A global data value-added service system supporting multi-domain trusted transactions

Abstract

The application relates to the technical field of data services, in particular to a global data value-added service system supporting multi-field trusted transactions. The system comprises a data acquisition module for acquiring real-time context data of a user; a cold start portrait construction module for combining the context data into an initial portrait when there is no historical behavior; a graph neural network inference module for aggregating group preference features to output a preference vector and aggregating negative aversion features to generate a repulsion vector; a real-time feedback self-calibration module for parallel display of recommended items on a first screen and switching subsequent display according to the spatiotemporal correlation of operation events; and a recommendation output module for outputting a recommendation result according to the initial portrait, the preference vector and the repulsion vector. Through the fusion of group positive and negative preferences and behavior entropy self-calibration, the application realizes accurate cold start recommendation and intention self-adaptive adjustment of a user without historical data.

Description

Technical Field

[0001] This invention relates to the field of data service technology, and in particular to a comprehensive data value-added service system that supports trusted transactions across multiple domains. Background Technology

[0002] Currently, with the development of the digital economy and comprehensive data services, providing accurate personalized recommendations to new users across multiple domains and channels has become one of the core challenges of data value-added service systems. Traditional recommendation systems, when faced with new users lacking historical behavioral data, typically employ default recommendation strategies based on popular items or simple rule-based cold-start methods. These approaches ignore the real-time contextual information when users enter the system, resulting in significant discrepancies between the recommendation results and the user's true intent. Some systems attempt to mine potential associations between users and items using algorithms such as collaborative filtering or matrix factorization; however, in the absence of historical interaction records of target users, these methods face severe data sparsity problems, making it difficult to generate effective personalized recommendations.

[0003] In recent years, graph neural networks have been introduced into the field of recommender systems to model higher-order relationships between users and items. However, existing graph neural network recommendation schemes typically only aggregate the positive preference features of neighboring nodes, ignoring negative information that users collectively dislike. This results in a lack of exclusionary constraints in the recommendation results, making it difficult to effectively avoid recommendations that users are not interested in. Furthermore, existing systems lack real-time self-calibration capabilities after the initial screen recommendation, failing to dynamically adjust subsequent recommendation strategies based on users' immediate actions. When the initial screen recommendation does not match user expectations, it can easily lead to user churn.

[0004] Chinese Patent Publication No. CN120066556A discloses a cold-start service recommendation method based on a multi-strategy pre-trained model, comprising the following steps: First, constructing a hypergraph based on labels, extracting cold-start service features, and constructing service embeddings; Second, obtaining enhanced service embeddings based on hypergraph contrastive learning using singular value decomposition; Third, obtaining content-optimized service embeddings based on contrastive learning using descriptive text; Fourth, fine-tuning the cold-start recommendation model through downstream recommendation tasks.

[0005] Therefore, the existing technology has the following problems: The system does not incorporate real-time contextual information of users entering the system as a basis for constructing cold start profiles, and it lacks a self-calibration mechanism based on negative group preference constraints and real-time operational behavior feedback, resulting in recommendation results that lack specificity and dynamic adaptability. Summary of the Invention

[0006] To address this, the present invention provides a full-domain data value-added service system that supports trusted transactions across multiple domains. This system overcomes the problems in existing technologies, such as the lack of real-time context information when users enter the system as the basis for constructing cold-start profiles, and the absence of a self-calibration mechanism based on negative group preference constraints and real-time operational behavior feedback, which leads to a lack of targeted and dynamic adaptation capabilities in recommendation results.

[0007] To achieve the above objectives, the present invention provides a comprehensive data value-added service system supporting trusted transactions across multiple domains, comprising: The data acquisition module is used to acquire real-time context data when a user enters the system. The real-time context data includes the user's entry channel identifier and the current scene tag. The cold start profile building module is used to combine the real-time context data into an initial intent tag vector in response to the detection that the target user has no historical behavior data, as the initial cold start profile of the target user. The graph neural network inference module is used to access the target user as a new node into the real-time relationship graph, and to aggregate the preference features of related group nodes that have the same channel identifier and / or the same scene label as the target user through the graph neural network, and output the group aggregated preference vector; it is also used to aggregate the negative aversion features of the related group nodes to generate a group rejection vector. The real-time feedback self-calibration module is configured to display multiple candidate recommendation items in parallel during the target user's first screen interaction, capture the operation event data triggered by the target user for the multiple candidate recommendation items in the current session, determine the adjustment strategy of the candidate recommendation items based on the temporal and spatial correlation between each operation event in the operation event data, and switch the subsequently displayed candidate recommendation items according to the adjustment strategy. The recommendation output module is used to determine and output the final recommendation result to the target user based on the initial cold start profile, the group aggregation preference vector, the group rejection vector, and the candidate recommendation items displayed subsequently.

[0008] Furthermore, the cold start profile construction module includes: The tag mapping unit is used to map the user entry channel identifier to a first tag subset according to a preset first mapping table, and to map the current scene tag to a second tag subset according to a preset second mapping table; A vector combination unit is used to concatenate the first subset of tags with the second subset of tags to generate the initial intent tag vector; The first mapping table and the second mapping table are pre-constructed based on the statistical patterns of group behavior in the historical interaction data of the entire domain and are updated regularly.

[0009] Furthermore, the graph neural network inference module includes: The graph access unit is used to access the target user as a new node into the real-time relationship graph, and to search for related group nodes with the same channel identifier and / or the same scene label in the real-time relationship graph according to the target user's channel identifier and scene label. The preference aggregation unit is used to aggregate the node embedding vectors of the associated group nodes through the message passing mechanism of the graph neural network to generate the group aggregation preference vector; The aversion aggregation unit is used to extract negative feedback records of the associated group nodes in historical interactions, perform clustering based on the content feature vectors corresponding to the negative feedback records, extract the cluster centers as the group's common aversion center vector, and generate the group rejection vector.

[0010] Furthermore, the real-time relationship graph is pre-constructed based on historical interaction data across the entire domain and updated in real time. The nodes in the real-time relationship graph include user nodes, content nodes, and channel nodes. The association edges in the real-time relationship graph are used to represent the channel association relationship between users, the interaction relationship between users and content, and the affiliation relationship between content and channels.

