Artificial intelligence-based recall method and device, electronic equipment and storage medium

Abstract

The disclosure provides an artificial intelligence-based recall method and device, electronic equipment and storage medium, relates to the technical field of data processing, and particularly relates to the field of artificial intelligence. The specific implementation scheme is: determining a first embedding value corresponding to a first user based on user features of the first user and resource features of information input by the first user, the first embedding value being used to represent the user features, the resource features and the degree of association between the user features and the resource features; determining a second user meeting a similar condition with the first user based on the first embedding value; and performing resource recall for the first user based on resource information corresponding to the second user.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and more particularly to artificial intelligence-based methods, apparatuses, electronic devices, and storage media. Background Technology

[0002] Artificial intelligence (AI) is a comprehensive technology in computer science that studies the design principles and implementation methods of various intelligent machines to enable them to have the functions of perception, reasoning, and decision-making.

[0003] Recall systems are an important application in the field of artificial intelligence. They can help users discover information that may be of interest to them in an information-overloaded environment, and then recall and push the information to users who are interested in it. Summary of the Invention

[0004] This disclosure provides a method, apparatus, electronic device, and storage medium for recall based on artificial intelligence.

[0005] According to a first aspect of this disclosure, an artificial intelligence-based recall method is provided, comprising: determining a first embedding value corresponding to the first user based on user characteristics of a first user and resource characteristics of information input by the first user, wherein the first embedding value is used to characterize the user characteristics, the resource characteristics, and the degree of association between the user characteristics and the resource characteristics;

[0006] Based on the first embedded value, a second user who meets similar conditions to the first user is identified;

[0007] Based on the resource information corresponding to the second user, resources are recalled for the first user.

[0008] According to a second aspect of this disclosure, an artificial intelligence-based recall device is provided, comprising:

[0009] The determining unit is configured to determine a first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user, wherein the first embedding value is used to characterize the user characteristics, the resource characteristics, and the degree of correlation between the user characteristics and the resource characteristics; and to determine a second user that meets similar conditions to the first user based on the first embedding value.

[0010] The recall unit is used to recall resources for the first user based on the resource information corresponding to the second user.

[0011] A third aspect of this disclosure provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the artificial intelligence-based recall method described above.

[0012] This disclosure provides a fourth aspect of a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to execute the artificial intelligence-based recall method described above.

[0013] The fifth aspect of this disclosure provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the artificial intelligence-based recall method described above.

[0014] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0015] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0016] Figure 1 This is a schematic diagram of an architecture of an artificial intelligence-based recall system provided in an embodiment of this disclosure;

[0017] Figure 2 This is another schematic diagram of the architecture of the AI-based recall system provided in this embodiment of the disclosure;

[0018] Figure 3 This is a schematic diagram of an optional process of the AI-based recall method provided in this embodiment of the disclosure;

[0019] Figure 4 This is a schematic diagram of another optional process for the AI-based recall method provided in this embodiment of the disclosure;

[0020] Figure 5 This is another schematic diagram of the architecture of the AI-based recall system provided in this disclosure embodiment;

[0021] Figure 6 This is a schematic diagram of an optional structure of the AI-based recall device provided in an embodiment of this application;

[0022] Figure 7 This is a block diagram of an electronic device used to implement the AI-based recall method of the embodiments of this disclosure. Detailed Implementation

[0023] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0024] In resource recommendation, the multi-matrix product ranking model (or multi-application resource retrieval model) is mainly implemented using the following three schemes:

[0025] 1) Based on the Content Base (CB) queue, use tags and first- and second-level categories for explicit resource retrieval;

[0026] 2) Based on the hot queue, inverted index resource recall is performed according to posterior information such as resource (item) exposure, clicks, and click-through rate (CTR);

[0027] 3) Based on the Graph Collaborative Filtering (GCF) queue, graph walk training is performed based on user clicks, comments and other behavioral information to confirm the similarity between the trained embedding and the trigger item embedding. Based on the similarity, the M most similar resources are recalled.

[0028] However, in resource recommendation, the recommendation platform empowers most matrix products, resulting in a relatively low number of Daily Active Users (DAU) for most products, with a large number of new and inactive users. New and inactive users refer to those whose behavioral information (such as logging into applications, clicking links within applications, searching through applications, watching videos, reading, etc.) is less than or equal to a second threshold. In this disclosure, new and inactive users can be collectively referred to as "cold start users." Compared to active users (those with behavioral information greater than the second threshold), cold start users have less behavioral information, resulting in a relatively sparse matrix of behavioral information.

[0029] For scheme 1), the CB queue is the most commonly used recall queue in related technologies, but its recall effect is poor in most matrix products, and it is often used as a fallback recall scheme. For scheme 2), the hot-priority queue is also the most commonly used recall queue, recalling high-priority and high-demand resources, but it lacks personalization and customization for a particular user. For scheme 3), the GCF queue uses existing user behavior for training, and its resource recall effect is good, and it also has personalization and customization. However, for most matrix products, cold start users account for a large proportion of the total users, that is, the matrix corresponding to the behavioral information of most users is relatively sparse, and the GCF queue can cover a small number of users. In addition, in related technologies, graph model-based recall methods can also be used to recall resources for users, but it requires rich user behavioral information to fully learn the vector representation of users and resources. Therefore, graph model-based recall methods cannot be used to recall resources for cold start users.

