Data recommendation method, device, server and storage medium

Abstract

An embodiment of the present application discloses a data recommendation method, device, server and storage medium, wherein the method includes: in response to a data acquisition request sent by a target account, obtaining associated account information corresponding to an associated account having an associated relationship with the target account, the associated account information including a target associated relationship and a target portrait data; determining an associated embedded vector corresponding to the associated account based on the target associated relationship, and determining a portrait embedded vector corresponding to the associated account based on the target portrait data; fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filtering a similar account set from an account set based on the target embedded vector; obtaining multimedia data under each similar account in the similar account set, and recommending data to the target account based on the multimedia data, which can make full use of the behavioral data and the portrait data of the account, thereby helping to improve the efficiency of data recommendation.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data recommendation method, device, server, and storage medium. Background Art

[0002] With the rapid development of electronic and internet technologies, multimedia data is becoming increasingly abundant, and recommendation systems have also seen rapid development. Recommendation systems can extract user-required data or current social hot topics from this massive amount of multimedia data. Content producers of multimedia data can include original content accounts. As user demands and requirements grow, multimedia platforms increasingly expect high-quality and promising original content accounts. However, original content accounts require a cold start process, meaning that it takes time for an original content account to accumulate followers, or for the multimedia content it publishes to be read, played, and forwarded. Currently, recommendation systems typically make recommendations based on account popularity, prioritizing high-profile accounts while often ignoring low-profile or new accounts. Therefore, for self-media platforms, accelerating account cold starts and effectively integrating them with recommendation systems to improve data recommendation efficiency has become a research hotspot. Summary of the Invention

[0003] The embodiments of the present application provide a data recommendation method, device, server and storage medium. By implementing the above method, the behavioral data and portrait data of the account can be fully utilized, thereby improving the speed of cold start distribution of multimedia data corresponding to the account, improving data recommendation efficiency, and improving user experience.

[0004] A first aspect of an embodiment of the present application discloses a data recommendation method, the method comprising:

[0005] In response to a data acquisition request sent by a target account, acquiring associated account information corresponding to an associated account associated with the target account, the associated account information including a target association relationship and target profile data, the target association relationship including an association relationship between the associated account and each account in the account set, and the target profile data indicating a category feature of the associated account;

[0006] Determining the associated embedded vector corresponding to the associated account based on the target association relationship, and determining the portrait embedded vector corresponding to the associated account based on the target portrait data;

[0007] fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filtering the account set to obtain a similar account set based on the target embedded vector;

[0008] Multimedia data under each similar account in the similar account set is obtained, and data is recommended to the target account based on the multimedia data.

[0009] A second aspect of an embodiment of the present application discloses a data recommendation device, comprising:

[0010] an acquisition unit, configured to, in response to a data acquisition request sent by a target account, acquire associated account information corresponding to an associated account associated with the target account, the associated account information including a target association relationship and target profile data, the target association relationship including an association relationship between the associated account and each account in the account set, the target profile data being used to indicate a category feature of the associated account;

[0011] a first determining unit, configured to determine an associated embedded vector corresponding to the associated account based on the target association relationship, and determine a portrait embedded vector corresponding to the associated account based on the target portrait data;

[0012] a second determining unit, configured to fuse the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filter the account set to obtain a similar account set based on the target embedded vector;

[0013] The recommendation unit is configured to obtain multimedia data under each similar account in the similar account set, and recommend data to the target account based on the multimedia data.

[0014] The third aspect of an embodiment of the present application discloses a server, including a processor, a memory and a network interface, wherein the processor, the memory and the network interface are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method of the above-mentioned first aspect.

[0015] A fourth aspect of an embodiment of the present application discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the method of the first aspect.

[0016] A fifth aspect of the present application discloses a computer program product or computer program, comprising computer instructions stored in a computer-readable storage medium. A processor of a server reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the server to perform the method of the first aspect.

[0017] In an embodiment of the present application, the server can respond to a data acquisition request sent by a target account and obtain associated account information corresponding to an associated account having an associated relationship with the target account. The associated account information includes a target associated relationship and a target portrait data. The target associated relationship includes an associated relationship between the associated account and each account in the account set. The target portrait data is used to indicate the category characteristics of the associated account. Then, the associated embedded vector corresponding to the associated account is determined based on the target associated relationship, and the portrait embedded vector corresponding to the associated account is determined based on the target portrait data. Then, the associated embedded vector and the portrait embedded vector are fused to obtain a target embedded vector, and a similar account set is obtained from the account set based on the target embedded vector. Further, multimedia data under each similar account in the similar account set is obtained, and data is recommended to the target account based on the multimedia data. By implementing the above method, the behavioral data between the user and the account and the portrait data of the account can be fully utilized, and the embedded vector of the account can be jointly determined based on the behavioral data and the portrait data to recall similar accounts of the account based on the embedded vector, thereby improving the speed of cold start distribution of multimedia data corresponding to the account, improving data recommendation efficiency, and improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0019] Figure 1a This is a schematic diagram of the architecture of a data recommendation system provided in an embodiment of the present application;

[0020] Figure 1b This is a schematic diagram of the architecture of a data recommendation system provided in an embodiment of the present application;

[0021] Figure 2 This is a flow chart of a data recommendation method provided in an embodiment of the present application;

[0022] Figure 3a This is a schematic diagram of a process for determining an associated embedded vector provided by an embodiment of the present application;

[0023] Figure 3b This is a schematic diagram of the structure of a similar account set provided in an embodiment of the present application;

[0024] Figure 3c This is a schematic diagram of the structure of a similar account set provided in an embodiment of the present application;

[0025] Figure 3d is a schematic diagram of a process for determining a target embedded vector provided by an embodiment of the present application;

[0026] Figure 4 This is a flow chart of a data recommendation method provided in an embodiment of the present application;

[0027] Figure 5a This is a schematic diagram of a process for determining a first vectorized model provided in an embodiment of the present application;

[0028] Figure 5b This is a schematic diagram of the structure of a weighted reference account relationship diagram provided in an embodiment of the present application;

[0029] Figure 5c This is a schematic diagram of a process for determining an associated embedded vector provided by an embodiment of the present application;

[0030] Figure 6 This is a structural diagram of a data recommendation device provided in an embodiment of the present application;

[0031] Figure 7 This is a structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0033] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0034] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0035] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0036] The solutions provided in the embodiments of this application involve technologies such as machine learning based on artificial intelligence, and are specifically described through the following embodiments:

[0037] Currently, multimedia platforms typically accelerate the cold start of potential accounts and integrate them with recommendation systems by having platform staff pre-configure a whitelist and then allocate a fixed amount of traffic to potential accounts on the whitelist for cold start attempts, or by using profile data to recall potential accounts. However, potential accounts often publish relatively little multimedia data, resulting in less corresponding profile data. Furthermore, due to the inability to quickly identify accurate potential accounts, traffic utilization efficiency is low on multimedia platforms. Due to the head effect, a large amount of traffic is diverted to high-traffic accounts, reducing the distribution of high-quality content and user perception of potential accounts. Consequently, the consumption behavior data between users and accounts, as well as the profile data of the accounts, are not fully utilized. Furthermore, the convergence rate of cold start rates for potential accounts is slow. If content exploration is used to continuously collect profile data for accounts and then use this data to recall similar accounts, data recommendation efficiency may be low, potentially leading to the loss of potential accounts and a decrease in their retention rate.