[0011] Furthermore, the preference aggregation unit determines the attention coefficient of each associated group node to the target user based on the channel identifier matching degree and scene tag matching degree between the associated group node and the target user, and aggregates the node embedding vector of each associated group node according to the attention coefficient to generate the group aggregation preference vector.

[0012] Furthermore, the negative feedback records include records of skipping, quickly swiping, and actively marking as uninterested by the associated group nodes in each historical session.

[0013] Furthermore, the real-time feedback self-calibration module includes: An operation event capture unit is used to display multiple candidate recommendation items in parallel during the target user's first screen interaction, and capture operation event data triggered by the target user for the multiple candidate recommendation items in the current session. The operation event data includes the trigger timestamp and screen coordinate position of each operation event. The entropy calculation unit is used to calculate the spatial dispersion based on the screen coordinate position of the operation event data, and to calculate the click rhythm uniformity based on the trigger timestamp of the operation event data, and to determine the intent locking coefficient based on the spatial dispersion and the click rhythm uniformity. The strategy adjustment unit is used to adjust the candidate recommendations to be displayed in the future according to the intent locking coefficient, determine the similarity expansion direction according to the positive click events in the operation event data, determine the rejection filtering direction according to the negative skip events in the operation event data, and switch the candidate recommendations to be displayed in the future according to the similarity expansion direction and the rejection filtering direction.

[0014] Furthermore, the entropy calculation unit includes: The spatial dispersion calculation subunit is used to obtain the coordinate position of each operation event on the screen in the operation event data, calculate the average vector magnitude and vector direction entropy of all coordinate positions relative to the center point of the screen, and determine the spatial dispersion based on the average vector magnitude and the vector direction entropy. The rhythm uniformity calculation subunit is used to obtain the trigger timestamp of each operation event in the operation event data, calculate the time interval sequence between adjacent operation events, and determine the click rhythm uniformity based on the variance of the time interval sequence. A fusion calculation subunit is used to determine the intent locking coefficient based on the spatial dispersion and the click rhythm uniformity.

[0015] Furthermore, the recommendation output module includes: The representation generation unit is used to take the initial cold-start profile as an anchor point, the group aggregation preference vector as a positive sample target, and the group rejection vector as a negative sample target, and iteratively optimize the anchor point to bring it closer to the positive sample target and move it away from the negative sample target in the embedding space, thereby generating a calibrated user preference representation. The recommendation generation unit is used to determine the similarity between each candidate recommendation item and the calibrated user preference representation based on the calibrated user preference representation and the candidate recommendation items subsequently displayed, and select a preset number of candidate recommendation items as the final recommendation result according to the similarity from high to low.

[0016] Furthermore, the similarity is obtained by calculating the cosine similarity between the calibrated user preference representation and the content feature vector of each candidate recommendation.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This embodiment introduces the channel identifier and scene tag when the user enters the system as the basis for constructing the cold start profile, so that the initial recommendation results have scene adaptability; by using a graph neural network to simultaneously aggregate the positive preference features and negative aversion features of related group nodes, a push-pull two-way constraint group knowledge transfer is formed, which makes up for the lack of information due to the absence of individual data; by displaying multiple candidate recommendation items in parallel in the first screen interaction and capturing the spatiotemporal correlation of user operation events, the subsequent recommendation strategy can be adjusted and optimized in real time.

[0018] Furthermore, this embodiment maps user entry channel identifiers and scene tags into a subset of tags through a tag mapping unit. Based on the statistical patterns of group behavior, a mapping table is constructed and updated periodically, enabling instantaneous context to be efficiently transformed into a structured profile, solving the problem of not being able to provide personalized recommendations when new users have no historical data. By embedding pre-trained words, tags are mapped to a dense vector space and concatenated to generate an initial intent tag vector that integrates channel and scene features, providing an accurate initial representation basis for subsequent group preference aggregation and recommendation output.

[0019] Furthermore, in this embodiment, the graph access unit connects the target user as a new node to the real-time relationship graph and retrieves associated group nodes with the same channel identifier and / or scene label, enabling new users without historical behavioral data to achieve cold-start recommendation by leveraging group knowledge. The preference aggregation unit extracts the embedding vectors of associated group nodes through the graph attention network layer, and classifies the associated group nodes into different candidate sets according to the matching degree of channel identifier and scene label. After aggregation, they are concatenated to generate a group aggregation preference vector, so that the group nodes more closely associated with the target user contribute more to the preference vector, while retaining the differentiated information of groups with different degrees of association. The aversion aggregation unit extracts the negative feedback records of associated group nodes, clusters the negative feedback content feature vectors, and selects the cluster center of the largest category as the group rejection vector, so that the features that the group commonly dislikes can be quantitatively represented and used for subsequent recommendation constraints.

[0020] Furthermore, this embodiment captures user's first-screen operation events in real time through the operation event capture unit, providing an immediate data foundation for strategy adjustments after the first-screen interaction; the spatial dispersion calculation subunit quantifies the spatial concentration of operations through average offset distance and vector direction entropy, the rhythm uniformity calculation subunit quantifies the rhythmic stability of operations through the variance of the time interval sequence, and the fusion calculation subunit determines the intent locking coefficient by combining monotonically decreasing and monotonically increasing functions, realizing joint quantitative analysis of the clarity of user intent; the strategy adjustment unit adjusts the category search range in stages according to the intent locking coefficient, and determines the similarity expansion direction and the rejection filtering direction according to positive click events and negative skip events, respectively, and switches subsequent recommendation items according to the common constraints of the two directions, so that subsequent recommendations can both align with the direction that users are interested in and avoid content that they are not interested in, improving the ability to quickly correct deviations after the first-screen recommendation and the user experience.

