Data fusion

Abstract

Disclosed in the embodiments of the present disclosure are a data fusion method and apparatus, and a storage medium and a terminal. The method comprises: inputting a relationship network graph of a target user into a fusion prediction model, wherein the relationship network graph takes user entities as nodes and relationships between the user entities as edges, and the user entities are generated on the basis of basic attributes of the users and external information; on the basis of the fusion prediction model, aggregating neighbor external information of the target user, so as to obtain target external information of the target user, and obtaining, on the basis of the fusion prediction model, an external evaluation corresponding to the target external information; acquiring an internal evaluation for the target user, and integrating the internal evaluation into the fusion prediction model; and fusing the external evaluation and the internal evaluation to obtain an overall evaluation that is output by means of the fusion prediction model. By means of objective relevance between users, external information is adaptively learned and propagated, thereby reducing the model maintenance costs while improving the utilization efficiency of the external information, significantly improving the utilization rate of the external information, and thus improving the reliability and accuracy of a final evaluation result.

Description

Data fusion TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of computer, and particularly, to data fusion. BACKGROUND

[0002] In the digital era, data as an indispensable core asset, various important data are often scattered in different enterprises and institutions, and these data drive the internal operation optimization and efficiency improvement of enterprises. In order to make better decisions, in addition to making full use of internal data, a single enterprise platform also needs to collect a large amount of high-quality external data from outside the enterprise. At this time, a data fusion method that can effectively fuse internal data and external data is needed, so that the enterprise can deeply mine the needs and preferences and behavior habits of consumers based on rich and diverse data, and thus provide more personalized and intelligent product and service experience for consumers. SUMMARY

[0003] Embodiments of the present disclosure provide a data fusion method, device, storage medium and terminal, which can solve the technical problem that internal information and external information are difficult to effectively fuse in related technologies.

[0004] In a first aspect, embodiments of the present disclosure provide a data fusion method, which comprises: inputting a relationship network graph of a target user into a fusion prediction model, the relationship network graph taking a user entity as a node and a relationship between each user entity as an edge connecting the nodes, and the user entity being generated based on basic attributes of the user and external information; aggregating neighbor external information of the target user based on the fusion prediction model to obtain target external information of the target user, and obtaining an external evaluation corresponding to the target external information based on the fusion prediction model; obtaining an internal evaluation of the target user and inputting the internal evaluation into the fusion prediction model, the internal evaluation being obtained by predicting based on internal information of the target user; and fusing the external evaluation and the internal evaluation based on the fusion prediction model to obtain an overall evaluation output by the fusion prediction model.

[0005] The beneficial effects brought by the technical solutions of the first aspect include: based on the node relationship in the relationship network graph, the external information of other neighbor users can be transmitted to the target user through the relationship, thereby the external information of the neighbor users is aggregated to obtain the external information of the target user, so that the external information is adaptively propagated and information gain is generated; then in the fusion prediction model, the external evaluation and the internal evaluation corresponding to the external information are further fused to output the final evaluation of the target user. In this way, the external information is adaptively learned and propagated based on the objective association between users, the utilization efficiency of the external information is improved, the model maintenance cost is reduced, the utilization rate of the external information is significantly improved, and the reliability and accuracy of the final evaluation result are improved.

[0006] In some possible implementation manners, before the relationship network graph of the target user is input into the fusion prediction model, the method further includes: obtaining user data of all users, wherein the user data includes basic attributes and external information; generating a user entity corresponding to each user and extracting a relationship between each user entity based on the user data of each user, and constructing a full-quantity relationship network graph corresponding to all users based on the user entity of each user and the relationship between the user entities; obtaining the relationship network graph of the target user, wherein the relationship network graph includes a target user entity of the target user and at least one neighbor user entity having a relationship with the target user entity.

[0007] Through the technical solutions in the possible implementation manners, the user entity of each user is constructed based on the basic attributes and the external information of all users, and the relationship between each user entity is extracted, so as to construct the full-quantity relationship network graph corresponding to all users by taking the user entity as a node and the relationship as an edge. On this basis, when the target user needs to be predicted, the relationship network graph corresponding to the target user can be directly searched in the full-quantity relationship network graph.

[0008] In some possible implementation manners, the external evaluation corresponding to the target external information is obtained based on the fusion prediction model, including: updating the target user entity of the target user based on the target external information, and predicting the target user entity based on the fusion prediction model to obtain the external evaluation corresponding to the target external information.

[0009] Through the technical solutions in the possible implementation manners, when the external information aggregation is specifically implemented, the target external information aggregated from other nodes can update the entity representation of the target user in the relationship network, and further, the model directly outputs the external evaluation of the target user based on the updated target user entity.

[0010] In some possible implementation manners, the target external information of the target user is obtained by aggregating external information of each neighbor user of the target user based on the fusion prediction model, and the method comprises: determining an aggregation weight corresponding to each neighbor user according to a relationship between the target user entity and each neighbor user entity in the relationship network graph, wherein the aggregation weight is positively correlated with a closeness of the relationship; and performing weighted average calculation on external information of each neighbor user in the fusion prediction model based on the aggregation weight of each neighbor user to obtain the target external information of the target user.

[0011] According to the technical solutions in the possible implementation manners, the closer the relationship between the target user and each neighbor user, the higher the aggregation weight of the external information of the neighbor user, so that the external information of each neighbor user is calculated by weighting to obtain the target external information of the target user, and the closeness of the relationship can be more appropriately reflected in the aggregation result.

[0012] In some possible implementation manners, the target external information of the target user is obtained by performing weighted average calculation on external information of each neighbor user in the fusion prediction model based on the aggregation weight of each neighbor user, and the method comprises: in the fusion prediction model, adding an attention weight to each external information based on a time of obtaining the external information, wherein the attention weight of the external information is greater when the time of obtaining the external information is closer to the current time; performing first weighted average calculation on the external information of each neighbor user based on the aggregation weight of each neighbor user to obtain first external information of the target user, and performing second weighted average calculation on the external information based on the attention weight of each external information to obtain second external information of the target user; and calculating the target external information of the target user according to the first external information and the second external information.

[0013] According to the technical solutions in the possible implementation manners, the time of obtaining the external information of each user can be different, and the external information obtained at a time close to the current time is more reliable than the external information obtained at a time far away from the current time for the current prediction, so that a monotonic and learnable attention weight is added to the external information, so that the closer the time of obtaining the external information, the weaker the attention to the information, thereby improving the stability of the model in continuous evaluation of the user and avoiding negative effects of long-term historical information on performance.

[0014] In some possible implementation manners, the internal evaluation is connected to the module in which the external evaluation is located in the fusion prediction model in a residual connection manner.

[0015] By the technical solutions in the possible implementation manners above, the internal evaluation obtained only based on internal information is introduced into the model when external evaluation is obtained through the residual connection mode, so as to avoid the disappearance of gradient information caused by the increase of model depth after the internal evaluation is input into the front end of the model, thereby improving the volume of original internal information when internal and external information is fused, and realizing more accurate and effective internal and external information fusion.

[0016] In some possible implementation manners, the fusion of the external evaluation and the internal evaluation based on the fusion prediction model to obtain the overall evaluation output by the fusion prediction model includes: in the fusion prediction model, the credibility of the external evaluation is determined according to the source of the external evaluation, the more real the source is, the higher the credibility is; the attention weight of the external evaluation and the internal evaluation is respectively assigned according to the credibility, the higher the credibility is, the higher the attention weight of the external evaluation is; and the external evaluation and the internal evaluation are weighted fused based on the attention weight of the external evaluation and the attention weight of the internal evaluation to obtain the overall evaluation output by the fusion prediction model.