[0038] In response to the above problems, an embodiment of the present application provides a method for data recommendation. The server can respond to a data acquisition request sent by a target account and obtain associated account information corresponding to an associated account that has an associated relationship with the target account. The associated account information includes a target associated relationship and a target portrait data. The target associated relationship includes the associated relationship between the associated account and each account in the account set. The target portrait data is used to indicate the category characteristics of the associated account. Then, the associated embedded vector corresponding to the associated account is determined based on the target associated relationship, and the portrait embedded vector corresponding to the associated account is determined based on the target portrait data. Then, the associated embedded vector and the portrait embedded vector are fused to obtain a target embedded vector, and a similar account set is obtained from the account set based on the target embedded vector. Further, multimedia data under each similar account in the similar account set is obtained, and data is recommended to the target account based on the multimedia data. By implementing the above method, the behavioral data between the user and the account and the portrait data of the account can be fully utilized, and the embedded vector of the account can be jointly determined based on the behavioral data and the portrait data to recall similar accounts of the account based on the embedded vector, thereby improving the speed of cold start distribution of multimedia data corresponding to the account, improving data recommendation efficiency, and improving user experience.

[0039] It should be noted that the data recommendation method provided by this solution can be specifically applied to data recommendation systems. Figure 1a , Figure 1a 101 and a server 102 are provided. The data recommendation method can be specifically executed by the server 102.

[0040] Taking terminal 101 as an example, terminal 101 receives a data acquisition request initiated by a target account through terminal 101 and sends the data acquisition request to server 102. Server 102 obtains associated account information corresponding to associated accounts associated with the target account in response to the data acquisition request. This associated account information may include the target association relationship and target profile data. The target association relationship may include the association relationship between the associated account and each account in the account set. The target profile data may be used to indicate the category characteristics of the associated account. Server 102 determines an associated embedded vector corresponding to the associated account based on the target association relationship and a profile embedded vector corresponding to the associated account based on the target profile data. Server 102 then fuses the associated embedded vector and the profile embedded vector to obtain a target embedded vector. Based on the target embedded vector, server 102 filters the account set to obtain a set of similar accounts. Server 102 obtains multimedia data from each similar account in the set of similar accounts and recommends data to the target account based on the multimedia data. Subsequently, server 102 transmits the determined multimedia data to terminal 101, which displays the multimedia data so that the target account obtains the required multimedia data.

[0041] Figure 1a The terminal 101 shown can be a smart device such as a mobile phone, tablet computer, laptop computer, PDA, mobile Internet device (MID), wearable device, etc. The terminal 101 and the server 102 can be directly or indirectly connected via wired or wireless communication, which is not limited in this application.

[0042] In one implementation scenario, this solution also provides another data recommendation system that applies this data recommendation method, such as Figure 1b This is a schematic diagram of the architecture of a data recommendation system provided by an embodiment of the present application. In the data recommendation system described in the embodiment of the present application, the data recommendation method can be specifically performed by a recommendation module, wherein the recommendation module can include an account embedded vectorization module, a similar account recall module, a recommendation recall module, and a recommendation ranking module.

[0043] In one implementation, the target embedded vector of the associated account can be determined by the account embedded vectorization module. Specifically, the account embedded vectorization module can obtain the associated account information corresponding to the associated account that has an associated relationship with the target account. The associated account information includes the target associated relationship and the target portrait data. The target associated relationship includes the associated relationship between the associated account and each account in the account set. The target portrait data is used to indicate the category characteristics of the associated account. Then, the associated embedded vector corresponding to the associated account is determined based on the target associated relationship, and the portrait embedded vector corresponding to the associated account is determined based on the target portrait data. Further, the associated embedded vector and the portrait embedded vector are fused to obtain the target embedded vector. After the account embedded vectorization module determines the target embedded vector, the target embedded vector can be uploaded to the similar account recall module so that the similar account recall module recalls similar accounts of the associated account. The specific implementation method of the recall can be to calculate the similarity between the target embedded vector of the associated account and the embedded vector of each account in the account set, so as to determine the similar account set of the associated account based on the similarity. Then, after the similar account recall module determines the similar account set, the similar account set can be uploaded to the recommendation recall module. After receiving the similar account set, the recommendation recall module can obtain the multimedia data under each similar account in the similar account set from the content database, and send the multimedia data to the recommendation sorting module, and the recommendation sorting module can sort at least one multimedia file included in the multimedia data to recommend data according to the sorting result. Specifically, the recommendation sorting module receives the multimedia data sent by the recommendation recall module, and sorts the multimedia data to display the multimedia data to the user according to the sorting result. Among them, the sorting method of the multimedia data is not limited in this application, and the multimedia data can be sorted according to any sorting rules. For example, the multimedia data can be sorted according to the historical number of clicks on the multimedia data, and the multimedia data can be output according to the sorting result.

[0044] In one implementation, the data recommendation system may further include a content production terminal, a content consumption terminal, an upstream and downstream content interface module, a content distribution export module, a content database, a scheduling center module, an audit module, a duplicate elimination module, a statistical reporting interface module, and a statistical database.

[0045] The content production end is used to provide the multimedia data required by the multimedia platform. The multimedia data can be, for example, graphics, text, or videos, and videos can include short videos and mini videos. Content producers (which can be understood as accounts) can upload multimedia data, where content producers can include professional generated content (PGC) or user generated content (UGC) or multi-channel network (MCN) or professional user generated content (PUGC). The content production end can obtain the upload module interface address through the communication connection with the upstream and downstream content interface modules, and then upload the multimedia data to be published through the interface address.

[0046] The content consumption terminal is used to communicate with the upstream and downstream content interface modules to obtain index information for accessing multimedia data, and then communicate with the upstream and downstream content interface modules and the content export module to obtain multimedia data. It can also report various user consumption behaviors of multimedia data to the statistics reporting interface module. For example, if the multimedia data is a video, consumption behaviors may include video playback duration, cache time, and various user interactions with the video content, such as comments, forwarding, sharing, favorites, likes, and reports.

[0047] The uplink and downlink content interface module is used to communicate directly with the content production end and can obtain multimedia data provided by the content production end. This information can be related to the multimedia data, which can be referred to as meta-information. For example, this meta-information can include the size of the multimedia data, a link to a cover image, a title, the release time, the publisher, the source channel, and so on. It can also include the classification of the multimedia data, such as the level classification and tag information of the multimedia data. For example, if the multimedia data is an article about watches, the first-level classification of the article may be technology, the second-level classification may be smart watches, and the third-level classification may be domestic watches. The tag information may specifically include the brand and model of the watch. The multimedia data and its meta-information can be stored in the content database. The module can also report the posting flow information of each account to the statistical reporting interface module. The posting flow information can include the posting time and content type of the multimedia data published by each account. The module can also send the multimedia data uploaded by the content production end to the scheduling module center, so that the scheduling module center can schedule according to the received multimedia data.