[0021] Furthermore, this embodiment uses the initial cold-start profile as an anchor point through a representation generation unit. It iterative optimization is performed with the group aggregation preference vector as the positive sample target and the group rejection vector as the negative sample target. This ensures that the anchor point is simultaneously subject to bidirectional constraints of positive convergence and negative distancing in the embedding space, generating a calibrated user preference representation. This solves the problems of recommendation bias caused by relying solely on the cold-start profile and individual feature overload caused by relying solely on group preferences, achieving an organic integration of group knowledge transfer and initial individual profiles. By using the ratio of the distance between the anchor point and the positive sample target to the distance between the anchor point and the negative sample target as the optimization objective, and minimizing this ratio as the iteration direction, the optimization process simultaneously considers the convergence of positive preferences and the avoidance of negative aversion, preventing recommendation imbalance caused by a bias in one direction. The recommendation generation unit calculates the cosine similarity between the calibrated user preference representation and each candidate recommendation item and sorts them from high to low similarity to select the final recommendation result, ensuring that the ranking logic of the recommendation output is consistent with the representation optimization objective in the embedding space. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a full-domain data value-added service system that supports trusted transactions across multiple domains, as described in this embodiment. Figure 2 This is a schematic diagram of the cold start profile construction module in the full-domain data value-added service system that supports trusted transactions in multiple domains, as described in this embodiment. Figure 3 This is a schematic diagram of the graph neural network inference module in the full-domain data value-added service system supporting trusted transactions across multiple domains, as described in this embodiment. Figure 4 This is a schematic diagram of the real-time feedback self-calibration module in the full-domain data value-added service system that supports trusted transactions in multiple fields, as described in this embodiment. Detailed Implementation

[0023] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0024] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] Please see Figure 1 As shown, it is a schematic diagram of the full-domain data value-added service system that supports trusted transactions in multiple domains in this embodiment.

[0026] This embodiment provides a comprehensive data value-added service system that supports trusted transactions across multiple domains, including: The data acquisition module is used to acquire real-time context data when a user enters the system. The real-time context data includes the user's entry channel identifier and the current scene tag. The cold start profile building module is used to combine the real-time context data into an initial intent tag vector in response to the detection that the target user has no historical behavior data, as the initial cold start profile of the target user. The graph neural network inference module is used to access the target user as a new node into the real-time relationship graph, and to aggregate the preference features of related group nodes that have the same channel identifier and / or the same scene label as the target user through the graph neural network, and output the group aggregated preference vector; it is also used to aggregate the negative aversion features of the related group nodes to generate a group rejection vector. The real-time feedback self-calibration module is configured to display multiple candidate recommendation items in parallel during the target user's first screen interaction, capture the operation event data triggered by the target user for the multiple candidate recommendation items in the current session, determine the adjustment strategy of the candidate recommendation items based on the temporal and spatial correlation between each operation event in the operation event data, and switch the subsequently displayed candidate recommendation items according to the adjustment strategy. The recommendation output module is used to determine and output the final recommendation result to the target user based on the initial cold start profile, the group aggregation preference vector, the group rejection vector, and the candidate recommendation items displayed subsequently.

[0027] This embodiment introduces the channel identifier and scene tag when the user enters the system as the basis for constructing the cold start profile, so that the initial recommendation results have scene adaptability; by using a graph neural network to simultaneously aggregate the positive preference features and negative aversion features of related group nodes, a push-pull two-way constraint group knowledge transfer is formed, which makes up for the lack of information due to the absence of individual data; by displaying multiple candidate recommendation items in parallel during the first screen interaction and capturing the spatiotemporal correlation of user operation events, the subsequent recommendation strategy can be adjusted and optimized in real time.

[0028] Specifically, the data acquisition module is triggered when a target user enters the system, acquiring real-time context data of the user's entry through front-end embedding SDKs or server-side gateway interception. Specifically, when a target user accesses the system through any channel such as a webpage, mobile application, mini-program, or third-party partner platform, the data acquisition module automatically captures the entry channel identifier of the current session. This entry channel identifier includes, but is not limited to, the platform name from which the user originated, the entry page URL, promotional activity ID, or QR code parameters, used to characterize the channel from which the user entered the system.

[0029] Simultaneously, the data acquisition module determines the user's current scenario tag by using the current page's routing information, URL parameters, the identifier of the entry component clicked by the user, or the name of the functional module where the page is located. These scenario tags include, but are not limited to, homepage recommendation scenarios, product detail scenarios, shopping cart scenarios, content browsing scenarios, search result scenarios, or promotional marketing scenarios, representing the specific business scenario when the user enters the system. Furthermore, the data acquisition module also obtains contextual auxiliary information for the current session, including but not limited to entry timestamps, user device information, network environment information, and user login status information. This real-time contextual data is encapsulated according to a preset data format and sent to the cold start profile building module for further processing. After completing one data acquisition, the data acquisition module continuously monitors user actions within the current session, transmitting subsequent operation event data to the real-time feedback self-calibration module in real time until the current session ends or the user exits the system.

[0030] Please see Figure 2 As shown, it is a schematic diagram of the cold start profile construction module in the full-domain data value-added service system that supports trusted transactions in multiple domains in this embodiment.

[0031] Specifically, the cold start profile building module includes: The tag mapping unit is used to map the user entry channel identifier to a first tag subset according to a preset first mapping table, and to map the current scene tag to a second tag subset according to a preset second mapping table; A vector combination unit is used to concatenate the first subset of tags with the second subset of tags to generate the initial intent tag vector; The first mapping table and the second mapping table are pre-constructed based on the statistical patterns of group behavior in the historical interaction data of the entire domain and are updated regularly.

[0032] In this embodiment, the tag mapping unit receives real-time context data sent by the data acquisition module, extracts the user entry channel identifier, and maps the user entry channel identifier to a first tag subset according to a preset first mapping table. The first mapping table stores the mapping relationship between each channel identifier and its corresponding tag subset. The first mapping table is constructed based on the statistical patterns of group behavior of users from different channel sources in the historical interaction data of the entire domain. That is, by statistically analyzing the high-frequency interaction categories of users from each channel source in the historical data, the most frequently occurring category tags are extracted to form the first tag subset corresponding to the channel. Specifically, the content of all users' historical interactions under each channel source is grouped and statistically analyzed according to categories. The frequency ratio of each category under the channel source is calculated. After sorting according to the frequency ratio from high to low, the top K1 category tags are selected to form the first tag subset corresponding to the channel. K1 is a preset positive integer that can be dynamically set according to the richness of the category distribution of the channel source. The more dispersed the distribution, the larger the value of K1; the more concentrated the distribution, the smaller the value of K1. The mapping relationship is updated according to the latest historical data at a preset period to ensure the timeliness and accuracy of the mapping relationship.