[0030] Therefore, in view of the defects in resource retrieval methods in related technologies, this disclosure provides an artificial intelligence-based retrieval method that can overcome some or all of the shortcomings of the prior art.

[0031] This disclosure provides an artificial intelligence-based recall method, apparatus, electronic device, and computer-readable storage medium, which can improve the accuracy of cold-start user resource recall. The following describes exemplary applications of the electronic device provided in this disclosure. The electronic device provided in this disclosure can be implemented as various types of terminal devices, as a server, or as a server and terminal device working together.

[0032] See Figure 1 , Figure 1 This is a schematic diagram of an artificial intelligence-based recall system 100 provided in this embodiment. To support an exemplary application, the terminal device 400 connects to the server 200 via the network 300, and the server 200 connects to the database 500. The network 300 can be a wide area network or a local area network, or a combination of both, and data transmission is achieved using wireless or wired links.

[0033] In some embodiments, taking an electronic device as a terminal device as an example, the AI-based recall method provided in this disclosure can be implemented by the terminal device. For example, the terminal device 400 runs a client 410, which can be a client for AI-based recall. The client can collect user characteristics of a first user and resource characteristics of information input by the first user, and transmit the user characteristics of the first user and resource characteristics of information input by the first user to the server 200 through the network 300.

[0034] In some embodiments, taking the electronic device as a server as an example, the AI-based recall method provided in this disclosure can be implemented collaboratively by the server and the terminal device. For example, the server 200 obtains the historical point log set and the user information set of historical users from the database 500. Then, the server 200 trains an embedding value model based on the obtained positive resource feature sample set, negative resource feature sample set, and user feature sample set.

[0035] When an AI-based recall is required, the client 410 obtains the user characteristics of the first user and the resource characteristics of the information input by the first user. The client 410 can confirm the characteristics of the first user based on the login information of the first user and / or the user information input by the first user. The client can also confirm the resource characteristics of the first user based on the information input by the first user (such as information entered in the search box of the application). Then, server 200 determines the type of the first user based on the first user's historical behavior information; if the first user is an active user, the user characteristics and / or resource characteristics of the information input by the first user are input into a hot queue or a CB queue for resource recall of active users; if the first user is a cold start user, the user characteristics and resource characteristics of the information input by the first user are input into a cold start queue, the cold start queue including an embedded value model (the embedded value model can exist in the form of a dictionary), the cold start queue inputs the user characteristics and resource characteristics of the information input by the first user into the embedded value model, and determines the first embedded value corresponding to the first user according to the output result of the embedded value model; the first embedded value is used to characterize the user characteristics, the resource characteristics, and the degree of correlation between the user characteristics and the resource characteristics; based on the first embedded value, a second user that meets similar conditions to the first user is determined; based on the resource information corresponding to the second user, resource recall is performed for the first user.

[0036] In some embodiments, the terminal device 400 or server 200 can implement the AI-based recall method provided in this disclosure by running a computer program. For example, the computer program can be a native program or software module in an operating system; it can be a native application (APP), i.e., a program that needs to be installed in the operating system to run; it can also be a mini-program, i.e., a program that only needs to be downloaded to a browser environment to run; or it can be a mini-program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module, or plugin.

[0037] In practical applications, server 200 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network or local area network to achieve data computation, storage, processing, and sharing. Terminal 400 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smart TV, smartwatch, etc., but is not limited to these. Terminal device 400 and server 200 can be directly or indirectly connected via wired or wireless communication, and this disclosure does not impose any restrictions.

[0038] Figure 2 This illustration shows another architecture diagram of the AI-based recall system provided in an embodiment of the present disclosure. To support an exemplary application, the terminal device 400 connects to the server 200 via the network 300, and the server 200 connects to the database 500. The network 300 can be a wide area network or a local area network, or a combination of both, and data transmission is achieved using wireless or wired links.

[0039] In some embodiments, taking the electronic device as a server as an example, the AI-based recall method provided in this disclosure can be implemented collaboratively by a terminal device and a server. For example, the terminal device 400 runs a client 410, which can be a client for AI-based recall. The client can collect user characteristics of a first user and resource characteristics of information input by the first user, and transmit a FEED refresh request carrying the user characteristics of the first user and the resource characteristics of information input by the first user to the resource aggregation module 201 included in the server 200 through the network 300; the resource aggregation module 201 sends the FEED refresh request to each downstream queue.

[0040] Each queue corresponds to a recall method, such as queue 1 corresponding to the hot-optimized recall method, queue 2 corresponding to the CB recall method, and queue k corresponding to the artificial intelligence-based recall method provided in this embodiment.

[0041] Each queue uses different strategies to recall resources based on the different attributes of the user characteristics and / or resource characteristics. For example, for cold-start users, resource recall can be performed using the AI-based recall method provided in this embodiment; it can also be performed using the AI-based recall method provided in this embodiment, as well as a recall method combining other recall methods; and the resources recalled by each queue are scored through stages such as coarse sorting, fine sorting, and resorting, and the resources and corresponding scores are returned to the resource aggregation module 201. For active users, recall can be performed through the queues corresponding to the popular recall method and the CB recall method respectively; and the resources recalled by each queue are scored through stages such as coarse sorting, fine sorting, and resorting, and the resources and corresponding scores are returned to the resource aggregation module 201. By using different recall methods for different users, the accuracy of resource recall for cold-start users is improved on the one hand, and the accuracy and personalization of resource recall for active users are also guaranteed on the other hand.