[0048] The content distribution export module is used to communicate with the recommendation and sorting module, obtain the multimedia data provided by the recommendation and sorting module, send the multimedia data to the content consumption end, and display the multimedia data in the message source list of the user terminal.

[0049] The content database is used to receive and store multimedia data and multimedia data metadata uploaded by the upstream and downstream content interface modules. For example, the target profile data of the associated account required by the account embedded vectorization module to determine the target embedded vector for the associated account can be obtained from the content database. This target profile data can specifically be obtained from the multimedia data metadata stored in the content database. The content database is also used to store the audit results provided by the audit module and the duplicate removal results provided by the duplicate removal module.

[0050] The dispatch center module is responsible for the entire scheduling process of multimedia data flow within the data recommendation system. It receives multimedia data to be stored in the content database through the upstream and downstream content interface modules and sends the received multimedia data and its metadata to the content database, enabling real-time updates of the content database. It also schedules the duplicate removal module to mark and filter duplicate multimedia data stored in the content database. It also schedules the audit module to audit multimedia data, addressing issues such as data security.

[0051] The audit module is used to read multimedia data from the content database through the scheduling module center and audit the multimedia data to ensure that the multimedia data pushed by the data recommendation system complies with specified standards, such as laws or policies, so as to filter out multimedia data that does not meet the specified standards. The audit module's audit results can also be written to the content database through the scheduling module center.

[0052] The deduplication module is used to filter multimedia data by deduplicating duplicate multimedia data stored in the content database through the scheduling module center. This deduplication process can include deduplicating multimedia data titles, images of multimedia cover images, content content, and video and audio fingerprints. The deduplication module typically vectorizes the titles and text of the graphic content.

[0053] The statistical reporting interface module is used to receive various consumption behavior data of users on multimedia data reported by the content consumption end, and is also used to receive the posting flow information reported by the uplink and downlink content interface modules, and is also used to write the reported data into the statistical database.

[0054] The statistical database is used to receive statistical data reports from content consumers, providing data support for subsequent statistical analysis and mining. It establishes communication with the account embedded vectorization module and provides the account embedded vectorization module with the data required to determine the account's embedded vector. It is also used to receive reports of posting information from content producers to collect information such as category labels of multimedia data published by accounts. For example, the associated account information required by the account embedded vectorization module to determine the target embedded vector for the associated account can be obtained from the statistical database.

[0055] See also Figure 2 , Figure 2 This is a flow chart of a data recommendation method provided by an embodiment of the present application. The method is applied to and can be executed by a server. As shown in the figure, the data recommendation method may include:

[0056] S201: In response to a data acquisition request sent by a target account, obtain associated account information corresponding to an associated account associated with the target account.

[0057] In one implementation, a server may receive a data acquisition request sent by a target account and obtain associated account information corresponding to an associated account associated with the target account. The target account may be understood as a user currently reading on a multimedia platform, such as a reading application. The associated account may then be the account corresponding to the multimedia data currently being read by the user. The associated account information may include a target association relationship and target profile data. The target association relationship may include the association relationship between the associated account and each account in an account set. For example, if the associated account is account A, and the account set includes account B, account C, and account D, the association relationship between account A and account B may be understood as the target account reading multimedia data posted by account A, followed by the target account reading multimedia data posted by account A. The target profile data may be used to indicate the category characteristics of the associated account. The category characteristics may include the category characteristics of the target multimedia data under the associated account, the target group characteristics of the associated account, and so on.

[0058] S202: Determine the associated embedded vector corresponding to the associated account based on the target association relationship, and determine the portrait embedded vector corresponding to the associated account based on the target portrait data.

[0059] In one implementation, the specific implementation method of determining the associated embedded vector corresponding to the associated account based on the target association relationship can be as follows: Figure 3a The process diagram shown is as follows: Figure 3aAs shown, a target account relationship graph can be constructed based on the target association relationship, where the target account relationship graph includes at least one account node. After constructing the target account relationship graph, a node random walk can be performed in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph. Then, the first vectorized model is called to process the target random walk sequence to obtain the associated embedded vector corresponding to the associated account.

[0060] In one implementation, the target portrait data may include multiple category labels under the associated account. For example, the category label may include the category label of the target multimedia data, the category label of the associated account, and the target group label of the associated account, etc. Then, a specific implementation method of determining the portrait embedded vector corresponding to the associated account based on the target portrait data may be to call the second vectorization model to process the multiple category labels under the associated account respectively to obtain the portrait embedded vector corresponding to each category label. Specifically, assuming that the category label includes the category label of the target multimedia data, the category label of the associated account, and the target group label of the associated account, the second vectorization model may be called to process the category label of the target multimedia data under the associated account to obtain a first portrait embedded vector, the second vectorization model may be called to process the category label of the associated account to obtain a second portrait embedded vector, and the second vectorization model may be called to process the target group label of the associated account to obtain a third portrait embedded vector. Then, after obtaining the first portrait embedded vector, the second portrait embedded vector, and the third portrait embedded vector, the portrait embedded vector corresponding to the associated account may be constructed based on the first portrait embedded vector, the second portrait embedded vector, and the third portrait embedded vector. Among them, the above-mentioned construction of the portrait embedded vector corresponding to the associated account can be understood as the portrait embedded vector corresponding to the associated account includes the first portrait embedded vector, the second portrait embedded vector and the third portrait embedded vector.

[0061] Optionally, the first vectorization model may be a related word vector model, such as a Skip-gram model, and the second vectorization model may also be a related word vector model, wherein the first vectorization model and the second vector model may be the same or different.

[0062] S203: Fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filtering the account set to obtain a similar account set based on the target embedded vector.

[0063] In one implementation, the server may fuse the associated embedded vector and the portrait embedded vector to obtain a target embedded vector corresponding to the associated account. After obtaining the target embedded vector, the server may filter out a similar account set corresponding to the associated account from the account set based on the target embedded vector.

[0064] For example, if Figure 3b and Figure 3c The following is a schematic diagram of the structure of a similar account set corresponding to an associated account provided by an embodiment of the present application. Figure 3b The accounts shown are some of the similar accounts in the similar account set corresponding to Account A (Teacher Tang teaches mathematics). It can be seen that the similar accounts are very similar to Account A in type and style, and the similar accounts are all about mathematics teaching. For example, the similar account corresponding to Account A can be Account A1 (Teacher Luo's mathematics class). Figure 3c The accounts shown are some of the similar accounts in the similar account set corresponding to account B (XXX food). It can be seen that the similar accounts are very similar to account B in type and style, and the similar accounts are all about food. It can also be seen that among similar accounts, some similar accounts have relatively few followers, that is, through the method of the embodiment of the present application, similar accounts with less followers can be recalled. In this case, the follow-up amount of high-heat accounts with more followers can be used as the target follow-up amount of similar potential accounts recalled by the target embedded vector corresponding to the account, thereby realizing the migration of follow-up amount and the cold start of potential accounts, which can effectively reduce the traffic required to test new accounts and speed up the convergence speed. It can be seen that the distribution of multimedia data corresponding to the account is better promoted through the user's consumption behavior data on the account and the portrait data of the account itself. Potential accounts can also get priority in obtaining directional traffic, and the utilization efficiency of traffic can be improved.