[0033] Simultaneously, the tag mapping unit queries the corresponding second tag subset in a preset second mapping table based on the current scene tag. The second mapping table pre-stores the mapping relationship between each scene tag and its corresponding tag subset. This second mapping table is also constructed based on the statistical patterns of user behavior in different scenes within the overall historical interaction data. Specifically, it calculates the frequency of user historical interaction content in each scene according to the proportion of category occurrence, and selects the top K2 category tags to form the second tag subset corresponding to that scene. Here, K2 is a preset positive integer that can be dynamically set according to the richness of the scene's category distribution. The second mapping table is updated periodically according to a preset cycle. After obtaining the first and second tag subsets, the vector combination unit converts each tag in the two subsets into a corresponding tag embedding vector according to a preset tag encoding rule. It then concatenates all the tag embedding vectors to generate a fixed-dimensional initial intent tag vector. The tag encoding rule is as follows: each tag is mapped to the dense vector space output by the system's pre-trained word embedding model. Each category tag corresponds to a fixed-dimensional real-valued vector, and the semantic similarity between different tags is measured by the cosine distance between their vectors.

[0034] This embodiment maps user entry channel identifiers and scene tags into a subset of tags through a tag mapping unit. Based on the statistical patterns of group behavior, a mapping table is constructed and updated periodically, enabling instantaneous context to be efficiently transformed into a structured profile, solving the problem of not being able to provide personalized recommendations when new users have no historical data. By embedding pre-trained words, tags are mapped to a dense vector space and concatenated to generate an initial intent tag vector that integrates channel and scene features, providing an accurate initial representation basis for subsequent group preference aggregation and recommendation output.

[0035] Please see Figure 3 As shown, it is a schematic diagram of the graph neural network inference module in the full-domain data value-added service system that supports trusted transactions in multiple domains in this embodiment.

[0036] Specifically, the graph neural network inference module includes: The graph access unit is used to access the target user as a new node into the real-time relationship graph, and to search for related group nodes with the same channel identifier and / or the same scene label in the real-time relationship graph according to the target user's channel identifier and scene label. The preference aggregation unit is used to aggregate the node embedding vectors of the associated group nodes through the message passing mechanism of the graph neural network to generate the group aggregation preference vector; The aversion aggregation unit is used to extract negative feedback records of the associated group nodes in historical interactions, perform clustering based on the content feature vectors corresponding to the negative feedback records, extract the cluster centers as the group's common aversion center vector, and generate the group rejection vector.

[0037] Specifically, the real-time relationship graph is pre-built based on historical interaction data across the entire domain and updated in real time. The nodes in the real-time relationship graph include user nodes, content nodes, and channel nodes. The association edges in the real-time relationship graph are used to represent the channel association relationship between users, the interaction relationship between users and content, and the affiliation relationship between content and channel.

[0038] Specifically, the preference aggregation unit determines the attention coefficient of each associated group node to the target user based on the channel identifier matching degree and scene tag matching degree between the associated group node and the target user, and aggregates the node embedding vector of each associated group node according to the attention coefficient to generate the group aggregation preference vector.

[0039] Specifically, the negative feedback records include records of skipping, quickly swiping, and actively marking as uninterested by the associated group nodes in each historical session.

[0040] In this embodiment, after receiving the initial cold-start profile of the target user sent by the cold-start profile construction module, the graph access unit adds the target user as a new node to the real-time relationship graph. The real-time relationship graph is pre-constructed based on historical interaction data across the entire domain and stored in a graph database, updating in real-time as new users and new interactions occur. The nodes in this real-time relationship graph include user nodes, content nodes, and channel nodes. Association edges represent channel associations between users, interaction relationships between users and content, and affiliation relationships between content and channels. Based on the target user's channel identifier, the graph access unit retrieves user nodes with the same channel identifier and the content nodes connected to these user nodes through interaction relationship edges in the real-time relationship graph. Simultaneously, based on the target user's scene tag, it retrieves other user nodes currently in the same scene and the content nodes they have interacted with in the real-time relationship graph, treating all retrieved user nodes and content nodes as associated group nodes.

[0041] The preference aggregation unit uses a graph attention network layer to aggregate the embedding vectors of associated group nodes. Specifically, the preference aggregation unit extracts the embedding vectors of each associated group node from the real-time relationship graph. The dimension of each node's embedding vector is d, and a total of M associated group nodes are retrieved, forming an M×d node embedding matrix. The preference aggregation unit traverses each associated group node one by one, determining whether its channel identifier is the same as the target user's channel identifier and whether its scene label is the same as the target user's scene label. If both the channel identifier and scene label are the same, the embedding vector of the associated group node is copied to the first candidate set; if only the channel identifier or only the scene label is the same, the embedding vector of the associated group node is copied to the second candidate set; if neither the channel identifier nor the scene label is the same, the embedding vector of the associated group node is copied to the third candidate set. The preference aggregation unit sums all the embedding vectors in the first candidate set element-wise and then averages them to obtain the first aggregated vector; it sums all the embedding vectors in the second candidate set element-wise and then averages them to obtain the second aggregated vector; and it sums all the embedding vectors in the third candidate set element-wise and then averages them to obtain the third aggregated vector. The preference aggregation unit concatenates the first aggregation vector, the second aggregation vector, and the third aggregation vector in a preset order to generate a group aggregated preference vector with a dimension of 3d.

[0042] The aversion aggregation unit collects content with negative feedback markers from the content nodes connected to the aforementioned associated group nodes in the real-time relationship graph. These negative feedback markers include skip operation records, quick swipe operation records, and records actively marked as uninteresting. Skip operation records indicate that the user swipes past the recommended item without any interaction; quick swipe operation records indicate that the user's dwell time on the recommended item is less than a preset time threshold; and records actively marked as uninteresting indicate that the user explicitly clicked the "uninterested" or "don't want to see" button. The aversion aggregation unit extracts the content feature vectors corresponding to all content with negative feedback markers, resulting in T negative feedback content feature vectors, each with dimension d. The aversion aggregation unit uses a clustering algorithm to cluster these T negative feedback content feature vectors. The number of clusters, K3, is dynamically determined based on the total number of negative feedback samples T, i.e., K3 is the integer part of the square root of T. After clustering is completed, the number of negative feedback content feature vectors contained in each category is counted. The category with the largest number of samples is selected, and the mean vector of all negative feedback content feature vectors in the category is calculated. This mean vector is used as the cluster center of the category, which is the common aversion center vector of the group. The group rejection vector is generated, and the dimension of the group rejection vector is d.