[0042] The resource aggregation module 201 integrates the resources of each queue and performs at least one of the following processes: deduplication, fusion, filtering, and sorting, to generate a final distribution list (resource recall list) and returns the list to the client 410, or returns the list to the client 410 through the server 200.

[0043] Based on the list, the client 410 presents the resources corresponding to the list to the first user through the display device corresponding to the client 410.

[0044] Among them, the queues included in the resource aggregation module 201 and the server 200, except for the queue corresponding to the AI-based recall method provided in this embodiment, are all designed and optimized for a single matrix product (application), requiring significant human and machine costs. The queues other than the queue corresponding to the AI-based recall method provided in this embodiment have strong generalization capabilities, can better combine the characteristics of different product lines (applications) to improve the relevance of resources recalled by cold-start users, and can be quickly applied to various matrix products, effectively saving machine and human resources.

[0045] Figure 3 This paper illustrates an optional flowchart of an AI-based recall method provided in an embodiment of this application, and will describe each step in detail.

[0046] Step S301: Determine the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user.

[0047] In some embodiments, an AI-based recall device (hereinafter referred to as the device) acquires user information of a first user; determines user characteristics of the first user based on the user information; receives information input by the first user; processes the information input by the first user to obtain resource characteristics; and determines a first embedding value corresponding to the first user based on the user characteristics and the resource characteristics. The first embedding value is used to characterize the user characteristics, the resource characteristics, and the degree of association between the user characteristics and the resource characteristics. The user information includes at least one of the user's identity document (ID), age, gender, income, primary category, secondary category, and interests; the user characteristic sample set includes at least one of the ID characteristics, age characteristics, gender characteristics, income characteristics, primary category characteristics, secondary category characteristics, and interests characteristics corresponding to the historical users. The interests and hobbies mentioned can be those entered by the user after the historical user launches the application, guided by the application. The ID features can be semantic and / or structural features, such as semantic features including combinations of letters and numbers that have special meaning to the user, strings corresponding to anime characters, etc.; structural features can include characters + numbers, AABB, sequential numbers or letters, etc.; the age feature can be an age range, such as 12-15 years old, 16-18 years old, 19-23 years old, etc.; the income feature can be an income range, such as monthly income of 0-1000 yuan, 1001 yuan-3000 yuan, etc. The specific age range or income range can be set to any value corresponding to the range based on actual needs, and this disclosure does not impose specific restrictions; the primary classification and the secondary classification can be the primary and secondary classifications stipulated by each field, for example, a major category of a certain field is determined as the primary classification, and the subcategories included in each major category are the secondary classifications; specifically, the primary classification (or primary discipline) of economics can include theoretical economics and applied economics; among which, theoretical economics includes the secondary classification of economics: political economy, history of economic thought, economic history, etc.; applied economics includes the secondary classification of economics: national economics, regional economics, and public finance, etc. The processing of the information input by the first user can include at least one of the following: word segmentation, removal of meaningless words (such as conjunctions, modal particles), and merging of adjacent and repeated words or phrases. For example, if the information input by the first user is "What should I eat for dinner during my weight loss period?", the input information can be processed based on a word segmentation model to obtain "during my weight loss period", "dinner", "what to eat", and "ya", and after removing the meaningless modal particle "ya", the resource features obtained are "during my weight loss period", "dinner", and "what to eat".

[0048] In some embodiments, the first embedding value is determined by an embedding value model. Before the device determines the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user, it may first train the embedding value model.

[0049] In specific implementation, the device obtains a positive resource feature sample set and a negative resource feature sample set based on the historical user's historical display log set; it obtains a user feature sample set based on the historical user's user information set; and it uses the features in the positive resource feature sample set, the negative resource feature sample set, and the user feature sample set as input to the embedding value model, and uses the embedding value as the output of the embedding value model to train the embedding value model. The historical user's historical display log includes logs showing resources to the historical user based on the historical user's input information, where the historical user clicked on the resource, or logs showing resources to the historical user based on the historical user's input information, where the historical user did not click on the resource. The positive resource feature sample set includes a set of logs showing resources to the historical user and the historical user clicking on the resources; the negative resource feature sample set includes a set of logs showing resources to the historical user and the historical user not clicking on the resources; the historical display logs of the historical user can be the historical display logs corresponding to all applications; optionally, due to the large amount of data in each application, for applications whose online time is less than a first time threshold, the historical display logs corresponding to the application can be traced back on a daily basis; for applications whose online time is greater than or equal to the first time threshold, the historical display logs corresponding to the application can be traced back on an hourly basis.

[0050] In some optional embodiments, the apparatus can further evaluate the embedding value model to obtain the model evaluation metric (Area Under the Curve, AUC) of the embedding value model; and adjust the parameters of the embedding value model based on the AUC of the embedding value model. Specifically, if the AUC of the embedding value model is less than a threshold, the parameters in the embedding value model are adjusted, and the model is retrained based on the features in the positive resource feature sample set, the negative resource feature sample set, and the user feature sample set. After training, the embedding value model is evaluated to obtain the AUC of the embedding value model, until the AUC of the embedding value model is greater than or equal to the threshold, confirming that the embedding value model training is complete.

[0051] The acquisition, storage, and application of user personal information (user information) involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0052] Step S302: Based on the first embedded value, determine a second user who meets similar conditions to the first user.