[0065] In one implementation, when the associated embedded vector and the portrait embedded vector are fused to obtain the target embedded vector, the fusion method can be weighted summation. Then, a specific implementation method for obtaining the target embedded vector can be to first obtain the associated weight corresponding to the associated embedded vector and the portrait weight corresponding to the portrait embedded vector. The associated weight and the portrait weight respectively represent the importance of the associated embedded vector and the portrait embedded vector when constructing the target embedded vector, and the associated weight and the portrait weight can be obtained through continuous learning of the model. After determining the associated weight and the portrait weight, the associated embedded vector can be weighted based on the associated weight to obtain a weighted associated embedded vector corresponding to the associated embedded vector. Similarly, the portrait embedded vector can be weighted based on the portrait weight to obtain a weighted portrait embedded vector corresponding to the portrait embedded vector. Then, after determining the weighted associated embedded vector and the weighted portrait embedded vector, the weighted associated embedded vector and the weighted portrait embedded vector can be spliced. Splicing can be understood as the summation of vectors, that is, the weighted associated embedded vector and the weighted portrait embedded vector are summed, and the summation result is the target embedded vector.

[0066] For example, if Figure 3d FIG. 1 is a flow chart of determining a target embedded vector according to an embodiment of the present application. Figure 3d As shown, sparse features can be understood as including the above-mentioned target association relationships and target portrait data, and embedded vectors can be understood as including associated embedded vectors determined based on the target association relationships, and portrait embedded vectors determined based on the target portrait data. Each embedded vector is multiplied by the corresponding weight, and the target embedded vector is obtained by summing up the multiplication results. Figure 3d a0, a1…an in are the weights corresponding to each embedded vector.

[0067] In one implementation, determining a similar account set can be based on the similarity between a target embedded vector and the embedded vectors corresponding to each account in the account set. Specifically, an embedded vector set corresponding to the account set can be obtained, the embedded vector set including the embedded vector corresponding to each account in the account set. After obtaining the embedded vector set, the similarity between the target embedded vector and each embedded vector in the embedded vector set can be determined, where the similarity can be obtained by calculating the cosine distance between the two vectors. After determining the similarity between the target embedded vector and each embedded vector, each similarity is compared with a preset similarity, and similarities with the target embedded vector that are greater than the preset similarity are selected. Vectors with similarities with the target embedded vector that are greater than the preset similarity can be referred to as similar embedded vectors. Through this selection process, at least one similar embedded vector can be obtained. After obtaining at least one similar embedded vector, the accounts corresponding to each similar embedded vector in the at least one similar embedded vector can be obtained. The accounts corresponding to the similar embedded vectors can be referred to as similar accounts, i.e., at least one similar account can be obtained. A similar account set can then be constructed based on the at least one similar account.

[0068] Optionally, when constructing a similar account set based on the at least one similar account, the at least one similar account can be added to the similar account set to obtain a similar account set. The at least one similar account can also be further filtered, and the similar accounts obtained after the filtering are added to the similar account set to obtain a similar account set. In the above filtering of similar accounts, the filtering can be based on the consumption behavior of the target account on the similar accounts. For example, if the target account's consumption data volume for a similar account is higher than a preset consumption data volume, the similar account is retained. If the target account's consumption data volume for a similar account is lower than a preset consumption data volume, the similar account is deleted. Then, according to the above filtering process, the retained similar accounts can be added to the similar account set to obtain a similar account set. Among them, the above-mentioned consumption data volume may refer to the data volume of the target account's consumption behavior on the account. The consumption behavior may refer to behaviors such as browsing, playing, commenting, collecting, forwarding, sharing, liking, etc. on the account. If the account is a video account, the consumption behavior may also include the average playback completion rate of the video content in the video account. If the account is a picture and text account, the consumption behavior may also include the average reading completion rate of the picture and text content in the picture and text account.

[0069] S204: Acquire multimedia data under each similar account in the similar account set, and recommend data to the target account based on the multimedia data.

[0070] In one implementation, multimedia data under each similar account in the similar account set can be obtained from the content database, wherein the multimedia data includes at least one multimedia file, wherein each multimedia file can be obtained by each similar account through Figure 1b The uplink and downlink content interface services shown are uploaded, and obtained after review and deduplication processing. After obtaining the multimedia data under each similar account, the multimedia files in the multimedia data can be sorted according to the preset sorting rules to obtain the sorting results corresponding to each multimedia file, and data recommendations can be made to the target account based on the sorting results. Among them, the preset sorting rules can be to sort each multimedia file according to the historical number of clicks corresponding to each multimedia file. Specifically, the historical number of clicks corresponding to each multimedia file can be obtained, and the historical number of clicks corresponding to each multimedia file can be sorted in descending order to obtain the sorting results. The multimedia files in the top N positions in the sorting results are used as the data to be recommended, and the multimedia files in the top N positions are recommended to the target account.

[0071] In one implementation, multimedia data from each similar account in a similar account set may be obtained to recommend data to a target account based on the obtained multimedia data. Optionally, when recommending data to a target account based on the obtained multimedia data, all of the obtained multimedia data may be recommended to the target account. Furthermore, considering that the amount of obtained multimedia data may be relatively large, the obtained multimedia data may be further filtered to recommend data to the target account based on the filtered multimedia data.

[0072] Optionally, multimedia data can be deleted based on its time, for example, the time can be its release time. Specifically, the release time of each multimedia data can be obtained, and then the release time of each multimedia data can be compared with a preset time condition. When the release time of the multimedia data meets the preset time condition, the multimedia data is retained. Correspondingly, when the release time of the multimedia data does not meet the preset time condition, the multimedia data is deleted. The preset time condition can be set in advance, and the preset time condition can specifically be that the difference between the release time of the multimedia and the time of the data acquisition request of the current target user does not exceed a preset time. The preset time can be set shorter, for example, the preset time can be 2 hours, or 1 hour, etc. It can be seen that by setting the above preset time condition, it can be ensured as much as possible that the multimedia data obtained by the target user is the latest data, more in line with the user's intention, rather than relatively old data, thereby improving the efficiency of data recommendation and improving the user experience.

[0073] Optionally, while performing the aforementioned filtering based on the time of multimedia data, a number of recommended data may be set to further filter the multimedia data based on the number of recommended data, thereby reducing data redundancy and improving data recommendation efficiency. Specifically, after obtaining the filtered multimedia data according to the time of multimedia data, the number of filtered multimedia data may be determined. If the number is greater than a preset number, some of the multimedia data may be deleted to reduce the number of multimedia data to less than or equal to the preset number.

