domain basic weights using the first environment sub-vector; and to locally modify the domain basic weights of the target product category using the second environment sub-vector.

[0129] According to an embodiment of this application, the product recommendation module 340 can also be used to adjust the features of the candidate products by weighting them using the dynamic weight distribution to obtain a comprehensive dynamic matching score for each candidate product; determine a recommendation risk safety threshold based on the first environment sub-vector; and select candidate products that meet the following conditions as product recommendation results: the comprehensive dynamic matching score is higher than a preset recommendation standard threshold, and the product risk level is lower than the recommendation risk safety threshold.

[0130] According to embodiments of this application, any multiple modules among the data acquisition module 310, dynamic user state vector acquisition module 320, environmental feature vector acquisition module 330, and product recommendation module 340 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 310, dynamic user state vector acquisition module 320, environmental feature vector acquisition module 330, and product recommendation module 340 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the data acquisition module 310, the dynamic user state vector acquisition module 320, the environmental feature vector acquisition module 330, and the product recommendation module 340 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0131] Figure 4 A block diagram schematically illustrates an electronic device suitable for implementing a product recommendation method according to an embodiment of this application.

[0132] like Figure 4 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0133] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0134] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0135] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0136] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0137] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the product recommendation method provided in the embodiments of this application.

[0138] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0139] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0140] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0141] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0142] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0143] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0144] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A product recommendation method, characterized in that, The method includes: Obtain user-authorized cross-domain behavioral data and external environment data; Based on the cross-domain behavior data, the first major model is used to perform unified semantic encoding and semantic tendency analysis to obtain high-dimensional semantic vectors and behavior tendency vectors. Time-aware attention is then used to fuse the high-dimensional semantic vectors and the behavior tendency vectors to obtain dynamic user state vectors. Based on the aforementioned external environment data, a second major model is used for natural language understanding and information extraction to obtain environmental feature vectors; and Based on the dynamic user state vector and the environmental feature vector, candidate product features and corresponding dynamic weight distributions are generated, and product recommendation results are determined based on the candidate product features and the dynamic weight distributions.

2. The method according to claim 1, characterized in that, The first major model is used for unified semantic encoding and semantic tendency analysis to obtain high-dimensional semantic vectors and behavioral tendency vectors, including: Obtain text-based data and non-text-based data from the cross-domain behavior data; The textual data is input into the first large-scale language model to obtain a first semantic feature vector; the non-textual data is processed descriptively and then input into the first large-scale language model to obtain a second semantic feature vector; the first semantic feature vector and the second semantic feature vector are fused in a unified semantic space to obtain the high-dimensional semantic vector; and The semantic tendency analysis task head based on the first large-scale language model performs semantic tendency analysis on the text data to generate the behavior tendency vector.

3. The method according to claim 1, characterized in that, The method further includes: The first large model is used to quantize interest intensity and perform time-aware encoding on the cross-domain behavior data to obtain a domain interest intensity vector and a time-decaying interest vector; and By using time-aware attention, the high-dimensional semantic vector, the behavioral tendency vector, the domain interest intensity vector, and the time-decayed interest vector are weighted and fused to obtain a dynamic user state vector.

4. The method according to claim 3, characterized in that, The step of using the first large model to quantify the interest intensity and time-aware encoding of the cross-domain behavior data includes: The first large model is used to perform semantic analysis on the interaction content of different domains in the cross-domain behavioral data. Based on the semantic analysis results, the user's interest intensity score in each domain is calculated to generate the corresponding domain interest intensity vector; and Based on the timestamp information of the cross-domain behavior data, the time decay function or time embedding coding mechanism is used to model the temporal features of historical behavior records to obtain the time decay interest vector.

5. The method according to claim 1, characterized in that, Based on the external environment data, the second major model is used for natural language understanding and information extraction to obtain environmental feature vectors, including: Based on the external environment data, the second major model is used to perform natural language understanding to identify environmental semantic elements and corresponding relationships. Signal decoupling is performed on the environmental semantic elements and their corresponding relationships to obtain a first environmental sub-vector and a second environmental sub-vector. The first environmental sub-vector reflects global environmental changes, and the second environmental sub-vector reflects local environmental changes related to the target product category. The first environment sub-vector and the second environment sub-vector are numerically encoded using the second major model and then combined and represented in a unified semantic space to generate the environment feature vector.

6. The method according to claim 5, characterized in that, The step of generating candidate product features and corresponding dynamic weight distributions based on the dynamic user state vector and the environmental feature vector includes: Based on the dynamic user state vector, candidate product features are extracted from the product database; The candidate product features are matched and calculated to obtain the corresponding domain-based weights; as well as The dynamic weight distribution is obtained by correcting the basic weights of the domain using a dual-signal nonlinear correction function.

7. The method according to claim 6, characterized in that, The method of correcting the domain-based weights using a dual-signal nonlinear correction function includes: The domain-based weights are globally modified using the first environment sub-vector; as well as The domain-based weights of the target product category are locally modified using the second environment sub-vector.

8. The method according to any one of claims 5 to 7, characterized in that, The process of determining product recommendation results based on the candidate product features and the dynamic weight distribution includes: The dynamic weight distribution is used to adjust the features of the candidate products to obtain a comprehensive dynamic matching score for each candidate product. Based on the first environment sub-vector, a recommended risk safety threshold is determined; and Candidate products that meet the following conditions are selected as product recommendation results: the comprehensive dynamic matching score is higher than the preset recommendation standard threshold, and the product risk level is lower than the recommendation risk safety threshold.

9. A product recommendation device, characterized in that, The device includes: The data acquisition module is used to: acquire user-authorized cross-domain behavioral data and external environment data; The dynamic user state vector acquisition module is used to: based on the cross-domain behavior data, perform unified semantic encoding and semantic tendency analysis using the first major model to obtain high-dimensional semantic vectors and behavior tendency vectors, and use time-aware attention to fuse the high-dimensional semantic vectors and the behavior tendency vectors to obtain dynamic user state vectors. The environmental feature vector acquisition module is used to: obtain environmental feature vectors based on the external environment data by performing natural language understanding and information extraction processing using a second major model; and The product recommendation module is used to: generate candidate product features and corresponding dynamic weight distributions based on the dynamic user state vector and the environmental feature vector, and determine the product recommendation result based on the candidate product features and the dynamic weight distribution.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. 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 steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

Images