[0043] In this embodiment, the graph neural network inference module needs to undergo a training phase before use. The specific training process is as follows: Constructing a training dataset: Extract a large number of existing users and their interaction records from the full-domain historical interaction data as training samples. For each training sample, take any existing user as the target user, and use its channel identifier and scene label as input. Extract the user's associated group nodes from the real-time relationship graph. Generate a group aggregation preference vector by the preference aggregation unit and a group rejection vector by the aversion aggregation unit in the same way as described above. Constructing a graph attention network model: The graph attention network model includes an input layer, a graph attention layer, and an output layer. The input layer receives the initial embedding vectors of the associated group nodes. The graph attention layer introduces an attention mechanism to dynamically calculate the attention coefficient of each associated group node's embedding vector during aggregation, and aggregates the embedding vectors of all associated group nodes according to the attention coefficient, outputting a group aggregation preference vector. The parameters of the graph attention layer include the embedding vector transformation matrix and attention weight vector of each node.

[0044] Construct the training loss function: For each training sample, extract the content feature vectors of the positive content that the target user actually interacted with from their historical interaction records as positive samples, and extract the content feature vectors of content that they explicitly skipped or marked as uninteresting as negative samples. Use the Euclidean distance between the aggregated preference vector and the content feature vectors of the positive samples as the first loss term, aiming to minimize this distance; use the Euclidean distance between the aggregated preference vector and the content feature vectors of the negative samples as the second loss term, aiming to maximize this distance; use the difference between the first and second loss terms as the overall loss function.

[0045] Training Optimization: A stochastic gradient descent optimization algorithm is employed. Training samples from the training dataset are input into the graph attention network model in batches. Forward propagation calculates the overall loss function value, and backpropagation calculates the gradient of the model parameters. The transformation matrix and attention weight vector of the graph attention layer are updated based on the gradient. This iterative process is repeated until the overall loss function value converges to below a preset convergence threshold or reaches a preset maximum number of iterations, resulting in a trained graph attention network model. The trained graph attention network model is deployed in the preference aggregation unit for group preference aggregation of real-time access target users during the actual inference stage. The graph neural network inference module sends the group aggregation preference vector and group rejection vector generated after training to the recommendation output module. The group aggregation preference vector acts as a positive approach to the target, used to align the initial cold-start profile towards this direction in subsequent recommendation decisions; the group rejection vector acts as a negative approach to the target, used to push the initial cold-start profile away from this direction. This bidirectional constraint achieves more accurate group knowledge transfer.

[0046] In this embodiment, the graph access unit connects the target user as a new node to the real-time relationship graph and retrieves related group nodes with the same channel identifier and / or scene tag, enabling new users without historical behavioral data to achieve cold-start recommendation by leveraging group knowledge. The preference aggregation unit extracts the embedding vectors of related group nodes through the graph attention network layer, and classifies the related group nodes into different candidate sets according to the matching degree of channel identifier and scene tag. After aggregation, the sets are concatenated to generate a group aggregation preference vector, so that the group nodes more closely associated with the target user contribute more to the preference vector, while retaining the differentiated information of groups with different degrees of association. The aversion aggregation unit extracts the negative feedback records of related group nodes, clusters the negative feedback content feature vectors, and selects the cluster center of the largest category as the group rejection vector, so that the features that the group commonly dislikes can be quantitatively represented and used for subsequent recommendation constraints.

[0047] Please see Figure 4 As shown, this is a schematic diagram of the real-time feedback self-calibration module in the full-domain data value-added service system that supports trusted transactions in multiple domains in this embodiment.

[0048] Specifically, the real-time feedback self-calibration module includes: An operation event capture unit is used to display multiple candidate recommendation items in parallel during the target user's first screen interaction, and capture operation event data triggered by the target user for the multiple candidate recommendation items in the current session. The operation event data includes the trigger timestamp and screen coordinate position of each operation event. The entropy calculation unit is used to calculate the spatial dispersion based on the screen coordinate position of the operation event data, and to calculate the click rhythm uniformity based on the trigger timestamp of the operation event data, and to determine the intent locking coefficient based on the spatial dispersion and the click rhythm uniformity. The strategy adjustment unit is used to adjust the candidate recommendations to be displayed in the future according to the intent locking coefficient, determine the similarity expansion direction according to the positive click events in the operation event data, determine the rejection filtering direction according to the negative skip events in the operation event data, and switch the candidate recommendations to be displayed in the future according to the similarity expansion direction and the rejection filtering direction.

[0049] Specifically, the entropy calculation unit includes: The spatial dispersion calculation subunit is used to obtain the coordinate position of each operation event on the screen in the operation event data, calculate the average vector magnitude and vector direction entropy of all coordinate positions relative to the center point of the screen, and determine the spatial dispersion based on the average vector magnitude and the vector direction entropy. The rhythm uniformity calculation subunit is used to obtain the trigger timestamp of each operation event in the operation event data, calculate the time interval sequence between adjacent operation events, and determine the click rhythm uniformity based on the variance of the time interval sequence. A fusion calculation subunit is used to determine the intent locking coefficient based on the spatial dispersion and the click rhythm uniformity.

[0050] In this embodiment, the operation event capture unit displays multiple candidate recommendation items in parallel during the target user's initial screen interaction. These multiple candidate recommendation items are simultaneously presented on the target user's display interface according to a preset layout. The operation event capture unit uses front-end tracking technology or a system event monitoring mechanism to capture operation event data triggered by the target user in response to the multiple candidate recommendation items within the current session in real time. The operation event data includes the trigger timestamp and screen coordinates of each operation event. Operation events include click events, skip events, swipe events, and dwell events. A click event indicates that the user clicks on a candidate recommendation item; a skip event indicates that the user swipes past or ignores a candidate recommendation item; a swipe event indicates that the user performs a vertical swipe operation on the screen; and a dwell event indicates that the user dwells on a candidate recommendation item for more than a preset time threshold.