[0074] Optionally, the multimedia data can be deleted based on its popularity, for example, the popularity can refer to the number of clicks. Specifically, the number of clicks on each multimedia data can be obtained first, wherein the number of clicks on each multimedia data within a preset time period can be obtained. The preset time period can be pre-set, for example, the preset time period can be 24 hours, or 12 hours before the data acquisition request time of the current target account, etc. After determining the number of clicks on each multimedia data, the multimedia data can be sorted in descending order of the number of clicks to obtain a sorting result. The first K multimedia data in the sorting result can be retained, and the remaining multimedia data can be deleted. The K value can be pre-set, for example, it can be 10 or 15, etc.

[0075] It should be noted that, in the embodiment of the present application, when further filtering the acquired multimedia data to recommend data to the target account based on the filtered multimedia data, the acquired multimedia data may be further filtered in the manner described above, or the acquired multimedia data may be further filtered based on other methods, which is not limited in the present application.

[0076] In an embodiment of the present application, in response to a data acquisition request sent by a target account, associated account information corresponding to an associated account having an associated relationship with the target account can be obtained, the associated account information including a target associated relationship and a target portrait data, the target associated relationship including an associated relationship between the associated account and each account in the account set, the target portrait data being used to indicate the category characteristics of the associated account, then, based on the target associated relationship, an associated embedded vector corresponding to the associated account is determined, and based on the target portrait data, a portrait embedded vector corresponding to the associated account is determined, then, the associated embedded vector and the portrait embedded vector are fused to obtain a target embedded vector, and based on the target embedded vector, a similar account set is obtained from the account set, further, multimedia data under each similar account in the similar account set is obtained, and data recommendation is performed for the target account based on the multimedia data. By implementing the above method, the behavioral data between the user and the account and the portrait data of the account can be fully utilized, and the embedded vector of the account can be jointly determined based on the behavioral data and the portrait data, so as to recall similar accounts of the account based on the embedded vector, thereby improving the speed of cold start of multimedia data distribution corresponding to the account, improving data recommendation efficiency, and improving user experience.

[0077] See also Figure 4 , Figure 4 This is a flow chart of a data recommendation method provided by an embodiment of the present application. The method is applied to and can be executed by a server. As shown in the figure, the data recommendation method may include:

[0078] S401: In response to a data acquisition request sent by a target account, obtain associated account information corresponding to an associated account associated with the target account.

[0079] S402: Construct a target account relationship graph based on the target association relationship.

[0080] S403: Perform a random walk of nodes in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph.

[0081] In step S402 and step S403, a target account relationship graph may be constructed based on the target association relationship. The target association relationship may include association relationships between the associated account and each account in the account set. The target account relationship graph may be as follows: Figure 3a The target account relationship graph shown in , includes account nodes corresponding to accounts in the account set that are associated with the linked account. For example, if a user plays a video posted by a linked account and then immediately plays a video posted by account B in the account set, then the linked account and account B are associated. To construct the target account relationship graph, a random walk of nodes can be performed in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph.

[0082] S404: Calling the first vectorization model to process the target random walk sequence to obtain the associated embedded vector corresponding to the associated account.

[0083] In one implementation, obtaining the first vectorized model can be accomplished by first acquiring training data, which may include relationships between accounts in an account set. After acquiring the training data, a reference account relationship graph can be constructed based on the relationships between accounts in the account set. The reference account relationship graph can include at least one account node. Next, a random walk of nodes is performed in the reference account relationship graph to obtain a reference random walk sequence corresponding to the reference account relationship graph. Once the reference random walk sequence is obtained, the initial vectorized model can be trained based on the reference random walk sequence to obtain the first vectorized model.

[0084] For example, Figure 5a The figure shows a flow chart of determining a first vectorized model according to an embodiment of the present application. The specific implementation process may include the following steps:

[0085] S1: Get the data to be trained.

[0086] Among them, obtaining the data to be trained may include the consumption behavior between the user and each account. For example, for a certain account, the user's consumption behavior towards the account may include following the account, and may also include the consumption behavior towards the multimedia data published by the account. For example, if the multimedia data is a video, the consumption behavior may include the playback time of the video, the cache time, and various interactive behaviors between the user and the video content, such as commenting, forwarding, sharing, collecting, liking, reporting, etc.

[0087] S2: Construct a reference account relationship diagram.

[0088] The reference account relationship graph includes account nodes and the connecting edges between account nodes. Account nodes can be Figure 5a Referring to the circles in the reference account relationship diagram, the connecting edges can be the connecting lines between the circles. For the connecting edges, if a user consumes a multimedia data Page1 published by account A, and then consumes a multimedia data Page2 published by account B, then there can be a connecting edge between the account node corresponding to account A and the account node corresponding to account B. Among them, the consumption can be understood as reading or playing. It should be understood that the consumption can be effective consumption. The effective consumption can specifically refer to the consumption time reaching a preset value. For example, for an article, the article reading time must reach a preset value. For example, for a video, the video time must reach a preset value. The following S2.1-S2.4 are explanations of the various attributes of the connecting edges. Optionally, the above-mentioned reference account relationship diagram can include two types, which can be isomorphic graphs and heterogeneous graphs. The account nodes in the isomorphic graph can be nodes of the same behavior type, and the account nodes in the heterogeneous graph can be nodes of multiple types. For example, they can be nodes of the following behavior type, nodes of the comment behavior type, and nodes of the sharing behavior type.

[0089] S2.1: Source of connection edges: For multimedia data in the form of videos, users’ effective video playback can be used to construct a reference account relationship graph; for multimedia data in the form of pictures and texts, users’ effective reading of pictures and texts can be used to construct a reference account relationship graph.

[0090] S2.2: Time window of the connection edge: used to specify the time period corresponding to the acquired data to be trained. For example, the time window can be 15 days, which means that the acquired data to be trained is data within the past 15 days.

[0091] S2.3: Truncation of connecting edges: This is used to control the length of the following reference random walk sequence. The sequence generated by the user's consumption behavior within half an hour can be used as a reference random walk sequence.

[0092] S2.4: Deleting Connection Edges: You can delete some connection edges in the reference account relationship graph to ensure their validity. This can prevent abnormal user behavior caused by, for example, account porting or inaccurate data recommendations. When constructing the reference account relationship graph, you can use the associated information corresponding to the account nodes to set filtering conditions to ensure that the constructed reference account relationship graph represents the user's actual consumption behavior on the account.

[0093] S3: Determine the reference random walk sequence.

[0094] Among them, a node random walk can be performed in the reference account relationship graph, then for each account node in the reference account relationship graph, multiple reference random walk sequences can be obtained. For example, taking account node A as the starting point, when performing a node random walk, reference random walk sequences ABEF, ABECB, etc. can be obtained. Optionally, when performing a node random walk in the reference account relationship graph, a random walk can be performed in accordance with a weighted random walk method, that is, a walk is performed based on the weight of the edge, for example, Figure 5b The figure shows a schematic diagram of the structure of a reference account relationship diagram with weights provided in an embodiment of the present application. The weight can measure the importance of the connection edge between two account nodes, such as Figure 5b As shown, the weight W1 of the edge connecting account nodes A and B is 0.1, and the weight W2 of the edge connecting account nodes A and C is 0.7. Therefore, when walking from account node A, the direction of the edge with the higher weight can be preferentially chosen, that is, walking towards account node C. The similarity between account nodes can be determined based on the obtained reference random walk sequence. Specifically, the more common account nodes between two account nodes, the higher the similarity between the two account nodes, and the more similar the two account nodes are. Alternatively, the shorter the length of the reference random walk sequence of two account nodes, the more similar the two account nodes are.