[0017] By the technical solutions in the possible implementation manners above, the attention mechanism is introduced to fuse internal and external information, for a user with real external information obtained from the outside, the weight of the external evaluation of the model is higher when the internal and external evaluations are fused; if the user has no external information, the weight of the internal evaluation is higher when the internal and external evaluations are fused; and for a user with external information aggregated by neighbors, the weights of the internal evaluation and the external evaluation are equal when the internal and external evaluations are fused, so that the weight of the external evaluation in the fusion is adjusted according to the credibility of the external evaluation of the user, so that the customer group with large external information gain can retain the gain as much as possible, and the customer group with small external information gain can maintain internal consistency.

[0018] In some possible implementation manners, the fusion prediction model is trained based on at least one sample data carrying a standard evaluation label; in the training process of the fusion prediction model, a predicted evaluation label is output for each sample data, a loss value of a loss function is calculated according to the predicted evaluation label and the standard evaluation label of each sample data, and the model parameters are adjusted based on the loss value until the fusion prediction model meets a preset convergence condition.

[0019] By the technical solutions in the possible implementation manners above, the fusion prediction model is trained based on sample data carrying a standard evaluation label, in the training process, the model calculates the loss according to the predicted evaluation label output for the sample data and the standard evaluation label of the sample data, and further adjusts the parameters in the model according to the loss value, so that the model capable of effectively fusing internal and external information and accurately predicting is obtained through multiple iterative training.

[0020] In some possible implementation manners, the loss function comprises a performance loss function and a consistency loss function; the performance loss function is used to calculate a performance loss value between the predicted evaluation label and the standard evaluation label of each sample data, and the consistency loss function is used to calculate a consistency loss value between the pre-internal predicted evaluation label and the pre-external predicted evaluation label in the predicted evaluation label of each sample data.

[0021] By the technical solutions in the possible implementation manners, a multi-objective loss function for performance and consistency is designed for the fusion prediction model, the optimization of the model is constrained from two aspects of performance and consistency, the consistency between the fusion evaluation and the original internal evaluation is effectively improved, the unstable change of the original internal evaluation caused by unreliable external information is avoided, and the gain caused by reliable external information is retained.

[0022] In some possible implementation manners, in the training process, the credibility of the sample external evaluation of the sample data is determined according to the source of the sample external evaluation, the more real the source is, the higher the credibility is; the performance loss weight and the consistency loss weight corresponding to the sample data are determined according to the credibility, the higher the credibility is, the higher the performance loss weight of the sample data is; and the loss value is the sum of the loss values of all sample data, and the loss value of each sample data is the weighted sum of the performance loss value and the consistency loss value of each sample.

[0023] By the technical solutions in the possible implementation manners, the performance loss and the consistency loss have different weights in the total loss, so as to adjust the balance between the prediction performance and the internal consistency of the model. Then, for the sample data used for training, in addition to a part of the external information of the sample data being aggregated from the neighbor external information, a part of the external information of the sample data is directly queried from the outside, so that the credibility of the external information is determined by the source of the external information, and based on the sample data with different credibilities, the sample data can have different weights in the calculation of the performance loss and the consistency loss. The performance loss weight of the sample with reliable external information is higher, so that the model based on the sample with reliable external information can pay more attention to learning how to perform fusion prediction, that is, pay more attention to training the prediction performance; the sample with unreliable external information has a greater consistency loss weight, so that the model based on the sample with unreliable external information can pay more attention to learning how to keep the predicted result at a high internal consistency, and by controlling the loss weights of different samples, the performance can be improved while the consistency is considered.

[0024] In some possible implementation manners, after the fusion prediction model is used to fuse the external evaluation and the internal evaluation to obtain the overall evaluation output by the fusion prediction model, the method further comprises: performing user individualization analysis and / or product recommendation on the target user based on the overall evaluation.

[0025] By the technical solutions in the possible implementation manners above, after the target user is evaluated by fusing the internal and external information of the target user through the fusion prediction model, subsequent personalized service and product recommendation can be further performed on this basis, and user experience is optimized.

[0026] In a second aspect, the embodiments of the present disclosure provide a data fusion device, which comprises: a data input module configured to input a relationship network graph of a target user into a fusion prediction model, the relationship network graph taking a user entity as a node and a relationship between each user entity as an edge connecting the nodes, and the user entity being generated based on basic attributes and external information of the user; an external information aggregation module configured to aggregate neighbor external information of the target user based on the fusion prediction model to obtain target external information of the target user, and obtain an external evaluation corresponding to the target external information based on the fusion prediction model; an internal information access module configured to obtain an internal evaluation of the target user and access the internal evaluation to the fusion prediction model, the internal evaluation being obtained by prediction based on internal information of the target user; and a prediction output module configured to fuse the external evaluation and the internal evaluation based on the fusion prediction model to obtain an overall evaluation output by the fusion prediction model.

[0027] In a third aspect, the embodiments of the present disclosure provide a computer program product containing instructions, which, when executed on a computer or a processor, cause the computer or the processor to perform the steps of the method described above.

[0028] In a fourth aspect, the embodiments of the present disclosure provide a computer storage medium storing a plurality of instructions, which are adapted to be loaded and executed by a processor to perform the steps of the method described above.

[0029] In a fifth aspect, the embodiments of the present disclosure provide a terminal comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being adapted to be loaded and executed by the processor to perform the steps of the method described above. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the related art, brief introductions will be given to the drawings needed to be used in the embodiments or the related art descriptions. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0031] FIG. 1 is an exemplary system architecture diagram of a data fusion method according to an embodiment of the present disclosure;

[0032] FIG. 2 is a flow diagram of a data fusion method according to an embodiment of the present disclosure;

[0033] FIG. 3 is a flow diagram of a data fusion method according to an embodiment of the present disclosure;

[0034] FIG. 4 is an implementation logic block diagram of a data fusion method according to an embodiment of the present disclosure;

[0035] FIG. 5 is a structural block diagram of a data fusion apparatus according to an embodiment of the present disclosure;

[0036] FIG. 6 is a structural diagram of a terminal according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0037] In order to make the features and advantages of the embodiments of the present disclosure more apparent and easy to understand, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only some of the embodiments of the present disclosure, but not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the embodiments of the present disclosure.

[0038] The following description refers to the accompanying drawings. Unless otherwise indicated, same or similar elements in different drawings are denoted by same or similar reference numerals. The implementations described in the following exemplary embodiments are not meant to represent all implementations consistent with the embodiments of the present disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present disclosure as detailed in the appended claims. And in the description of the embodiments of the present disclosure, unless otherwise specified, " / " means or, for example, A / B can mean A or B: "and / or" in the text only represents a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases of A alone, A and B together, and B alone. In addition, in the description of the embodiments of the present disclosure, "multiple" means two or more than two.

[0039] Hereinafter, the terms "first" and "second" are used only for descriptive purposes, and cannot be understood as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features.