[0051] The preset time threshold is used to distinguish between effective attention and rapid swiping. It is determined by combining the minimum cognitive reaction time of the human eye to visual information with the average browsing speed in mobile interaction scenarios, ensuring that operation events with a dwell time exceeding this threshold reflect effective user attention behavior. The operation event capture unit records the trigger timestamps and screen coordinates of each operation event according to their trigger order, forming a sequence of operation events within the current session, and transmits this sequence to the entropy calculation unit in real time.

[0052] The spatial discreteness calculation subunit extracts the screen coordinates of all operation events from the operation event sequence. It calculates the offset distance of each operation event's screen coordinates relative to the screen center point, and averages these offset distances to obtain the average offset distance. Simultaneously, based on the offset direction of each operation event relative to the screen center point, the subunit divides the entire circumference into multiple directional intervals at preset angular intervals. It counts the number of operation events falling within each directional interval, and uses the ratio of the number of operation events in each interval to the total number of operation events as the probability value for that interval. Based on the probability values ​​of each directional interval, it calculates the information entropy to obtain the vector direction entropy.

[0053] The preset angle interval is used to control the fineness of the directional interval division. It is determined by comprehensively considering the spatial resolution of screen touch events and the statistical requirements of the operation event sample size. An excessively large angle interval will mask the distribution differences in different directions, while an excessively small angle interval will result in sparse samples and statistical instability within each directional interval. Under the conditions of typical mobile screen size and expected number of operation events, the smallest angle interval that ensures sufficient statistical sample size for most directional intervals is selected. The spatial dispersion calculation subunit determines the spatial dispersion based on the average offset distance and vector direction entropy: a smaller average offset distance indicates that the operation position is more concentrated in the central area of ​​the screen, and a smaller vector direction entropy indicates that the operation direction is more concentrated. A lower average offset distance and vector direction entropy indicate lower spatial dispersion, signifying that the user's initial screen operation is spatially concentrated in a specific area; conversely, a higher average offset distance and vector direction entropy indicate higher spatial dispersion, signifying that the user's initial screen operation is spatially scattered throughout the screen.

[0054] The rhythm uniformity calculation subunit extracts the trigger timestamps of each operation event from the operation event sequence. It then calculates the time intervals between adjacent operation events in chronological order of the trigger timestamps, resulting in a time interval sequence. The rhythm uniformity calculation subunit calculates the average of all time intervals in this time interval sequence, then squares the difference between each time interval and the average time interval, and averages all squared values ​​to obtain the variance of the time interval sequence. The rhythm uniformity calculation subunit determines the click rhythm uniformity based on the variance of the time interval sequence: a smaller variance indicates a more uniform distribution of operation events over time, resulting in higher click rhythm uniformity and a more stable user operation rhythm; a larger variance indicates a more uneven distribution of operation events over time, resulting in lower click rhythm uniformity and greater fluctuations in user operation rhythm.

[0055] The fusion computing subunit receives the spatial dispersion from the spatial dispersion calculation subunit and the click rhythm uniformity from the rhythm uniformity calculation subunit. The fusion computing subunit determines the intent locking coefficient as follows: It constructs a first monotonically decreasing function with spatial dispersion as the independent variable and the intent locking coefficient as the dependent variable, where the intent locking coefficient decreases as spatial dispersion increases. Simultaneously, it constructs a second monotonically increasing function with click rhythm uniformity as the independent variable and the intent locking coefficient as the dependent variable, where the intent locking coefficient increases as click rhythm uniformity increases. The fusion computing subunit inputs the spatial dispersion into the first monotonically decreasing function to calculate the first component, and inputs the click rhythm uniformity into the second monotonically increasing function to calculate the second component. Then, it calculates the average of the first and second components to obtain the intent locking coefficient.

[0056] In the first monotonically decreasing function, as the spatial dispersion increases, the first component decreases accordingly, and the rate of decrease gradually slows down. In the second monotonically increasing function, as the click rhythm uniformity increases, the second component increases accordingly, and the rate of increase gradually slows down. The mapping relationship between the two functions is determined in advance through system calibration, and their specific parameters are obtained by fitting the statistical relationship between spatial dispersion and user satisfaction, and the statistical relationship between click rhythm uniformity and user satisfaction in historical data.

[0057] In this embodiment, the strategy adjustment unit receives an intent locking coefficient sent by the fusion calculation subunit. The value range of this intent locking coefficient is a preset interval. A higher value indicates that the user's current intent is more explicit and stable, while a lower value indicates that the user's current intent is more scattered or fluctuating. The strategy adjustment unit determines the category search range of subsequent candidate recommendations based on the intent locking coefficient: when the intent locking coefficient is higher than a preset first judgment threshold, it is determined that the user's intent is explicit, and the category search range of subsequent candidate recommendations is narrowed, so that the subsequent candidate recommendations are concentrated in categories with higher similarity to the content clicked on the user's first screen; when the intent locking coefficient is lower than a preset second judgment threshold, it is determined that the user's intent is scattered, and the category search range of subsequent candidate recommendations is expanded, so that the subsequent candidate recommendations cover more diverse categories; when the intent locking coefficient is between the first judgment threshold and the second judgment threshold, it is determined that the explicitness of the user's intent is moderate, and the scaling of the category search range is correspondingly between the two extreme cases mentioned above.

[0058] The specific values ​​of the first and second judgment thresholds are predetermined as follows: During the testing phase before the system goes live, a large amount of operational event data from test users in the initial screen interaction is collected, the intent lock coefficient of each test user is calculated, and the final recommendation satisfaction feedback in subsequent sessions is recorded; using the intent lock coefficient as a feature and the subsequent recommendation satisfaction feedback as a label, a receiver operating characteristic (ROC) curve is plotted, and the intent lock coefficient value corresponding to the maximum value of the Youden index is selected as the boundary reference between the first and second judgment thresholds. Specifically, the intent lock coefficient value corresponding to the maximum sum of recall and precision on the ROC curve is selected as the first judgment threshold, and the intent lock coefficient value corresponding to the second maximum sum of recall and precision on the same curve is selected as the second judgment threshold. The specific scaling range of the category search scope is predetermined by the system configuration parameters. These configuration parameters define the mapping relationship between the various value ranges of the intent lock coefficient and the scaling factor of the category search scope. The value is determined during system initialization based on the business scenario's preference for exploration breadth and utilization precision, and is dynamically adjusted during system operation based on actual recommendation effect feedback.