[0095] S4: Train the initial vectorized model.

[0096] The initial vectorization model may be a Skip-gram model. Specifically, the reference random walk sequence obtained above may be input into the Skip-gram model to train the Skip-gram model, thereby obtaining a trained initial vectorization model, which is also the first vectorization model.

[0097] In one implementation, after determining the first vectorization model, the associated embedded vector corresponding to each account can be determined using the first vectorization model. Figure 5cThe figure shows a flow chart of determining associated embedded vectors, where the associated relationship is specifically the attention behavior. Figure 5c As can be seen from the figure, we can obtain the following behaviors between accounts and then construct a reference account relationship graph based on these behaviors. After determining the reference account relationship graph, we can perform a random walk of nodes in the reference account relationship graph to obtain a reference random walk sequence corresponding to the reference account relationship graph. Then, after obtaining the reference random walk sequence, we input the reference random walk sequence into the reference account relationship graph to obtain the associated embedded vectors corresponding to each account.

[0098] S405: Determine the portrait embedded vector corresponding to the associated account based on the target portrait data.

[0099] In one implementation, the above process of determining the associated embedded vector mainly uses the interaction behavior information between the user and the associated account. However, for a new account, the content published by the new account needs to go through a cold start process. During this process, the interaction behavior information between the user and the new account, such as the amount of attention and consumption behavior data, is very sparse. In this case, the amount of target association relationship information required to determine the associated embedded vector of the new account is too small. In this case, the account portrait data can be introduced so that the obtained target embedded vector can better characterize the characteristics of the account.

[0100] In one implementation, the target profile data may include category labels for target multimedia data under an associated account, category labels for associated accounts, and target group labels for associated accounts, and may also include other data used to characterize associated accounts, which is not limited in this application. This application uses the example of target profile data including at least one of the category labels for target multimedia data under an associated account, the category labels for associated accounts, and the target group labels for associated accounts as an example for illustration.

[0101] The category labels of the target multimedia data under the associated account may specifically refer to the category labels of the multimedia data published by the associated account within a preset time period, which are ranked in the top L in the category label ranking. For example, the preset time period may be the period within the three months prior to the current time. The categories of multimedia data published by the associated account in the past three months include social, financial, and other categories. The social category may also include categories such as agriculture, rural areas, and farmers, industry, and safety incidents, and the financial category may also include categories such as housing prices, commodity prices, and inflation. Thus, the aforementioned categories of agriculture, rural areas, and farmers, industry, safety incidents, housing prices, commodity prices, and inflation can all serve as category labels, and each of these category labels can be represented using a one-hot encoding. Specifically, taking the associated account A as an example, a statistics server can count the number of multimedia data corresponding to each category label within the past three months prior to the current time. The category labels are then ranked from highest to lowest according to the number of multimedia data corresponding to each category label, and the top L category labels in the ranking result are used as the category labels of the target multimedia data under account A. The aforementioned L can be preset, for example, 3, 20, or the like. For example, assuming L is 3, the category labels of the target multimedia data under account A may include industry, safety accidents, and housing prices.

[0102] Among them, the category labels of the associated accounts can specifically include first-level category labels, second-level category labels, and third-level category labels. The first-level category labels, second-level category labels, and third-level category labels can all be represented by one-hot encoding. Specifically, for an account, the account can publish various multimedia data. It can be understood that multimedia data has corresponding first-level categories (first-level category labels). For example, taking the multimedia data as an article as an example, the first-level categories of the article include society, technology, finance, education, sports, etc., and each first-level category can also have second-level categories (second-level category labels). For example, the second-level category of technology can include Internet technology, mobile technology, mechanical technology, artificial intelligence, etc., and each second-level category can also have third-level categories (third-level category labels). For example, the third-level category of artificial intelligence can include machine learning, deep learning, computer vision, natural language processing, etc.

[0103] In one implementation, the category tags of the associated account in an embodiment of the present application may include N1 first-level category tags, N2 second-level category tags, and N3 third-level category tags, wherein the N1 first-level category tags are the first-level category tags with a cumulative number of tags being the top N1 among the first-level category tags corresponding to the multimedia data published by the associated account, the N2 second-level category tags are the second-level category tags with a cumulative number of tags being the top N2 among the second-level category tags corresponding to the multimedia data published by the associated account, and the N3 third-level category tags are the third-level category tags with a cumulative number of tags being the top N3 among the third-level category tags corresponding to the multimedia data published by the associated account. N1, N2, and N3 can be set in advance. Specifically, assuming that N1, N2, and N3 can all be 5, the number of first-level category tags, the number of second-level category tags, and the number of third-level category tags corresponding to the multimedia data published by the associated account can be counted, and the first-level category tags can be sorted in descending order according to the number of first-level category tags to obtain a first sorting result, the second-level category tags can be sorted in descending order according to the number of second-level category tags to obtain a second sorting result, and the third-level category tags can be sorted in descending order according to the number of third-level category tags to obtain a third sorting result. Furthermore, the top 5 first-level category tags are obtained from the first sorting result, the top 5 second-level category tags are obtained from the second sorting result, and the top 5 third-level category tags are obtained from the third sorting result.

[0104] The target groups of linked accounts can include any of the following: teenagers, young adults, middle-aged adults, and seniors. They can also be categorized by age group, for example, those between 10 and 15 or between 16 and 20. The target group labels of linked accounts can be the labels corresponding to the aforementioned target groups. Each target group can be assigned a label, which can be represented using a one-hot encoding. The target group label can be understood as indicating which accounts the linked account is more suitable for following or spending with.

[0105] In one implementation, if in a specific implementation scenario, there may be multiple vectors corresponding to the category labels of the target multimedia data under the associated account and the vectors corresponding to the category labels of the associated account. For example, if the number of category labels of the target multimedia data under the associated account is set to 20, then there are 20 vectors corresponding to the category labels of the target multimedia data under the associated account; if the number of category labels of the associated account is set to 15, then there are 15 vectors corresponding to the category labels of the associated account. However, in the process of inputting the above-mentioned second vectorization model, considering that the vectors corresponding to the same category label are the same under different accounts, the vectors corresponding to the category labels of the target multimedia data under the associated account can be constructed as a whole into one vector, and the vectors corresponding to the category labels of the associated account can be constructed as a whole into one vector. Through the above processing, the coverage of the category labels can be increased, and it will not be affected by the input order of the category labels. The vector corresponding to the target association relationship, the vector corresponding to the category label of the target multimedia data under the associated account, and the vector corresponding to the category label of the associated account can also be added with a specified vector. The vector can be a vector of all zeros. The specified vector can be used as an empty input placeholder. When there is an empty vector in the input of the second vectorization model, the final target embedded vector is not affected by the empty value.