[0040] In the digital era, data, as an indispensable core asset, drives the internal operation optimization and efficiency improvement of enterprises. In this context, important data is often scattered in different enterprises and institutions. For a single enterprise platform, in addition to making full use of internal data, it also needs to obtain high-quality external data. Based on rich internal and external data, better decisions can be made to better solve user problems. External data, often referred to as information data collected from outside the enterprise, includes but is not limited to public data resources, social media data, deep cooperation data shared with business partners, market research and analysis reports, etc. For example, in the financial credit scenario, in order to comprehensively and accurately evaluate the credit characteristics of users, the credit information of users held by partners needs to be obtained to score the credit rating of users, and the credit limit of the user is adjusted based on the credit score of the user. Therefore, it can be seen that these external data can reflect the real face and potential dynamics of the market from different dimensions, so that enterprises can timely understand the industry dynamics and thus develop more accurate and effective market strategies. Through the deep integration of big data and artificial intelligence technology, enterprises can deeply mine the demand preferences, behavior habits and even emotional tendencies of consumers based on rich internal and external data, thereby providing more personalized and intelligent product and service experiences for consumers.

[0041] However, in the context of data-driven decision-making, obtaining high-quality external data is not easy, which often requires considerable cost, including data source procurement, complex data cleaning and processing procedures, and multiple measures to ensure data compliance and security. Therefore, the external data resources that enterprises can obtain in actual operation are often strictly limited, making it difficult to achieve comprehensive and in-depth coverage of a large user group. Enterprises usually adopt targeted acquisition strategies for external data, and query external data for a part of users each time they need to use external data. This group of users may be randomly selected from the total number of users, or may be selected under certain screening conditions. Further, the customer groups that enterprises focus on at different times will change, which makes it difficult for enterprises to continuously track the same user. The discontinuity of external information acquisition directly affects the stability and reliability of information when enterprises make decisions based on the fusion of internal and external data, increasing the uncertainty and risk in the decision-making process.

[0042] In addition to the challenges in data acquisition, there are also significant differences in format, structure, and quality between external information and internal data, which are particularly prominent in the data fusion process. If there is a lack of effective fusion strategies and model frameworks, not only is it difficult to fully realize the value of external data, but also it may lead to significant deviations in user evaluation and decision-making due to information mismatch or conflict, affecting the accuracy and effectiveness of decisions.

[0043] Therefore, the embodiments of the present disclosure provide a data fusion method to solve the technical problem that internal information and external information are difficult to be effectively fused.

[0044] Please refer to FIG. 1, which is an exemplary system architecture diagram of a data fusion method provided by the embodiments of the present disclosure.

[0045] As shown in FIG. 1, the system architecture can include a terminal 101, a network 102 and a server 103. The network 102 is used to provide a communication link medium between the terminal 101 and the server 103. The network 102 can include various types of wired communication links or wireless communication links, for example, the wired communication links include optical fiber, twisted pair or coaxial cable, and the wireless communication links include Bluetooth communication link, Wireless-Fidelity (Wi-Fi) communication link or microwave communication link, etc.

[0046] The terminal 101 can interact with the server 103 through the network 102 to receive messages from the server 103 or send messages to the server 103, or the terminal 101 can interact with the server 103 through the network 102 to receive messages or data sent by other users to the server 103. The terminal 101 can be hardware or software. When the terminal 101 is hardware, it can be various electronic devices, including but not limited to smart watches, smart phones, tablet computers, laptop computers and desktop computers, etc. When the terminal 101 is software, it can be installed in the above-mentioned electronic devices, which can be implemented as multiple software or software modules (for example, used to provide distributed services) or a single software or software module, which is not specifically limited here.

[0047] In the embodiments of the present disclosure, the terminal 101 first inputs the relationship network graph of the target user into the fusion prediction model, the relationship network graph takes the user entity as the node and the relationship between each user entity as the edge connecting the nodes, and the user entity is generated based on the basic attributes and external information of the user; then, the terminal 101 can aggregate the neighbor external information of the target user based on the fusion prediction model to obtain the target external information of the target user, and obtain the external evaluation corresponding to the target external information based on the fusion prediction model; next, the terminal 101 obtains the internal evaluation of the target user and inputs the internal evaluation into the fusion prediction model, and the internal evaluation is obtained based on the internal information of the target user; finally, the terminal 101 can fuse the external evaluation and the internal evaluation based on the fusion prediction model to obtain the overall evaluation output by the fusion prediction model.

[0048] The server 103 can be a service server providing various services. It should be noted that the server 103 can be hardware or software. When the server 103 is hardware, it can be implemented as a distributed server cluster composed of multiple servers or as a single server. When the server 103 is software, it can be implemented as multiple software or software modules (for example, to provide distributed services) or as a single software or software module, which is not limited here.

[0049] Alternatively, the system architecture can also not include the server 103, in other words, the server 103 can be an optional device in the embodiments of the present disclosure, that is, the method provided by the embodiments of the present disclosure can be applied to a system structure including only the terminal 101, which is not limited by the embodiments of the present disclosure.

[0050] It should be understood that the number of terminals, networks and servers in FIG. 1 is only illustrative, and can be any number of terminals, networks and servers according to the needs of implementation.

[0051] Please refer to FIG. 2, which is a flowchart of a data fusion method provided by the embodiments of the present disclosure. The execution subject of the embodiments of the present disclosure can be a terminal executing data fusion, a processor in the terminal executing the data fusion method, or a data fusion service in the terminal executing the data fusion method. For convenience of description, the specific execution process of the data fusion method is introduced below with the execution subject being the processor in the terminal as an example.

[0052] As shown in FIG. 2, the data fusion method can at least include the following steps.

[0053] S202, input the relationship network graph of the target user into the fusion prediction model, the relationship network graph takes the user entity as the node and the relationship between the user entities as the edge connecting the nodes, and the user entity is generated based on the basic attributes and external information of the user.

[0054] Optionally, when fusing data from different data sources, a machine learning technique such as a tree model is usually used to train the internal and external information to obtain a prediction result based on the fused information, thereby replacing the model result based on the internal information. However, due to limitations in actual scenarios, the acquisition of external information can have discontinuity, for example, a user had external information last month, but no new external information was queried for the user this month, so the external information for the user this month is missing. In the face of this information missing problem, there are two common processing methods, one is to select the missing external information for null processing, which is more direct and simple in technical implementation, but this also leads to a significant jump or fluctuation in the score or behavior pattern of the user predicted based on internal and external information between the months with complete and missing information, and the discontinuity of the information directly affects the stability and prediction ability of the model, making it difficult to maintain the accuracy of the model output. The other processing method is to directly continue to use the historical external information that can be retrieved for the user to make predictions, although this method ensures the continuity of the information, but in the long run, it ignores the sensitivity of user behavior over time, especially in a rapidly changing market environment or user preferences, this "stagnant" information usage method is easy to make the model appear inadequate when capturing the current decision preference of the user, thereby adversely affecting the decision performance and reducing the timeliness and adaptability of the model.

[0055] Optionally, directly nulling or directly using historical external information for the user will make it difficult for the model to output stable and accurate prediction results for the user. Therefore, in order to ensure that reliable external information can be used as a basis for predicting the user each time, some real external information possessed by users can be diffused to users without real external information after each query for external information, thereby improving the utilization rate of limited external information and enabling each user to have external information with high timeliness.