[0059] The strategy adjustment unit determines the similarity expansion direction based on positive click events in the operation event data. Specifically, it extracts the content feature vector of candidate recommendations where the user generates a positive click event, and uses the direction pointed to by this content feature vector as the similarity expansion direction. The strategy adjustment unit also determines the rejection filtering direction based on negative skip events in the operation event data. Specifically, it extracts the content feature vector of candidate recommendations where the user generates a negative skip event, and uses the opposite direction pointed to by this content feature vector as the rejection filtering direction. The strategy adjustment unit switches the subsequently displayed candidate recommendations according to both the similarity expansion direction and the rejection filtering direction. That is, during the selection process of subsequently displayed recommendations, it prioritizes candidate recommendations whose content feature direction is consistent with the similarity expansion direction and opposite to the rejection filtering direction.

[0060] This embodiment captures user's first-screen operation events in real time through an operation event capture unit, providing an immediate data foundation for strategy adjustments after the first-screen interaction. The spatial dispersion calculation subunit quantifies the spatial concentration of operations through average offset distance and vector direction entropy, while the rhythm uniformity calculation subunit quantifies the rhythmic stability of operations through the variance of the time interval sequence. The fusion calculation subunit determines the intent locking coefficient by combining monotonically decreasing and monotonically increasing functions, achieving joint quantitative analysis of the clarity of user intent. The strategy adjustment unit adjusts the category search range in stages according to the intent locking coefficient, and determines the similarity expansion direction and the rejection filtering direction based on positive click events and negative skip events, respectively. Subsequent recommendations are switched according to the common constraints of the two directions, so that subsequent recommendations can both align with the user's interests and avoid uninteresting content, improving the ability to quickly correct deviations after the first-screen recommendation and enhancing the user experience.

[0061] Specifically, the recommendation output module includes: The representation generation unit is used to take the initial cold-start profile as an anchor point, the group aggregation preference vector as a positive sample target, and the group rejection vector as a negative sample target, and iteratively optimize the anchor point to bring it closer to the positive sample target and move it away from the negative sample target in the embedding space, thereby generating a calibrated user preference representation. The recommendation generation unit is used to determine the similarity between each candidate recommendation item and the calibrated user preference representation based on the calibrated user preference representation and the candidate recommendation items subsequently displayed, and select a preset number of candidate recommendation items as the final recommendation result according to the similarity from high to low.

[0062] Specifically, the similarity is obtained by calculating the cosine similarity between the calibrated user preference representation and the content feature vector of each candidate recommendation.

[0063] In this embodiment, the representation generation unit receives the initial cold-start profile from the cold-start profile construction module, and the group aggregation preference vector and group rejection vector from the graph neural network inference module. The representation generation unit uses the initial cold-start profile as an anchor point, which represents the initial position of the target user in the embedding space. The representation generation unit uses the group aggregation preference vector as a positive sample target, which represents the collective preference direction of a group of users with the same channel source and / or scene context as the target user; this is the target direction that the anchor point should gravitate towards. The representation generation unit uses the group rejection vector as a negative sample target, which represents the content direction that the aforementioned group of users are collectively uninterested in; this is the target direction that the anchor point should avoid.

[0064] The representation generation unit adjusts the anchor point's position in the embedding space through iterative optimization. Specifically, starting from the current embedding position of the initial cold-start profile, the representation generation unit calculates the first distance between the anchor point and the positive sample target, and the second distance between the anchor point and the negative sample target in each iteration. The representation generation unit constructs an optimization objective based on the ratio of the first distance to the second distance. The smaller the first distance and the larger the second distance, the smaller this ratio, indicating that the anchor point is closer to the positive sample target and farther away from the negative sample target. The representation generation unit aims to minimize this ratio by gradually updating the anchor point's embedding position using backpropagation and gradient descent algorithms, causing the anchor point to move towards the positive sample target and in the opposite direction to the negative sample target after each iteration. The representation generation unit repeats this iterative process until a preset maximum number of iterations is reached or the change in the optimization objective is lower than a preset convergence threshold. The iteration stops when the anchor point embedding position after the iteration is completed is used as the calibrated user preference representation.

[0065] The recommendation generation unit receives the calibrated user preference representation sent by the representation generation unit, as well as the candidate recommendations for subsequent display sent by the real-time feedback self-calibration module. The recommendation generation unit extracts the content feature vectors of each candidate recommendation, with each content feature vector residing in the same embedding space as the calibrated user preference representation. The recommendation generation unit calculates the cosine similarity between the calibrated user preference representation and the content feature vector of each candidate recommendation, which is obtained by dividing the dot product of the two vectors by the product of their magnitudes. A higher similarity value indicates that the candidate recommendation and the calibrated user preference representation are more aligned in the embedding space, meaning the candidate recommendation better matches the target user's interests. The recommendation generation unit sorts all subsequent candidate recommendations from highest to lowest similarity value and selects a predetermined number of the top-ranked candidate recommendations as the final recommendation result, outputting it to the target user's display interface.

[0066] The preset maximum number of iterations and convergence threshold are determined as follows: During the system training phase, the recommendation accuracy on the validation set is used as the evaluation metric. The model performance is tested under different combinations of maximum iterations and convergence thresholds. The maximum number of iterations and convergence threshold corresponding to the peak recommendation accuracy are selected as the optimal configuration, ensuring that the anchor point optimization process achieves optimal results within a reasonable computational cost. The preset number is determined based on the number of recommendation items that the target user's display interface can accommodate on a single screen, ensuring that the final recommendation results are fully presented within the user's current screen area without additional scrolling or page-turning operations. This number can be adaptively adjusted according to the screen size and resolution of different terminal devices.