[0053] In some embodiments, the device may filter the historical click log set, including historical click logs of click operations and user information corresponding to the click operations; based on the historical click logs including click operations and user information corresponding to the click operations, determine an embedding value set; the embedding value set is used to determine a second user who meets similar conditions to the first user.

[0054] In some embodiments, the device determines a second embedding value in the embedding value set that satisfies the similarity condition based on the first embedding value; and determines the user corresponding to the second embedding value as the second user. The second user can be a single user, or multiple users determined based on similarity ranking. For example, the embedding values ​​in the embedding value set are sorted from high to low similarity, and the K users corresponding to the top K embedding values ​​are identified as the second user.

[0055] In specific implementation, the device can determine a second embedding value that satisfies the similarity condition by determining the similarity between the first embedding value and at least one embedding value included in the embedding value set. The algorithms for determining the similarity between the first embedding value and at least one embedding value included in the embedding value set may differ, and the similarity conditions may also differ. The embedding value can be represented by a vector (e.g., a vector with one row and N columns, where each embedding value has the same length, and N can be adjusted based on actual needs by setting parameters of the embedding value model). Therefore, determining the similarity between the first embedding value and at least one embedding value included in the embedding value set can be achieved through methods including, but not limited to, one of the following:

[0056] 1) The device uses Euclidean distance calculation to determine the similarity between the first embedding value and at least one embedding value included in the set of embedding values. The similarity condition may be that the Euclidean distance between the first embedding value and the first embedding value in space is the smallest. The device confirms that the embedding value in the set of embedding values ​​with the smallest Euclidean distance between the first embedding value and the first embedding value in space is the second embedding value.

[0057] 2) The device uses the Manhattan distance calculation method to determine the similarity between the first embedding value and at least one embedding value included in the set of embedding values. The similarity condition can be that the Manhattan distance with the first embedding value in space is the smallest. The device confirms that the embedding value in the set of embedding values ​​with the smallest Manhattan distance with the first embedding value in space is the second embedding value.

[0058] 3) The device determines the similarity between the first embedding value and at least one embedding value included in the set of embedding values ​​by calculating the Chebyshev distance. The similarity condition can be that the Chebyshev distance with the first embedding value in space is the smallest. The device confirms that the embedding value in the set of embedding values ​​with the smallest Chebyshev distance with the first embedding value in space is the second embedding value.

[0059] 4) The device uses the method of calculating the standardized Euclidean distance to determine the similarity between the first embedding value and at least one embedding value included in the set of embedding values. The similarity condition can be that the standardized Euclidean distance with the first embedding value in space is the smallest. The device confirms that the embedding value in the set of embedding values ​​with the smallest standardized Euclidean distance with the first embedding value in space is the second embedding value.

[0060] 5) The device determines the similarity between the first embedded value and at least one embedded value included in the set of embedded values ​​by calculating the cosine of the angle. The similarity condition can be that the cosine of the angle between the embedded value and the first embedded value is the largest (i.e., the angle between the embedded value and the first embedded value is the smallest, and the smallest angle is 0 degrees, at which point the vectors corresponding to the two embedded values ​​coincide in space); the device confirms that the embedded value with the largest cosine of the angle between the embedded value and the first embedded value in the set of embedded values ​​is the second embedded value;

[0061] 6) The device uses Hamming distance to determine the similarity between the first embedding value and at least one embedding value included in the set of embedding values. The similarity condition may be that the number of replacements is the fewest when converted to the first embedding value. The device confirms that the embedding value in the set of embedding values ​​that has the fewest replacements when converted to the first embedding value is the second embedding value.

[0062] 7) The device uses the Jaccard Similarity Coefficient to determine the similarity between the first embedded value and at least one embedded value included in the set of embedded values. The similarity condition may be that the proportion of the intersection element with the first embedded value in the union element with the first embedded value is the largest. The device confirms that the embedded value in the set of embedded values ​​with the largest proportion of the intersection element with the first embedded value in the union element with the first embedded value is the second embedded value.

[0063] Those skilled in the art will understand that the methods 1) to 7) described above for confirming similarity are merely examples, and other methods for confirming the similarity between two vectors can also be used to confirm the embedding values ​​that meet the similarity conditions from the set of embedding values. This disclosure does not make any specific limitations.

[0064] Step S303: Based on the resource information corresponding to the second user, perform resource recall for the first user.

[0065] In some embodiments, the device recalls resources for the first user based on resource information corresponding to the second user.

[0066] In some optional embodiments, the device recalls at least one resource information corresponding to each second user based on the historical click logs of the second user, including click operations. It then merges the at least one resource corresponding to all second users, records the total number of clicks for each resource, sorts the total number of clicks for all resources, and identifies the resource with the highest total number of clicks (a first threshold) from all the resource information of the second users. This resource with the highest total number of clicks (a first threshold) is then recalled for the first user. In this way, the resource with the most clicks can be preferentially recommended to the first user, allowing the first user to quickly obtain resources that may be of interest, thereby increasing the application's DAU (Daily Active Users).