[0056] Further, at present, similarity retrieval and diffusion through homogeneous K-Nearest Neighbor algorithm is a common strategy in data diffusion. This strategy uses the similarity of the intrinsic characteristics of users as a bridge to realize the integration and sharing of external information resources between similar users based on the similarity between users. However, this method still faces two major challenges in actual operation. One is that similarity is difficult to measure, and the similarity derived based on the intrinsic characteristics of users does not always accurately map to their external information environment. Those external information valuable to a user may have some degree of difference with the internal information rather than simple repetition. That is, the conduction of similarity is not necessarily true, and the similarity calculated based on the internal characteristics is difficult to determine as external similarity. The other challenge comes from the limitation of the external information query group. The distribution of the external information query group is often biased, which makes it difficult to ensure the universality and balance of information diffusion.

[0057] In the embodiments of the present disclosure, for the problem that similarity is difficult to directly quantify, the concept of relationship network is introduced, which marks the in-depth exploration of the connection between complex data. The relationship network is a multi-dimensional and dynamic mapping system, which contains the objective and comprehensive correlation between users. For the problem of biased query users and difficult information extrapolation, the introduction of relationship network can make the information diffusion not be limited to "similarity", but adaptively learn the correlation between users.

[0058] Specifically, the relationship network graph of users can be constructed by using the complex relationship between users as a medium. That is, first, the user data of all users is obtained, and the user data includes basic attributes and external information. The basic attributes are used to represent the simple characteristics of the user, also known as user portrait information, such as occupation, city, etc. According to the user data of each user, the user entity corresponding to each user can be generated, and the relationship between each user entity can also be extracted based on the specific information of the user entity. This relationship is extracted based on attributes and external information, and therefore can reflect the objective correlation between users. Further, taking the user entity as a node and the relationship as an edge, the full-amount relationship network graph corresponding to all users can be obtained. When a target user needs to be predicted, the relationship network graph corresponding to the target user can be directly retrieved in the full-amount relationship network graph.

[0059] Optionally, the relationship network graph is processed by a graph neural network (GNN) model to realize adaptive diffusion of external information. The graph neural network model is a kind of deep learning model for processing graph structure data. Entities in the graph structure data are represented in the form of nodes, and the relationships between the entities are represented in the form of edges. The GNN can learn useful representations in the graph structure data and use these representations for various tasks. The core idea of the graph neural network is to aggregate neighbor information of a node through a message-passing mechanism and update the feature representation of the node. Therefore, the graph neural network model is very suitable for processing the relationship network graph to aggregate external information of neighbor users to a target user. Based on this, in the embodiments of the present disclosure, a graph neural network is pre-trained as a fusion prediction model. After obtaining the relationship network graph containing the target user entity and the neighbor user entities having a relationship with the target user, the target user's relationship network graph can be input into the fusion prediction model. Through the fusion prediction model, external information of neighbor nodes can be aggregated for the target user's node without direct external information sources.

[0060] In S204, neighbor external information of the target user is aggregated based on the fusion prediction model to obtain target external information of the target user, and an external evaluation corresponding to the target external information is obtained based on the fusion prediction model.

[0061] Optionally, after the relationship network graph is input into the fusion prediction model, the fusion prediction model can aggregate neighbor external information based on the attributes of the user and the neighbor and the external information, and the information of the edge to obtain target external information of the target user. According to the target external information, the target user entity of the target user can be updated. Further, the target external information corresponding to the external evaluation can be obtained by predicting the target user entity based on the fusion prediction model. In this way, the external information is adaptively learned and propagated based on the objective association between users, which improves the utilization efficiency of the external information while reducing the model maintenance cost, and significantly improves the utilization rate of the external information.

[0062] In the embodiments of the present disclosure, only for the users who do not obtain real external information directly from the external information source, external information can be given through the diffusion of neighbor external information, while for the users who have directly obtained real external information, external evaluation and internal-external information fusion can be directly based on the real external information. In another possible implementation, neighbor external information can also be aggregated for all users. For the users who have directly obtained real external information, a lower weight can be given to the external information aggregated from the neighbor external information, and a higher weight can be given to the real external information owned by the user himself. Then, the external information of the two parts is calculated again, and finally the target external information of the user is obtained.

[0063] S206, obtain internal evaluation of the target user and input the internal evaluation into the fusion prediction model, the internal evaluation being obtained based on internal information of the target user.

[0064] Optionally, after obtaining the external evaluation of the target user based on the relationship network graph, the model further incorporates internal evaluation obtained based on internal information of the target user, and fuses the external evaluation and the internal evaluation, so as to realize fusion of internal and external information.

[0065] S208, fuse the external evaluation and the internal evaluation based on the fusion prediction model to obtain overall evaluation output by the fusion prediction model.

[0066] Optionally, after the internal evaluation and the external evaluation are both clear, the fusion prediction model fuses the external evaluation and the internal evaluation to output the overall evaluation corresponding to the target user. The overall evaluation is the prediction result fused with internal and external information, and therefore, based on the overall evaluation, user personalized analysis and / or product recommendation are performed on the target user, and the overall evaluation corresponding to the user is applied to user personalized service and product recommendation, so that user behavior information and feature information implied in the overall evaluation are used to provide services and products that are more in line with the needs of the user, thereby optimizing user experience.

[0067] In the embodiments of the present disclosure, a data fusion method is provided, which inputs a relationship network graph of a target user into a fusion prediction model, the relationship network graph taking a user entity as a node and a relationship between each user entity as an edge connecting the nodes, and the user entity being generated based on basic attributes and external information of the user; target external information of the target user is obtained by aggregating neighbor external information of the target user based on the fusion prediction model, and an external evaluation corresponding to the target external information is obtained based on the fusion prediction model; internal evaluation of the target user is obtained and input into the fusion prediction model, the internal evaluation being obtained based on internal information of the target user; and the external evaluation and the internal evaluation are fused based on the fusion prediction model to obtain overall evaluation output by the fusion prediction model. Since the node relationship in the relationship network graph can transmit external information of other neighbor users to the target user through the relationship, the external information of the neighbor users is aggregated to obtain the external information of the target user, so that the external information is adaptively propagated and information gain is generated; then in the fusion prediction model, the external evaluation corresponding to the external information and the internal evaluation are further fused to output the final evaluation of the target user. In this way, the external information is adaptively learned and propagated based on the objective association between users, the utilization efficiency of the external information is improved while the model maintenance cost is reduced, the utilization rate of the external information is significantly improved, and the reliability and accuracy of the final evaluation result are improved.

[0068] Please refer to FIG. 3, which is a flow diagram of a data fusion method according to an embodiment of the present disclosure.

[0069] As shown in FIG. 3, the data fusion method can at least include the following steps.

[0070] S302, input the relationship network graph of the target user into the fusion prediction model, the relationship network graph takes the user entity as a node and the relationship between each user entity as an edge connecting the nodes, and the user entity is generated based on the basic attributes and external information of the user.

[0071] For step S302, please refer to the detailed description in step S202, which will not be repeated here.

[0072] S304, determine the aggregation weight corresponding to each neighbor user according to the relationship between the target user entity and each neighbor user entity in the relationship network graph, and the size of the aggregation weight is positively correlated with the closeness of the relationship.

[0073] Optionally, please refer to FIG. 4, which is an implementation logic framework diagram of a data fusion method according to an embodiment of the present disclosure. As shown in FIG. 4, for the fusion prediction model, it includes an input layer, a model layer, and an output layer. The input layer is a module for data input, and the relationship network graph of the target user is input from the input layer to the fusion prediction model. In the model layer of the fusion prediction model, a series of processing is performed on the input data, including but not limited to external information diffusion based on nodes and edges in the relationship network, fusion of internal evaluation and external evaluation, etc. The output layer is a module for outputting the overall evaluation in the fusion prediction model.