[0067] This embodiment uses the initial cold-start profile as an anchor point through a representation generation unit. It iterative optimization is performed with the group aggregation preference vector as the positive sample target and the group rejection vector as the negative sample target. This ensures that the anchor point is simultaneously constrained by both positive convergence and negative distancing in the embedding space, generating a calibrated user preference representation. This solves the problems of recommendation bias caused by relying solely on the cold-start profile and individual feature overload caused by relying solely on group preferences, achieving an organic integration of group knowledge transfer and initial individual profiles. By using the ratio of the distance between the anchor point and the positive sample target to the distance between the anchor point and the negative sample target as the optimization objective, and minimizing this ratio as the iteration direction, the optimization process simultaneously considers the convergence of positive preferences and the avoidance of negative aversion, preventing recommendation imbalance caused by a bias in one direction. The recommendation generation unit calculates the cosine similarity between the calibrated user preference representation and each candidate recommendation item, and sorts the results from high to low similarity to select the final recommendation result. This ensures that the ranking logic of the recommendation output is consistent with the representation optimization objective in the embedding space.

[0068] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A comprehensive data value-added service system supporting trusted transactions across multiple domains, characterized in that, include: The data acquisition module is used to acquire real-time context data when a user enters the system. The real-time context data includes the user's entry channel identifier and the current scene tag. The cold start profile building module is used to combine the real-time context data into an initial intent tag vector in response to the detection that the target user has no historical behavior data, as the initial cold start profile of the target user. The graph neural network inference module is used to access the target user as a new node into the real-time relationship graph, and to aggregate the preference features of related group nodes that have the same channel identifier and / or the same scene label as the target user through the graph neural network, and output the group aggregated preference vector; it is also used to aggregate the negative aversion features of the related group nodes to generate a group rejection vector. The real-time feedback self-calibration module is configured to display multiple candidate recommendation items in parallel during the target user's first screen interaction, capture the operation event data triggered by the target user for the multiple candidate recommendation items in the current session, determine the adjustment strategy of the candidate recommendation items based on the temporal and spatial correlation between each operation event in the operation event data, and switch the subsequently displayed candidate recommendation items according to the adjustment strategy. The recommendation output module is used to determine and output the final recommendation result to the target user based on the initial cold start profile, the group aggregation preference vector, the group rejection vector, and the candidate recommendation items displayed subsequently.

2. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 1, characterized in that, The cold start profile construction module includes: The tag mapping unit is used to map the user entry channel identifier to a first tag subset according to a preset first mapping table, and to map the current scene tag to a second tag subset according to a preset second mapping table; A vector combination unit is used to concatenate the first subset of tags with the second subset of tags to generate the initial intent tag vector; The first mapping table and the second mapping table are pre-constructed based on the statistical patterns of group behavior in the historical interaction data of the entire domain and are updated regularly.

3. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 1, characterized in that, The graph neural network inference module includes: The graph access unit is used to access the target user as a new node into the real-time relationship graph, and to search for related group nodes with the same channel identifier and / or the same scene label in the real-time relationship graph according to the target user's channel identifier and scene label. The preference aggregation unit is used to aggregate the node embedding vectors of the associated group nodes through the message passing mechanism of the graph neural network to generate the group aggregation preference vector; The aversion aggregation unit is used to extract negative feedback records of the associated group nodes in historical interactions, perform clustering based on the content feature vectors corresponding to the negative feedback records, extract the cluster centers as the group's common aversion center vector, and generate the group rejection vector.

4. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 3, characterized in that, The real-time relationship graph is pre-built based on the full-domain historical interaction data and updated in real time. The nodes in the real-time relationship graph include user nodes, content nodes and channel nodes. The association edges in the real-time relationship graph are used to represent the channel association relationship between users, the interaction relationship between users and content, and the affiliation relationship between content and channel.

5. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 3, characterized in that, The preference aggregation unit determines the attention coefficient of each associated group node to the target user based on the channel identifier matching degree and scene tag matching degree between the associated group node and the target user, and aggregates the node embedding vector of each associated group node according to the attention coefficient to generate the group aggregated preference vector.

6. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 3, characterized in that, The negative feedback records include records of skipping, quickly swiping, and actively marking items as uninteresting by the associated group nodes in each historical session.

7. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 1, characterized in that, The real-time feedback self-calibration module includes: An operation event capture unit is used to display multiple candidate recommendation items in parallel during the target user's first screen interaction, and capture operation event data triggered by the target user for the multiple candidate recommendation items in the current session. The operation event data includes the trigger timestamp and screen coordinate position of each operation event. The entropy calculation unit is used to calculate the spatial dispersion based on the screen coordinate position of the operation event data, and to calculate the click rhythm uniformity based on the trigger timestamp of the operation event data, and to determine the intent locking coefficient based on the spatial dispersion and the click rhythm uniformity. The strategy adjustment unit is used to adjust the candidate recommendations to be displayed in the future according to the intent locking coefficient, determine the similarity expansion direction according to the positive click events in the operation event data, determine the rejection filtering direction according to the negative skip events in the operation event data, and switch the candidate recommendations to be displayed in the future according to the similarity expansion direction and the rejection filtering direction.

8. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 7, characterized in that, The entropy calculation unit includes: The spatial dispersion calculation subunit is used to obtain the coordinate position of each operation event on the screen in the operation event data, calculate the average vector magnitude and vector direction entropy of all coordinate positions relative to the center point of the screen, and determine the spatial dispersion based on the average vector magnitude and the vector direction entropy. The rhythm uniformity calculation subunit is used to obtain the trigger timestamp of each operation event in the operation event data, calculate the time interval sequence between adjacent operation events, and determine the click rhythm uniformity based on the variance of the time interval sequence. A fusion calculation subunit is used to determine the intent locking coefficient based on the spatial dispersion and the click rhythm uniformity.

9. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 1, characterized in that, The recommendation output module includes: The representation generation unit is used to take the initial cold-start profile as an anchor point, the group aggregation preference vector as a positive sample target, and the group rejection vector as a negative sample target, and iteratively optimize the anchor point to bring it closer to the positive sample target and move it away from the negative sample target in the embedding space, thereby generating a calibrated user preference representation. The recommendation generation unit is used to determine the similarity between each candidate recommendation item and the calibrated user preference representation based on the calibrated user preference representation and the candidate recommendation items subsequently displayed, and select a preset number of candidate recommendation items as the final recommendation result according to the similarity from high to low.

10. The full-domain data value-added service system supporting trusted transactions across multiple domains as described in claim 9, characterized in that, The similarity is obtained by calculating the cosine similarity between the calibrated user preference representation and the content feature vector of each candidate recommendation.

Images