[0067] Thus, by using the AI-based recall method provided in this disclosure, the first embedding value of the first user with a relevance is determined based on user characteristics and resource characteristics. This improves the accuracy of the parameters (first embedding value) used when subsequently determining the second user, thereby improving the accuracy of subsequent user determination, enhancing the relevance of the resources recalled by the user, and improving the user experience. For cold-start users with sparse user behavior, the second user with high similarity to the cold-start user can be accurately identified based solely on user characteristics and resource characteristics of user input information. This allows for accurate recall of resources that the cold-start user may be interested in, enabling the cold-start user to trigger more actions and become an active user, thus increasing the DAU of the matrix product. Furthermore, when training the embedding value model, historical point-and-click logs of all applications (or matrix products) are used as the resource feature sample set. This broadens the coverage of the model training samples and allows the trained embedding value model to be adapted to all applications. This gives the embedding value model provided in this disclosure strong generalization ability, enabling it to better combine the characteristics of different applications to improve the relevance of the resources recalled by the cold-start user. It can be quickly applied to various applications, saving machine and human resources.

[0068] Figure 4 This illustration shows another optional process diagram of the AI-based recall method provided in an embodiment of this disclosure. Figure 5This illustration shows another schematic diagram of an AI-based recall system provided in an embodiment of this disclosure, which will be combined with... Figure 4 and Figure 5 Please provide an explanation.

[0069] Step S401: Determine the training sample set.

[0070] In some embodiments, the training sample set includes a resource feature sample set and a user feature sample set; the resource feature sample set includes a positive resource feature sample set and a negative resource feature sample set; the AI-based recall device uses the historical display logs of all applications (or matrix products) as the resource feature sample set; the user feature sample set is obtained based on the user information set of all historical users of the applications; wherein, the historical display logs of historical users include logs of resources displayed to and clicked by the historical user based on the input information of the historical user, and logs of resources displayed to and not clicked by the historical user. Wherein, the log set of resources displayed to the historical user and clicked by the historical user is a positive resource feature sample set; the log set of resources displayed to the historical user and not clicked by the historical user is a negative resource feature sample set; wherein, the user information includes at least one of the user's ID, age, gender, income, primary category, secondary category, interests, etc.; the user feature sample set includes at least one of the historical user's ID feature, age feature, gender feature, income feature, primary category feature, secondary category feature, and interests feature; the resource features may include at least one of the resource's ID, primary category, secondary category, points of interest, and title word segmentation, etc. The interests and hobbies mentioned can be user input based on application guidance after the historical user launches the application. The ID features can be semantic and / or structural features of the ID. For example, semantic features can include combinations of letters and numbers that have special meaning to the user, strings corresponding to anime characters, etc., and structural features can include character + number, AABB, sequential numbers or letters, etc. The age feature can be an age range, such as 12-15 years old, 16-18 years old, 19-23 years old, etc. The income feature can be an income range, such as monthly income of 0-1000 yuan, 1001 yuan-3000 yuan, etc. Specific age or income ranges can be set to any value corresponding to a range based on actual needs, and this disclosure does not impose specific restrictions. The primary and secondary classifications can be the primary and secondary classifications stipulated for each field; for example, a major category of a certain field can be defined as the primary classification, and the subcategories included in each major category can be the secondary classifications. Specifically, the primary classification (or primary discipline) of economics can include theoretical economics and applied economics. Theoretical economics includes secondary classifications of economics such as political economy, history of economic thought, and economic history; applied economics includes secondary classifications of economics such as national economics, regional economics, and public finance.

[0071] In some optional embodiments, the training sample set can be updated based on time changes, and step S402 (i.e., training the embedding value model) is periodically executed based on the updated training sample set. The trained embedding value model is then periodically delivered online (or delivered to the server, or delivered to a queue included in the server). Since users' interests and content are constantly changing, if the embedding value model is trained only based on historical point data within a certain period, and there is no targeted training during model use, after a long period, the embedding value model cannot accurately confirm the user's embedding value, and therefore cannot accurately identify similar users. For example, at one time, "Want to go hiking?" might be a simple travel invitation, but at another time, it might mean a death invitation. By periodically updating the training set samples according to this disclosure, the embedding value model can continuously learn the correlation between user features and resource features, more accurately confirm the user's embedding value, and improve the accuracy of resource recall for cold-start users.

[0072] In some optional embodiments, for applications whose online time is less than a first time threshold, the historical display logs corresponding to the application can be backtracked on a daily basis; for applications whose online time is greater than or equal to the first time threshold, the historical display logs corresponding to the application can be backtracked on an hourly basis. On the one hand, this ensures that the number of historical display logs is appropriate, and on the other hand, it enables the embedding value model to continuously learn the correlation between user features and resource features, more accurately confirm the user's embedding value, and improve the accuracy of resource recall for cold-start users.

[0073] In some alternative embodiments, the device can also continuously train the embedding value model based on the recall resources and click resource information of cold-start users during the application of the embedding value model, so that the embedding value model can continuously learn the correlation between user features and resource features, more accurately confirm the user's embedding value, and improve the accuracy of resource recall for cold-start users.

[0074] The acquisition, storage, and application of user personal information (user information) involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0075] Step S402: Train the embedding value model.

[0076] In some embodiments, the apparatus uses features from the positive resource feature sample set, the negative resource feature sample set, and the user feature sample set as inputs to the embedding value model, and uses the embedding values ​​as outputs of the embedding value model to train the embedding value model.

[0077] In other embodiments, the apparatus may further evaluate the embedding value model to obtain its AUC; and adjust the parameters of the embedding value model based on the AUC. Specifically, if the AUC of the embedding value model is less than a threshold, the parameters in the embedding value model are adjusted, and the model is retrained based on features from the positive resource feature sample set, the negative resource feature sample set, and the user feature sample set. After training, the embedding value model is evaluated to obtain its AUC, until the AUC of the embedding value model is greater than or equal to the threshold, confirming that the embedding value model training is complete.