[0074] Specifically, in the model layer, the fusion prediction model first aggregates the neighbor external information of the target user to obtain the target external information of the target user. Considering that the closeness of the relationship between users is not the same, the more closely related two users are, the more homogeneous their external information is. Based on this, when aggregating the external information of multiple neighbors of the target user, the relationship between each neighbor user and the target user in the relationship network graph can be considered, so that the neighbor user with a closer relationship with the target user can obtain a higher aggregation weight for its external information. In this way, controlling the aggregation weight of each neighbor external information based on the closeness of the relationship can make each neighbor external information have a more appropriate influence on the aggregation result based on the closeness of the relationship.

[0075] S306, in the fusion prediction model, the external information of each neighbor user is weighted and averaged based on the aggregation weight of each neighbor user, the target external information of the target user is obtained, and the external evaluation corresponding to the target external information is obtained based on the fusion prediction model.

[0076] Optionally, each neighbor external information is given different aggregation weights based on the closeness of the relationship with the target user, and then the external information of each neighbor user is weighted and averaged in the fusion prediction model to obtain the target external information of the target user.

[0077] Optionally, in order to continuously provide personalized services to users, the evaluation of the user is updated at regular intervals, and in order to improve the stability of the model when the user is scored multiple times, historical external information can be introduced to output the evaluation each time. However, too much historical information or too long time in a long window can easily have a negative impact on the performance of the model. Therefore, in the embodiment of the present disclosure, considering that the acquisition time of the external information of each user can be different, and for the current prediction, the external information acquired at a time close to the current time is more reliable than the external information acquired long ago, the attention weight of each neighbor external information can be controlled based on the timeliness of each neighbor external information when aggregating the neighbor external information, so that the external information acquired at a time closer to the current time has a larger attention weight, thereby improving the stability of the continuous evaluation of the model on the user, and avoiding the negative impact of long-term historical information on the performance.

[0078] Further, referring to FIG. 4, the fusion prediction model can perform time series information processing while the neighbor external information is diffused. That is, the first weighted average calculation of the external information of each neighbor user based on the aggregation weight of each neighbor user obtains the first external information of the target user, and the second weighted average calculation of each external information based on the attention weight of each external information obtains the second external information of the target user. According to the first external information and the second external information, the target external information of the target user is finally obtained. Through the adaptive mechanism, the timeliness of the external information is processed, the stability of the continuous evaluation output by the model is optimized, and the noise caused by the historical information in the long window is avoided.

[0079] Optionally, referring to FIG. 4, the fusion prediction model can predict the target user entity based on the target external information to obtain the external evaluation corresponding to the target external information. In this way, the external information is adaptively learned and propagated based on the objective association between users, the utilization efficiency of the external information is improved, the maintenance cost of the model is reduced, and the utilization rate of the external information is significantly improved.

[0080] S308, obtaining the internal evaluation of the target user and inputting the internal evaluation into the fusion prediction model, and inputting the internal evaluation into the module where the external evaluation is located in the fusion prediction model in the form of residual connection.

[0081] Optionally, after obtaining the external evaluation of the target user based on the relationship network graph, the model also incorporates the internal evaluation obtained based on the internal information of the target user. Please refer to FIG. 4, the internal evaluation can be connected to the module where the external evaluation is located in the fusion prediction model through a residual connection. The residual connection is generally used to solve the problem of gradient disappearance or gradient explosion as the network depth increases. The core idea of the residual connection is to introduce the output of the previous layer as part of the input of the subsequent layer, which can directly propagate the gradient and help the training of the deep network. Through the residual connection, the internal evaluation can be directly introduced into the network layer where the external evaluation is located in the model, which helps the gradient to flow back more effectively during the training process, avoids the loss of gradient information as the model depth increases after the input of the internal evaluation in the front end of the model, and thus improves the volume of the original internal information during the fusion of internal and external information, and realizes more accurate and effective fusion of internal and external information.

[0082] S310, in the fusion prediction model, the credibility of the external evaluation is determined according to the source of the external evaluation, and the more real the source is, the higher the credibility is.

[0083] Optionally, when fusing the internal evaluation and the external evaluation, the fusion prediction model also considers that if the user has real external information, the external information is consistent with the real situation of the user, and the credibility is higher, so the gain for the final overall evaluation is larger; if the target external information is aggregated from other neighbor external information, the credibility is not as high as the real external information; for the user who has no external information and is difficult to aggregate enough external information from the neighbors, the credibility of the external information is low. The credibility of the external information determines the credibility of the external evaluation, that is, the fusion prediction model can determine the credibility of the external evaluation according to the source of the external evaluation, and the more real the source is, the higher the credibility is.

[0084] S312, the attention weight of the external evaluation and the internal evaluation is respectively assigned according to the credibility, and the higher the credibility is, the higher the attention weight of the external evaluation is.

[0085] Optionally, the attention weight of the external evaluation and the internal evaluation can be respectively assigned according to the credibility of the external evaluation, and the higher the credibility is, the higher the attention weight of the external evaluation is, for example, for the user who has real external information obtained from the outside, the weight of the external evaluation is higher when the model fuses the internal and external evaluations; if the user has no external information, the weight of the internal evaluation is higher when fusing; and for the user whose external information is aggregated from the neighbors, the weights of the internal and external evaluations are equal when fusing, so that the weight of the external evaluation is adjusted according to the credibility of the external evaluation of the user when fusing, so that the customer group with large gain of external information can retain the gain as much as possible, and the customer group with small gain can maintain the internal consistency.

[0086] S314, based on the external evaluation attention weight and the internal evaluation attention weight, weighting and fusing the external evaluation and the internal evaluation to obtain the overall evaluation of the fusion prediction model output.

[0087] Optionally, based on the external evaluation attention weight and the internal evaluation attention weight, weighting and fusing the external evaluation and the internal evaluation to obtain the overall evaluation of the fusion prediction model output, so that different credibility of external evaluation of different customer groups has a proper impact on the overall evaluation, and the gain effect of effective external information is ensured.

[0088] In the embodiments of the present disclosure, the closer the relationship between the target user and the neighbor user, the higher the aggregation weight of the external information of the neighbor user can be obtained. The target external information of the target user is calculated by weighting the external information of each neighbor user, so that each neighbor external information has a more appropriate impact on the aggregation result based on the closeness of the relationship. A monotonous learnable attention weight is added to the external information, so that the longer the acquisition time of the external information, the weaker the attention to this part of information, thereby improving the stability of the model for continuous evaluation of the user, and avoiding the negative impact of long-term historical information on the performance. The internal evaluation obtained based on the internal information is introduced into the model through a residual connection when the external evaluation is obtained, so as to avoid the disappearance of gradient information caused by the increase of model depth after the internal evaluation is input at the front end of the model, thereby improving the volume of the original internal information during the fusion of internal and external information, and realizing more accurate and effective fusion of internal and external information. According to the credibility of the external evaluation of the user, the weight of the external evaluation during the fusion is adjusted, so that the customer group with large gain of external information can retain the gain as much as possible, and the customer group with small gain can maintain internal consistency.

[0089] In the introduction of the above embodiments, the fusion prediction model is used for the diffusion of external information and the fusion of internal and external information. Before the fusion prediction model is used in the application scenario, the fusion prediction model needs to be trained first, and after the fusion prediction model converges to the expected performance, it is deployed in the actual scene to complete the fusion and prediction of internal and external information.