[0106] In one implementation, target profile data for the associated account can be obtained. This target profile data can include at least one of a category label for the target multimedia data under the associated account, a category label for the associated account, and a target group label for the associated account. After obtaining the target profile data, a second vectorization model can be invoked to process the target profile data to obtain a profile embedded vector corresponding to the associated account.

[0107] For example, if the target portrait data includes one of the category label of the target multimedia data under the associated account, the category label of the associated account, and the target group label of the associated account, assuming that the target portrait data is the category label of the target multimedia data under the associated account, the second vectorization model can be called to process the category label of the target multimedia data under the associated account, and the resulting embedded vector is the portrait embedded vector.

[0108] For example, if the target portrait data includes any two of the category labels of the target multimedia data under the associated account, the category labels of the associated account, and the target group labels of the associated account. Assuming that the target portrait data is the category label of the target multimedia data under the associated account and the category label of the associated account, the second vectorization model can be called to process the category label of the target multimedia data under the associated account to obtain a first portrait embedded vector, and the second vectorization model can be called to process the category label of the associated account to obtain a second portrait embedded vector. Then, after obtaining the first portrait embedded vector and the second portrait embedded vector, the portrait embedded vector corresponding to the associated account can be constructed based on the first portrait embedded vector and the second portrait embedded vector. Among them, the above-mentioned construction of the portrait embedded vector corresponding to the associated account can be understood as the portrait embedded vector corresponding to the associated account including the first portrait embedded vector and the second portrait embedded vector.

[0109] For example, if the target portrait data includes the category label of the target multimedia data under the associated account, the category label of the associated account, and the target group label of the associated account, then the second vectorization model can be called to process the category label of the target multimedia data under the associated account to obtain a first portrait embedded vector, the second vectorization model can be called to process the category label of the associated account to obtain a second portrait embedded vector, and the second vectorization model can be called to process the target group label of the associated account to obtain a third portrait embedded vector; then, after obtaining the first portrait embedded vector, the second portrait embedded vector, and the third portrait embedded vector, the portrait embedded vector corresponding to the associated account can be constructed based on the first portrait embedded vector, the second portrait embedded vector, and the third portrait embedded vector. Among them, the above-mentioned construction of the portrait embedded vector corresponding to the associated account can be understood as the portrait embedded vector corresponding to the associated account including the first portrait embedded vector, the second portrait embedded vector, and the third portrait embedded vector.

[0110] S406: Fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filtering the account set to obtain a similar account set based on the target embedded vector.

[0111] S407: Acquire multimedia data under each similar account in the similar account set, and recommend data to the target account based on the multimedia data.

[0112] The specific implementation of steps S401 , S406 and S407 can refer to the specific description of steps S201 , S203 and S204 in the above embodiment, which will not be repeated here.

[0113] In an embodiment of the present application, in response to a data acquisition request sent by a target account, associated account information corresponding to associated accounts associated with the target account is obtained. A target account relationship graph is constructed based on the target association relationship. A node random walk is performed in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph. A first vectorization model is invoked to process the target random walk sequence to obtain an associated embedded vector corresponding to the associated account. A portrait embedded vector corresponding to the associated account is determined based on the target portrait data. The associated embedded vector and the portrait embedded vector are fused to obtain a target embedded vector, and a similar account set is filtered from the account set based on the target embedded vector. Multimedia data for each similar account in the similar account set is obtained, and data recommendations are made to the target account based on the multimedia data. By implementing the above method, the behavioral data between the user and the account and the account's portrait data can be fully utilized. The account's embedded vector is determined based on the behavioral data and the portrait data, and similar accounts of the account are recalled based on the embedded vector. This can improve the cold start speed of multimedia data distribution corresponding to the account, improve data recommendation efficiency, and enhance the user experience.

[0114] See also Figure 6 , Figure 6 This is a schematic diagram of a data recommendation device according to an embodiment of the present application. The data recommendation device described in this embodiment includes:

[0115] An acquisition unit 601 is configured to, in response to a data acquisition request sent by a target account, acquire associated account information corresponding to an associated account associated with the target account, wherein the associated account information includes a target association relationship and target profile data. The target association relationship includes an association relationship between the associated account and each account in an account set. The target profile data is used to indicate a category feature of the associated account.

[0116] A first determining unit 602 is configured to determine an associated embedded vector corresponding to the associated account based on the target association relationship, and determine a portrait embedded vector corresponding to the associated account based on the target portrait data;

[0117] A second determining unit 603 is configured to fuse the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filter the account set to obtain a similar account set based on the target embedded vector;

[0118] The recommendation unit 604 is configured to obtain multimedia data from each similar account in the similar account set, and recommend data to the target account based on the multimedia data.

[0119] In one implementation, the first determining unit 602 is specifically configured to:

[0120] Building a target account relationship graph based on the target association relationship, wherein the target account relationship graph includes at least one account node;

[0121] Performing a random walk of nodes in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph;

[0122] The first vectorized model is called to process the target random walk sequence to obtain an associated embedded vector corresponding to the associated account.

[0123] In one implementation, the target portrait data includes at least one of a category label of the target multimedia data under the associated account, a category label of the associated account, and a target group label of the associated account; and the first determining unit 602 is specifically configured to:

[0124] Calling the second vectorization model to process the category labels of the target multimedia data under the associated account to obtain a first portrait embedded vector;

[0125] Calling the second vectorization model to process the category label of the associated account to obtain a second portrait embedded vector;

[0126] Calling the second vectorization model to process the pointing group label of the associated account to obtain a third portrait embedded vector;

[0127] Based on the first portrait embedded vector, the second portrait embedded vector and the third portrait embedded vector, a portrait embedded vector corresponding to the associated account is constructed.

[0128] In one implementation, the apparatus further includes a training unit 605, wherein the training unit 605 is specifically configured to:

[0129] Acquire data to be trained, where the data to be trained includes associations between accounts in the account set;

[0130] Building a reference account relationship graph based on the association relationship between each account in the account set, wherein the reference account relationship graph includes at least one account node;

[0131] Performing a random walk of nodes in the reference account relationship graph to obtain a reference random walk sequence corresponding to the reference account relationship graph;

[0132] The initial vectorized model is trained based on the reference random walk sequence to obtain a first vectorized model.

[0133] In one implementation, the second determining unit 603 is specifically configured to:

[0134] Obtaining an association weight corresponding to the association embedded vector and a portrait weight corresponding to the portrait embedded vector;

[0135] Determining a weighted association embedded vector corresponding to the association embedded vector based on the association weight, and determining a weighted portrait embedded vector corresponding to the portrait embedded vector based on the portrait weight;

[0136] The weighted association embedded vector and the weighted portrait embedded vector are concatenated to obtain a target embedded vector.