[0078] In specific implementation, the embedding value model can be an Alpha Factor Machine (alphaFM) model. The alphaFM model can be trained using a physical machine with user click behavior as the target. The input of the alphaFM model includes user features input from the user side and resource features input from the resource side. The user features, after being input into the alphaFM model, generate latent vectors corresponding to the user features; the resource features, after being input into the alphaFM model, generate latent vectors corresponding to the resource features. The latent vectors corresponding to the user features and the resource features have the same length and can both be vectors with one row and multiple columns. The latent vectors corresponding to the user features and the resource features are linearly added together to obtain the embedding value corresponding to the user.

[0079] During training, the device can input user-side features and resource-side features corresponding to a user into the alphaFM model. After the alphaFM model generates latent vectors corresponding to the user-side features and the resource-side features, the latent vectors corresponding to the user features and the resource features are linearly added together to obtain the embedding value corresponding to the user. Alternatively, the device can input both user features and resource features corresponding to a user into the alphaFM model. After the alphaFM model generates latent vectors corresponding to the user features and the resource features, the latent vectors corresponding to the user features and the resource features are linearly added together to obtain the embedding value corresponding to the user.

[0080] The formula corresponding to the alphaFM model can be expressed by the following equation:

[0081]

[0082]

[0083] Where y is the latent vector corresponding to the feature, and ω0 and ω i v represents the parameters of the alphaFM model. i Let x be the latent vector of the i-th feature, where <,> represents the vector dot product, and the length of the latent vector is k (k << n). i As a characteristic. x h x i and x i x j The coefficients are respectively <v h ,v i >and <v i ,v j > They have a common term v i .

[0084] Step S403: Distribute the embedded value model online; filter the embedded values ​​of some users and distribute them to the index server.

[0085] In some embodiments, the device delivers the trained embedded value model to an online platform (or to a server, or to a cold start recall queue included in the server). The online platform includes multiple recall queues, each corresponding to a recall method, such as a hot-optimized recall method for one recall queue and a cold start recall method for another recall queue.

[0086] Optionally, the device can deliver the embedded value model online in the form of a dictionary; wherein the dictionary includes at least latent vectors corresponding to user features and latent vectors corresponding to resource features. After receiving the user features and resource features of the first user, based on the latent vectors corresponding to user features and resource features included in the dictionary, all latent vectors corresponding to the first user are determined, and all latent vectors corresponding to the first user are linearly added together to obtain the first embedded value corresponding to the first user.

[0087] In other embodiments, the device determines the embedding value corresponding to the historical user based on the historical user's historical click log and the embedding value model; filters the embedding values ​​of users with clicks, and stores the embedding values ​​of the users with clicks and the historical click information index corresponding to the users with clicks in the index server (which may be the server corresponding to the Vec-Index service). The users with clicks include users with historical click behavior (e.g., if a resource is displayed to a historical user and the historical user clicks on the resource, then the historical user is a user with clicks and has historical click behavior). The index server is used to, in response to the embedding value of a cold-start user, determine users who meet similar conditions to the cold-start user, and the historical click information index corresponding to the users who meet the similar conditions.

[0088] Step S404: Obtain the historical click information of some users and write it to the Redis server.

[0089] In some embodiments, the device acquires the historical click information and / or the historical click information index of a user, and stores the historical click information and / or the historical click information index of the user in the Redis server.

[0090] In some alternative embodiments, the device responds to the embedding value of the cold-start user, determines from the index server a user that meets similar conditions to the cold-start user, and a historical click information index corresponding to the user; based on the historical click information index, it determines the user's historical click information from the Redis server, and recalls the resources corresponding to the historical click information to the cold-start user.

[0091] In some alternative embodiments, steps S401 to S404 can be performed offline, while steps S405 to S408 can be performed online.

[0092] Step S405: Confirm the user type of the first user.

[0093] In some embodiments, the device, in response to information input by the first user, determines the user type of the first user based on the amount of historical behavior information of the first user.

[0094] If the amount of historical behavior information of the first user is less than or equal to the second threshold, then the first user is a cold start user; if the amount of historical behavior information of the first user is greater than the second threshold, then the first user is an active user.

[0095] The historical behavior information includes at least one of the following: information about logging into (or opening) an application, information about clicking links in the application (such as the corresponding Uniform Resource Locator and Hypertext Markup Language, and the features corresponding to the Uniform Resource Locator and the Hypertext Markup Language), information about searching through the application (such as the features of the searched information, the number of searches, etc.), information about watching videos (such as the features of the videos, the viewing duration, etc.), and information about reading (such as the features of the text read, the reading duration, etc.).

[0096] If the first user has not been cold-started, then resources will be recalled for the first user based on the cold-start queue in the server, which may specifically include:

[0097] Step S406: Determine the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user.

[0098] In some embodiments, the device acquires user information of the first user; determines user characteristics of the first user based on the user information; processes the information input by the first user based on the information input by the first user to obtain resource characteristics; and inputs the user characteristics and resource characteristics of the first user into the embedding value model to obtain a first embedding value corresponding to the first user.