[0090] In one specific embodiment provided by the present disclosure, the fusion prediction model is trained based on at least one sample data carrying a standard evaluation label. During the training process of the fusion prediction model, the model outputs a predicted evaluation label for each sample data, and then calculates a loss value of a loss function according to the predicted evaluation label and the standard evaluation label of each sample data, so as to adjust the model parameters based on the loss value, until the fusion prediction model meets a preset convergence condition. After multiple iterations of training, a model capable of effectively fusing internal and external information and accurately predicting is finally obtained.

[0091] Specifically, a multi-objective loss function for performance and consistency rate can be designed for the fusion prediction model to constrain the optimization of the model from both performance and consistency. That is, the loss function of the fusion prediction model can include two parts: a performance loss function and a consistency loss function; wherein the performance loss function is used to calculate the performance loss value between the predicted evaluation label and the standard evaluation label of each sample data, and the consistency loss function is used to calculate the consistency loss value between the predicted internal evaluation label and the predicted external evaluation label in the predicted evaluation label of each sample data. In this way, the consistency between the fused evaluation and the original internal evaluation can be effectively improved, the unstable changes caused by unreliable external information to the original internal evaluation can be avoided, and the gain brought by reliable external information can be retained.

[0092] Further, in the above training process, in order to balance the prediction performance and internal-external consistency rate of the model, the performance loss weight of the performance loss function and the consistency loss weight of the consistency loss function can be allocated to each sample data adaptively. For the sample data used for training, in addition to a part of the sample data whose external information is aggregated from neighbor external information, there is also a part of the sample data whose external information is directly queried from the outside, so the credibility of the sample external evaluation can be determined according to the source of the sample external evaluation, and the more real the source is, the higher the credibility is. Further, for these sample data with different credibility, different weights can be given to the calculation of performance loss and consistency loss, the higher the credibility is, the higher the performance loss weight of the sample data is, that is, these sample data can be used to focus on improving the prediction performance of the model; the lower the credibility is, the higher the consistency loss weight of the sample data is, that is, these sample data can be used to focus on making the model have higher internal consistency. The final loss value is the sum of the loss values of all sample data, and the loss value of each sample data is the weighted sum of the performance loss value and the consistency loss value of each sample. The higher the performance loss weight of the sample with reliable external information is, the more the model based on the sample with reliable external information can focus on learning how to make fusion prediction, that is, it focuses more on training the prediction performance; the consistency loss weight of the sample with less reliable external information is larger, so that the model can focus more on learning how to keep the prediction result with higher internal consistency based on the sample with unreliable external information. By controlling the loss weight of different samples, the performance can be improved while the consistency rate is considered.

[0093] Referring to FIG. 5, FIG. 5 is a structural block diagram of a data fusion device provided by an embodiment of the present disclosure. As shown in FIG. 5, the data fusion device 500 includes: a data input module 510, configured to input a relationship network graph of a target user into a fusion prediction model, the relationship network graph taking a user entity as a node and a relationship between user entities as an edge connecting the nodes, and the user entity being generated based on basic attributes of the user and external information; an external information aggregation module 520, configured to aggregate neighbor external information of the target user based on the fusion prediction model to obtain target external information of the target user, and obtain an external evaluation corresponding to the target external information based on the fusion prediction model; an internal information access module 530, configured to obtain an internal evaluation of the target user and access the internal evaluation into the fusion prediction model, the internal evaluation being obtained by prediction based on internal information of the target user; and a prediction output module 540, configured to fuse the external evaluation and the internal evaluation based on the fusion prediction model to obtain an overall evaluation output by the fusion prediction model.

[0094] Optionally, the data fusion device 500 further includes a relationship network construction module, configured to obtain user data of all users, the user data including the basic attributes and the external information; generate a user entity corresponding to each user and extract a relationship between the user entities based on the user data of the users, and construct a full-quantity relationship network graph corresponding to all users based on the user entities and the relationship between the user entities; and obtain a relationship network graph of the target user, the relationship network graph including a target user entity of the target user and at least one neighbor user entity having a relationship with the target user entity.

[0095] Optionally, the external information aggregation module 520 is further configured to update the target user entity of the target user according to the target external information, and obtain the external evaluation corresponding to the target external information by prediction on the target user entity based on the fusion prediction model.

[0096] Optionally, the external information aggregation module 520 is further configured to determine an aggregation weight corresponding to each neighbor user according to the relationship between the target user entity and each neighbor user entity in the relationship network graph, the aggregation weight being positively correlated with the closeness of the relationship, and perform weighted average calculation on the external information of each neighbor user based on the aggregation weight of each neighbor user in the fusion prediction model to obtain the target external information of the target user.

[0097] Optionally, the external information aggregation module 520 is further configured to, in the fusion prediction model, add attention weights to each external information based on the acquisition time of the external information of each neighbor user, and the closer the acquisition time is to the current time, the greater the attention weight of the external information; perform first weighted average calculation on the external information of each neighbor user based on the aggregation weight of each neighbor user to obtain first external information of the target user, and perform second weighted average calculation on each external information based on the attention weight of each external information to obtain second external information of the target user; and calculate target external information of the target user according to the first external information and the second external information.

[0098] Optionally, the internal information access module 530 is further configured to access the internal evaluation into the module where the external evaluation is located in the fusion prediction model in a residual connection manner.

[0099] Optionally, the prediction output module 540 is further configured to, in the fusion prediction model, determine the credibility of the external evaluation according to the source of the external evaluation, and the more real the source is, the higher the credibility is; assign attention weights to the external evaluation and the internal evaluation according to the credibility, and the higher the credibility is, the higher the attention weight of the external evaluation is; and perform weighted fusion on the external evaluation and the internal evaluation based on the attention weight of the external evaluation and the attention weight of the internal evaluation to obtain the overall evaluation output by the fusion prediction model.

[0100] Optionally, the fusion prediction model is trained based on at least one sample data carrying a standard evaluation label; in the training process of the fusion prediction model, a predicted evaluation label is output for each sample data, a loss value of a loss function is calculated according to the predicted evaluation label and the standard evaluation label of each sample data, and the model parameters are adjusted based on the loss value until the fusion prediction model meets a preset convergence condition.

[0101] Optionally, the loss function includes a performance loss function and a consistency loss function; the performance loss function is used to calculate a performance loss value between the predicted evaluation label and the standard evaluation label of each sample data, and the consistency loss function is used to calculate a consistency loss value between the predicted internal evaluation label and the predicted external evaluation label in the predicted evaluation label of each sample data.

[0102] Optionally, in the training process, the credibility of the sample external evaluation is determined according to the source of the sample external evaluation, and the more real the source is, the higher the credibility is; the performance loss weight and the consistency loss weight corresponding to the sample data are determined according to the credibility, and the higher the credibility is, the higher the performance loss weight of the sample data is; and the loss value is the sum of the loss values of all sample data, and the loss value of each sample data is the weighted sum of the performance loss value and the consistency loss value of each sample.

[0103] Optionally, the data fusion apparatus 500 further comprises an application module, configured to perform user individual analysis and / or product recommendation on the target user based on the overall evaluation.

[0104] The embodiments of the present disclosure provide a computer program product containing instructions, which, when executed on a computer or processor, cause the computer or processor to perform the steps of the method of any of the above embodiments.

[0105] The embodiments of the present disclosure further provide a computer storage medium, which can store a plurality of instructions, and the instructions are suitable for being loaded and executed by a processor to perform the steps of the method of any of the above embodiments.

[0106] Please refer to FIG. 6, which is a structural schematic diagram of a terminal according to an embodiment of the present disclosure. As shown in FIG. 6, the terminal 600 can include at least one terminal processor 601, at least one network interface 604, a user interface 603, a memory 605, and at least one communication bus 602.

[0107] The communication bus 602 is configured to realize the connection and communication among the components.

[0108] The user interface 603 can include a display screen (Display) and a camera (Camera), and optionally, the user interface 603 can further include a standard wired interface and a wireless interface.

[0109] The network interface 604 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0110] The terminal processor 601 can include one or more processing cores. The terminal processor 601 connects various parts within the terminal 600 through various interfaces and lines, executes various functions of the terminal 600 and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 605, and calling data stored in the memory 605. Alternatively, the terminal processor 601 can be implemented in at least one of a hardware form of a digital signal processing (DSP), a field-programmable gate array (FPGA), and a programmable logic array (PLA). The terminal processor 601 can be integrated with a combination of one or more of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU is mainly used to process operating systems, user interfaces, and application programs; the GPU is used to render and draw the content to be displayed on the display screen; and the modem is used to process wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the terminal processor 601, but can be implemented by a separate chip.

[0111] The memory 605 can include a random access memory (RAM) and a read-only memory (ROM). Optionally, the memory 605 includes a non-transitory computer-readable storage medium. The memory 605 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 605 can include a program storage area and a data storage area, wherein the program storage area can store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area can store data involved in the above-mentioned various method embodiments, etc. The memory 605 can also be at least one storage device located away from the aforementioned terminal processor 601. As shown in FIG. 6, the memory 605 as a computer storage medium can include an operating system, a network communication module, a user interface module, and a data fusion program.

[0112] In the terminal 600 shown in FIG. 6, the user interface 603 is mainly used to provide an interface for the user to input and obtain data input by the user; and the terminal processor 601 can be used to call the data fusion program stored in the memory 605 and specifically perform the following operations: inputting the relationship network graph of the target user into the fusion prediction model, the relationship network graph taking the user entity as a node and the relationship between the user entities as an edge connecting the nodes, and the user entity being generated based on the basic attributes and external information of the user; obtaining the target external information of the target user based on the fusion prediction model by aggregating the neighbor external information of the target user, and obtaining the external evaluation corresponding to the target external information based on the fusion prediction model; obtaining the internal evaluation of the target user and inputting the internal evaluation into the fusion prediction model, the internal evaluation being obtained by predicting based on the internal information of the target user; and fusing the external evaluation and the internal evaluation based on the fusion prediction model to obtain the overall evaluation output by the fusion prediction model.

[0113] In some embodiments, before the terminal processor 601 performs the operation of inputting the relationship network graph of the target user into the fusion prediction model, it specifically performs the following steps: obtaining the user data of all users, the user data including the basic attributes and external information; generating the user entity corresponding to each user and extracting the relationship between each user entity based on the user data of each user, and constructing the full-quantity relationship network graph corresponding to all users based on each user entity and the relationship between each user entity; and obtaining the relationship network graph of the target user, the relationship network graph containing the target user entity of the target user and at least one neighbor user entity having a relationship with the target user entity.

[0114] In some embodiments, when the terminal processor 601 performs the operation of obtaining the external evaluation corresponding to the target external information based on the fusion prediction model, it specifically performs the following steps: updating the target user entity of the target user according to the target external information, and obtaining the external evaluation corresponding to the target external information by predicting the target user entity based on the fusion prediction model.

[0115] In some embodiments, when the terminal processor 601 performs the operation of obtaining the target external information of the target user based on the fusion prediction model by aggregating the neighbor external information of the target user, it specifically performs the following steps: determining the aggregation weight corresponding to each neighbor user according to the relationship between the target user entity and each neighbor user entity in the relationship network graph, the aggregation weight being positively correlated with the closeness of the relationship; and performing weighted average calculation on the external information of each neighbor user in the fusion prediction model based on the aggregation weight of each neighbor user to obtain the target external information of the target user.

[0116] In some embodiments, the terminal processor 601, when performing the weighted average calculation of the external information of each neighbor user based on the aggregation weight of each neighbor user in the fusion prediction model to obtain the target external information of the target user, specifically performs the following steps: in the fusion prediction model, attention weights are added to the external information of each neighbor user based on the acquisition time of each external information, and the closer the acquisition time of the external information to the current time, the greater the attention weight obtained by the external information; the first external information of the target user is obtained by performing the first weighted average calculation on the external information of each neighbor user based on the aggregation weight of each neighbor user, and the second external information of the target user is obtained by performing the second weighted average calculation on the external information based on the attention weight of each external information; and the target external information of the target user is calculated according to the first external information and the second external information.

[0117] In some embodiments, the terminal processor 601, when performing the internal evaluation into the fusion prediction model, specifically performs the following steps: the internal evaluation is connected to the module where the external evaluation is located in the fusion prediction model in the form of a residual connection.

[0118] In some embodiments, the terminal processor 601, when performing the fusion of the external evaluation and the internal evaluation based on the fusion prediction model to obtain the overall evaluation output by the fusion prediction model, specifically performs the following steps: in the fusion prediction model, the credibility of the external evaluation is determined according to the source of the external evaluation, and the more real the source, the higher the credibility; the attention weight of the external evaluation and the internal evaluation is respectively assigned according to the credibility, and the higher the credibility, the higher the attention weight of the external evaluation; and the external evaluation and the internal evaluation are weighted and fused based on the attention weight of the external evaluation and the attention weight of the internal evaluation to obtain the overall evaluation output by the fusion prediction model.

[0119] In some embodiments, the fusion prediction model is trained based on at least one sample data carrying a standard evaluation label; in the training process of the fusion prediction model, a predicted evaluation label is output for each sample data, a loss value of a loss function is calculated according to the predicted evaluation label and the standard evaluation label of each sample data, and the model parameters are adjusted based on the loss value until the fusion prediction model meets a preset convergence condition.

[0120] In some embodiments, the loss function includes a performance loss function and a consistency loss function; the performance loss function is used to calculate the performance loss value between the predicted evaluation label and the standard evaluation label of each sample data, and the consistency loss function is used to calculate the consistency loss value between the predicted internal evaluation label and the predicted external evaluation label in the predicted evaluation label of each sample data.

[0121] In some embodiments, in the training process, the credibility of the out-of-sample evaluation of the sample data is determined according to the source of the out-of-sample evaluation of the sample data, the more real the source is, the higher the credibility is; the performance loss weight and the consistency loss weight corresponding to the sample data are determined according to the credibility, the higher the credibility is, the higher the performance loss weight of the sample data is; the loss value is the sum of the loss values of all sample data, and the loss value of each sample data is the weighted sum of the performance loss value and the consistency loss value of each sample.

[0122] In some embodiments, after the terminal processor 601 performs fusion of the external evaluation and the internal evaluation based on the fusion prediction model to obtain the overall evaluation of the fusion prediction model output, the terminal processor 601 further specifically performs the following steps: performing user personalization analysis and / or product recommendation on the target user based on the overall evaluation.

[0123] In several embodiments provided in the present disclosure, it should be understood that the disclosed apparatus and method can be implemented in other manners. For example, the division of the above-described apparatus embodiments is merely illustrative, and for example, the division of the modules is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different modules can be indirect couplings or communication connections through some interfaces, devices or modules, and can be electrical, mechanical or in other forms.