[0099] The processing of the information input by the first user may include at least one of the following: word segmentation, removal of meaningless words (such as conjunctions and modal particles), and merging of adjacent and repeated words or phrases. For example, if the information input by the first user is "What should I eat for dinner during my weight loss period?", the input information can be processed based on a word segmentation model to obtain "during weight loss period", "dinner", "what to eat", and "ya". After removing the meaningless modal particle "ya", the resulting resource features are "during weight loss period", "dinner", and "what to eat".

[0100] In specific implementation, such as Figure 5 As shown, after the user-side features (user features of the first user) are input into the alphaFM model, a latent vector corresponding to the user features is generated; after the resource features are input into the alphaFM model, a latent vector corresponding to the resource features is generated; wherein, the latent vector corresponding to the user features and the latent vector corresponding to the resource features have the same length, both being vectors with one row and multiple columns.

[0101] In some optional embodiments, the device may input the user features and resource features corresponding to the first user into the alphaFM model respectively. After the alphaFM model generates the latent vectors corresponding to the user features and the resource features respectively, they are linearly added together to determine the first embedding value corresponding to the first user. Alternatively, the device may input all the user features and resource features corresponding to the first user into the alphaFM model. After the alphaFM model generates the latent vectors corresponding to the user features and the resource features respectively, they are linearly added together to determine the first embedding value corresponding to the first user.

[0102] Step S407: Based on the first embedding value, determine a second user from the index server who meets similar conditions to the first user.

[0103] In some embodiments, the apparatus determines, based on the first embedding value, a second embedding value from the set of embedding values ​​that satisfies the similarity condition; and determines the user corresponding to the second embedding value as the second user.

[0104] In specific implementation, such as Figure 5 As shown, the device can determine a second embedding value that satisfies a similarity condition by determining the similarity between the first embedding value and at least one embedding value included in the embedding value set in the index server. The algorithms for determining the similarity between the first embedding value and at least one embedding value included in the embedding value set are different, and the similarity conditions are different; the embedding value can be represented by a vector (e.g., a vector with one row and N columns, where each embedding value has the same length, and N can be set based on actual needs through parameters of the embedding value model).

[0105] The second embedding value may include multiple embedding values.

[0106] Step S408: Obtain the resource information corresponding to the second user from the Redis server, and perform resource recall for the first user.

[0107] In some embodiments, in response to the first embedding value of the first user, the device determines, from the index server, a second user that meets similar conditions to the first user, and a historical click information index corresponding to the second user, based on the first embedding value of the first user; based on the historical click information index, it determines the historical click information of the second user from the Redis server, and recalls the resources corresponding to the historical click information to the first user, thereby realizing the resource recall of the first user.

[0108] Thus, the AI-based recall method provided in this disclosure, which determines the first embedding value of the first user based on user characteristics and resource characteristics, can improve the accuracy of the parameters (first embedding value) used when subsequently determining the second user, thereby improving the accuracy of subsequent user determination, enhancing the relevance of the resources recalled by the user, and improving the user experience. For cold-start users with sparse user behavior, the second user with high similarity to the cold-start user can be accurately identified based solely on user characteristics and resource characteristics of user input information, thereby accurately recalling resources that the cold-start user may be interested in, enabling the cold-start user to trigger more operations and become an active user, increasing the DAU of the matrix product. During the model training phase, the historical point display logs of all applications (or matrix products) are used as the resource feature sample set. On the one hand, this makes the coverage of the model training samples wider, and on the other hand, it allows the trained embedding value model to be adapted to all applications, giving the embedding value model provided in this disclosure strong generalization ability. It can better combine the characteristics of different applications to improve the relevance of the resources recalled by the cold-start user, and can be quickly applied to various applications, effectively saving machine and human resources.

[0109] Figure 6 A schematic diagram of an optional structure of an AI-based recall device provided in an embodiment of this application is shown, and the details will be explained based on each part.

[0110] In some embodiments, the AI-based recall device 600 includes a determination unit 601 and a recall unit 602.

[0111] The determining unit 601 is configured to determine a first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user, wherein the first embedding value is used to characterize the user characteristics, the resource characteristics, and the degree of correlation between the user characteristics and the resource characteristics; and to determine a second user that meets similar conditions to the first user based on the first embedding value.

[0112] The recall unit 602 is used to recall resources for the first user based on the resource information corresponding to the second user.

[0113] In some embodiments, the AI-based recall device 600 may further include an acquisition unit 603.

[0114] The acquisition unit 603 is used to acquire user information of the first user before determining the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics corresponding to the information input by the first user; and to determine the user characteristics of the first user based on the user information.

[0115] In some embodiments, the AI-based recall device 600 may further include a receiving unit 604.

[0116] The receiving unit 604 is configured to receive information input by the first user before determining the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user; and to process the information input by the first user to obtain the resource characteristics.

[0117] In some embodiments, the AI-based recall device 600 may further include a model training unit 605.

[0118] The model training unit 605 is used to obtain a positive resource feature sample set and a negative resource feature sample set based on the historical point log set of historical users before determining the first embedding value corresponding to the first user based on the user features of the first user and the resource features corresponding to the information input by the first user; obtain a user feature sample set based on the user information set of the historical users; and train the embedding value model by using the features in the positive resource feature sample set, the negative resource feature sample set and the user feature sample set as the input of the embedding value model and the embedding value as the output of the embedding value model.

[0119] The model training unit 605 is also used to evaluate the embedding value model to obtain the AUC of the embedding value model; and to adjust the parameters of the embedding value model based on the AUC of the embedding value model.