[0124] The modules illustrated as separate components can or can not be physically separate, and the components illustrated as modules can or can not be physical modules, i.e., can be located in one place, or can be distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0125] In the above embodiments, all or part can be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part can be implemented in the form of a computer program product. The above computer program product includes one or more computer instructions. When loaded and executed on a computer, the above computer program instructions generate all or part of the processes or functions described in the above embodiments of the present disclosure. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted by the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that includes one or more available media sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a digital versatile disc (DVD)), or a semiconductor medium (for example, a solid state disk (SSD)), etc.

[0126] It should be noted that, for the foregoing method embodiments, in order to facilitate description, they are all described as a combination of a series of actions, but those skilled in the art should know that the embodiments of the present disclosure are not limited by the order of the described actions, because according to the embodiments of the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of the present disclosure.

[0127] In addition, it should be further noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of the present disclosure are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data need to comply with relevant laws, regulations, and standards of relevant countries and regions. For example, the external information, basic attributes, internal information, user data, etc. involved in the present disclosure are obtained under sufficient authorization.

[0128] The above describes specific embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order and still achieve desirable results. Additionally, the processes depicted in the figures do not necessarily require the particular order shown or sequential order in order to achieve the desired results. In some implementations, multitasking and parallel processing can be advantageous or necessary.

[0129] In the above embodiments, the description of each embodiment is focused on, and the part not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0130] The above is a description of a data fusion method, device, storage medium and terminal provided by an embodiment of the present disclosure. For those skilled in the art, according to the idea of the embodiment of the present disclosure, there will be changes in specific implementation manners and application ranges. In summary, the present disclosure should not be understood as a limitation on the embodiments of the present disclosure.

Claims

1. A data fusion method, the method comprising: inputting a relationship network graph of a target user into a fusion prediction model, the relationship network graph taking user entities as nodes and relationships between the user entities as edges connecting the nodes, the user entities being generated based on basic attributes and external information of the user; aggregating neighbor external information of the target user based on the fusion prediction model to obtain target external information of the target user, and obtaining an external evaluation corresponding to the target external information based on the fusion prediction model; obtaining an internal evaluation of the target user and inputting the internal evaluation into the fusion prediction model, the internal evaluation being obtained based on internal information of the target user; fusing the external evaluation and the internal evaluation based on the fusion prediction model to obtain an overall evaluation output by the fusion prediction model.

2. The method of claim 1, before the inputting a relationship network graph of a target user into a fusion prediction model, further comprising: obtaining user data of all users, the user data including basic attributes and external information; generating user entities corresponding to each user and extracting relationships between the user entities based on the user data of each user, and constructing a full-quantity relationship network graph corresponding to all users based on the user entities and the relationships between the user entities; obtaining a relationship network graph of a target user, the relationship network graph including a target user entity of the target user and at least one neighbor user entity having a relationship with the target user entity.

3. The method of claim 1, the obtaining an external evaluation corresponding to the target external information based on the fusion prediction model, comprising: updating the target user entity of the target user based on the target external information, and obtaining the external evaluation corresponding to the target external information based on the fusion prediction model and the target user entity.

4. The method of claim 1, the aggregating neighbor external information of the target user based on the fusion prediction model to obtain target external information of the target user, comprising: determining an aggregation weight corresponding to each neighbor user based on the relationships between the target user entity and each neighbor user entity in the relationship network graph, the aggregation weight being positively correlated with the closeness of the relationships; performing weighted average calculation on the external information of each neighbor user based on the aggregation weight of each neighbor user in the fusion prediction model to obtain the target external information of the target user.

5. The method of claim 4, the performing weighted average calculation on the external information of each neighbor user based on the aggregation weight of each neighbor user in the fusion prediction model to obtain the target external information of the target user, comprising: in the fusion prediction model, adding an attention weight to the external information of each neighbor user based on the acquisition time of the external information, the attention weight being greater for the external information acquired closer to the current time. performing a first weighted average calculation on the external information of each neighbor user based on an aggregated weight of each neighbor user to obtain first external information of the target user, and performing a second weighted average calculation on the external information based on an attention weight of each external information to obtain second external information of the target user; calculating target external information of the target user according to the first external information and the second external information.

6. The method of claim 1, wherein the internal evaluation is connected to the external evaluation in the fusion prediction model in a residual connection manner.

7. The method of claim 1, wherein the fusion of the external evaluation and the internal evaluation by the fusion prediction model to obtain the overall evaluation output by the fusion prediction model comprises: determining a credibility of the external evaluation according to a source of the external evaluation in the fusion prediction model, wherein the more real the source is, the higher the credibility is; assigning an attention weight to the external evaluation and the internal evaluation according to the credibility, wherein the higher the credibility is, the higher the attention weight of the external evaluation is; and performing a weighted fusion of the external evaluation and the internal evaluation based on the attention weight of the external evaluation and the attention weight of the internal evaluation to obtain the overall evaluation output by the fusion prediction model.

8. The method of claim 1, wherein the fusion prediction model is trained based on at least one sample data carrying a standard evaluation label. In the training process of the fusion prediction model, a predicted evaluation label is output for each sample data, a loss value of a loss function is calculated according to the predicted evaluation label and the standard evaluation label of each sample data, and a model parameter is adjusted based on the loss value until the fusion prediction model meets a preset convergence condition.

9. The method of claim 8, wherein the loss function comprises a performance loss function and a consistency loss function. The performance loss function is used to calculate a performance loss value between the predicted evaluation label and the standard evaluation label of each sample data, and the consistency loss function is used to calculate a consistency loss value between a pre-internal evaluation label and a predicted external evaluation label in the predicted evaluation label of each sample data.

10. The method of claim 9, wherein in the training process, a credibility of a sample external evaluation of a sample data is determined according to a source of the sample external evaluation, wherein the more real the source is, the higher the credibility is; a performance loss weight and a consistency loss weight corresponding to the sample data are determined according to the credibility, wherein the higher the credibility is, the higher the performance loss weight of the sample data is; and the loss value is a sum of loss values of all sample data, and the loss value of each sample data is a weighted sum of a performance loss value and a consistency loss value of each sample.

11. The method of claim 1, further comprising: ​ ​ ​ ​ ​ ​ performing user individual analysis and / or product recommendation on the target user based on the overall evaluation. 12.A data fusion device, comprising: a data input module configured to input a relationship network graph of a target user into a fusion prediction model, the relationship network graph having user entities as nodes and relationships between the user entities as edges connecting the nodes, the user entities being generated based on basic attributes and external information of the user; an external information aggregation module configured to aggregate neighbor external information of the target user based on the fusion prediction model to obtain target external information of the target user, and obtain an external evaluation corresponding to the target external information based on the fusion prediction model; an internal information access module configured to obtain an internal evaluation of the target user and access the internal evaluation into the fusion prediction model, the internal evaluation being obtained based on internal information of the target user; a prediction output module configured to fuse the external evaluation and the internal evaluation based on the fusion prediction model to obtain an overall evaluation output by the fusion prediction model. 13.A computer program product comprising instructions which, when executed on a computer or processor, cause the computer or processor to carry out the steps of the method of any one of claims 1-11. 14.A computer storage medium storing a plurality of instructions adapted to be loaded and executed by a processor to perform the steps of the method of any one of claims 1-11. 15.A terminal comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the method of any one of claims 1-11 when executing the computer program.

Images