[0120] The determining unit 601 is further configured to filter the historical click-through log set, including historical click-through logs of click operations and user information corresponding to the click operations; and determine an embedding value set based on the historical click-through logs including click operations and user information corresponding to the click operations; the embedding value set is used to determine a second user who meets similar conditions to the first user.

[0121] The determining unit 601 is specifically used to determine, based on the first embedding value, a second embedding value in the set of embedding values ​​that satisfies the similarity condition; and to determine the user corresponding to the second embedding value as the second user.

[0122] The recall unit 602 is specifically used to sort the resource information corresponding to the second user based on the historical display log of the second user including click operations; determine at least one resource that meets the first condition; and recall the at least one resource that meets the first condition for the first user.

[0123] In some embodiments, the AI-based recall device 600 may further include a user attribute confirmation unit 606.

[0124] User attribute confirmation unit 606 is used to determine the first user as a cold start user based on the historical behavior information of the first user before determining the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user; the cold start user includes users whose historical behavior information is less than or equal to a second threshold.

[0125] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0126] Figure 7A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0127] like Figure 7 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. The RAM 803 may also store various programs and data required for the operation of the electronic device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0128] Multiple components in electronic device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of displays, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows electronic device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0129] The computing unit 801 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as AI-based recall methods. For example, in some embodiments, the AI-based recall method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the AI-based recall method described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform AI-based recall methods by any other suitable means (e.g., by means of firmware).

[0130] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0131] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0132] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0133] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0134] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0135] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0136] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0137] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An artificial intelligence-based recall method, comprising: Based on the first user's historical behavior information, the first user was determined to be a cold start user; The cold start users include users whose historical behavior information is less than or equal to the second threshold. A first embedding value is determined based on the user characteristics of the first user and the resource characteristics of the information input by the first user. The first embedding value is used to characterize the user characteristics, the resource characteristics, and the degree of correlation between the user characteristics and the resource characteristics. The first embedding value is determined based on an embedding value model. An embedding value model is trained using a set of positive and negative resource feature samples obtained from historical user display logs, a user feature sample set, and embedding values. The positive resource feature sample set includes logs of displayed and clicked resources, while the negative resource feature sample set includes logs of displayed and unclicked resources. The embedding value model is evaluated to obtain its evaluation metric. If the evaluation metric is less than a preset threshold, the parameters of the embedding value model are adjusted, and the model is retrained until the evaluation metric is greater than or equal to the preset threshold, confirming the completion of the embedding value model training. Based on the first embedded value, a second user who meets similar conditions to the first user is identified; Based on the resource information corresponding to the second user, resource recall is performed for the first user, including: recalling at least one resource information corresponding to each second user, wherein the at least one resource information is the clicked resource information; merging at least one resource corresponding to all second users, recording the total number of clicks for each resource, identifying the resource with the highest total number of clicks, and recalling the resource for the first user.

2. The method according to claim 1, wherein, Before determining the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics corresponding to the information input by the first user, the method further includes: Obtain the user information of the first user; The user characteristics of the first user are determined based on the user information.

3. The method according to claim 1, wherein, Before determining the first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user, the method further includes: Receive information input by the first user; The information input by the first user is processed to obtain the resource characteristics.

4. The method according to claim 1, wherein, The method further includes: The historical click-through log set is filtered to include historical click-through logs of click operations and user information corresponding to the click operations; Based on the historical display logs including the click operations and the user information corresponding to the click operations, determine the set of embedded values; The set of embedded values ​​is used to determine a second user who meets similar conditions to the first user.

5. The method according to claim 1 or 4, wherein, The step of determining a second user who meets similar conditions to the first user based on the first embedded value includes: Based on the first embedding value, determine a second embedding value in the set of embedding values ​​that satisfies the similarity condition; The user corresponding to the second embedded value is determined to be the second user.

6. An artificial intelligence-based recall device, comprising: The user attribute confirmation unit is used to determine that the first user is a cold start user based on the first user's historical behavior information; The cold start users include users whose historical behavior information is less than or equal to the second threshold. The determining unit is configured to determine a first embedding value corresponding to the first user based on the user characteristics of the first user and the resource characteristics of the information input by the first user. The first embedding value is used to characterize the user characteristics, the resource characteristics, and the degree of correlation between the user characteristics and the resource characteristics. Based on the first embedding value, a second user who meets similar conditions to the first user is identified; wherein, the first embedding value is determined based on an embedding value model; the embedding value model is trained based on a positive resource feature sample set and a negative resource feature sample set obtained from the historical display log set of historical users, as well as a user feature sample set and the embedding value; wherein, the positive resource feature sample set includes a log set of displayed resources and clicked resources, and the negative resource feature sample set includes a log set of displayed resources and unclicked resources; the embedding value model is evaluated to obtain a model evaluation index of the embedding value model; if the model evaluation index is less than a preset threshold, the parameters of the embedding value model are adjusted, and the embedding value model is retrained until the model evaluation index is greater than or equal to the preset threshold, confirming that the embedding value model training is complete; The recall unit is used to recall resources for the first user based on the resource information corresponding to the second user, including: recalling at least one resource information corresponding to each second user, wherein the at least one resource information is the clicked resource information; merging at least one resource corresponding to all second users, recording the total number of clicks for each resource, identifying the resource with the most total clicks, and recalling the resource for the first user.

7. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-5.

8. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-5.

9. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the method according to any one of claims 1-5.

Images