[0137] In one implementation, the second determining unit 603 is specifically configured to:

[0138] Obtaining an embedded vector set corresponding to the account set, the embedded vector set including an embedded vector corresponding to each account in the account set;

[0139] Determining the similarity between the target embedded vector and each embedded vector in the embedded vector set, and screening out at least one similar embedded vector having a similarity with the target embedded vector greater than a preset similarity;

[0140] Obtaining an account corresponding to each similar embedded vector in the at least one similar embedded vector to obtain at least one similar account;

[0141] A similar account set is constructed based on the at least one similar account.

[0142] In one implementation, the recommendation unit 604 is specifically configured to:

[0143] Acquiring multimedia data under each similar account in the similar account set from a content database; wherein the multimedia data includes at least one multimedia file, each multimedia file being uploaded by each similar account through an upstream and downstream content interface service and obtained after undergoing review and duplicate elimination processing;

[0144] sorting each multimedia file in the multimedia data according to a preset sorting rule to obtain a sorting result corresponding to each multimedia file;

[0145] The multimedia files ranked in the top N positions are recommended to the target account, where N is a positive integer.

[0146] It is understood that the division of units in the embodiments of the present application is schematic and is merely a logical functional division. In actual implementation, other division methods may be used. The functional units in the embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0147] See also Figure 7 , Figure 7 701 , a memory 702 , and a network interface 703 . The processor 701 , the memory 702 , and the network interface 703 can exchange data.

[0148] The processor 701 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0149] The memory 702 may include a read-only memory and a random access memory, and provides program instructions and data to the processor 701. A portion of the memory 702 may also include a non-volatile random access memory. When the processor 701 calls the program instructions, it is used to execute:

[0150] In response to a data acquisition request sent by a target account, acquiring associated account information corresponding to an associated account associated with the target account, the associated account information including a target association relationship and target profile data, the target association relationship including an association relationship between the associated account and each account in the account set, and the target profile data indicating a category feature of the associated account;

[0151] Determining the associated embedded vector corresponding to the associated account based on the target association relationship, and determining the portrait embedded vector corresponding to the associated account based on the target portrait data;

[0152] fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filtering the account set to obtain a similar account set based on the target embedded vector;

[0153] Multimedia data under each similar account in the similar account set is obtained, and data is recommended to the target account based on the multimedia data.

[0154] In one implementation, the processor 701 is specifically configured to:

[0155] Building a target account relationship graph based on the target association relationship, wherein the target account relationship graph includes at least one account node;

[0156] Performing a random walk of nodes in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph;

[0157] The first vectorized model is called to process the target random walk sequence to obtain an associated embedded vector corresponding to the associated account.

[0158] In one implementation, the target portrait data includes at least one of a category label of the target multimedia data under the associated account, a category label of the associated account, and a target group label of the associated account; the processor 701 is specifically configured to:

[0159] Calling the second vectorization model to process the category labels of the target multimedia data under the associated account to obtain a first portrait embedded vector;

[0160] Calling the second vectorization model to process the category label of the associated account to obtain a second portrait embedded vector;

[0161] Calling the second vectorization model to process the pointing group label of the associated account to obtain a third portrait embedded vector;

[0162] Based on the first portrait embedded vector, the second portrait embedded vector and the third portrait embedded vector, a portrait embedded vector corresponding to the associated account is constructed.

[0163] In one implementation, the processor 701 is further configured to:

[0164] Acquire data to be trained, where the data to be trained includes associations between accounts in the account set;

[0165] Building a reference account relationship graph based on the association relationship between each account in the account set, wherein the reference account relationship graph includes at least one account node;

[0166] Performing a random walk of nodes in the reference account relationship graph to obtain a reference random walk sequence corresponding to the reference account relationship graph;

[0167] The initial vectorized model is trained based on the reference random walk sequence to obtain a first vectorized model.

[0168] In one implementation, the processor 701 is specifically configured to:

[0169] Obtaining an association weight corresponding to the association embedded vector and a portrait weight corresponding to the portrait embedded vector;

[0170] Determining a weighted association embedded vector corresponding to the association embedded vector based on the association weight, and determining a weighted portrait embedded vector corresponding to the portrait embedded vector based on the portrait weight;

[0171] The weighted association embedded vector and the weighted portrait embedded vector are concatenated to obtain a target embedded vector.

[0172] In one implementation, the processor 701 is specifically configured to:

[0173] Obtaining an embedded vector set corresponding to the account set, the embedded vector set including an embedded vector corresponding to each account in the account set;

[0174] Determining the similarity between the target embedded vector and each embedded vector in the embedded vector set, and screening out at least one similar embedded vector having a similarity with the target embedded vector greater than a preset similarity;

[0175] Obtaining an account corresponding to each similar embedded vector in the at least one similar embedded vector to obtain at least one similar account;

[0176] A similar account set is constructed based on the at least one similar account.

[0177] In one implementation, the processor 701 is specifically configured to:

[0178] Acquiring multimedia data under each similar account in the similar account set from a content database; wherein the multimedia data includes at least one multimedia file, each multimedia file being uploaded by each similar account through an upstream and downstream content interface service and obtained after undergoing review and duplicate elimination processing;

[0179] sorting each multimedia file in the multimedia data according to a preset sorting rule to obtain a sorting result corresponding to each multimedia file;

[0180] The multimedia files ranked in the top N positions are recommended to the target account, where N is a positive integer.

[0181] The embodiment of the present application further provides a computer storage medium in which program instructions are stored. When the program is executed, the program may include: Figure 2 or Figure 4 Part or all of the steps of the data recommendation method in the corresponding embodiment.

[0182] It should be noted that for the aforementioned various method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0183] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0184] The present application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a server reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the server to perform the steps performed in the above-described method embodiments.

[0185] The above is a detailed introduction to a data recommendation method, device, server and storage medium provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core ideas. At the same time, for general technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A data recommendation method, characterized in that: include: In response to a data acquisition request sent by a target account, associated account information corresponding to an associated account associated with the target account is acquired, the associated account information including a target association relationship and target profile data, the target association relationship including an association relationship between the associated account and each account in the account set, the target profile data being used to indicate a category characteristic of the associated account; the associated account being the account corresponding to the multimedia data currently being read by the user, and the association relationship between the associated account and a first account in the account set indicating that the target account has generated valid consumption behavior with the associated account and the first account, respectively; The target portrait data includes the category label of the target multimedia data under the associated account, the category label of the associated account, and the target group label of the associated account; Determining the associated embedded vector corresponding to the associated account based on the target association relationship, and determining the portrait embedded vector corresponding to the associated account based on the target portrait data; fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filtering the account set to obtain a similar account set based on the target embedded vector; Multimedia data under each similar account in the similar account set is obtained, and data is recommended to the target account based on the multimedia data.

2. The method according to claim 1, characterized in that The determining, based on the target association relationship, the associated embedded vector corresponding to the associated account includes: Building a target account relationship graph based on the target association relationship, wherein the target account relationship graph includes at least one account node; Performing a random walk of nodes in the target account relationship graph to obtain a target random walk sequence corresponding to the target account relationship graph; The first vectorized model is called to process the target random walk sequence to obtain an associated embedded vector corresponding to the associated account.

3. The method according to claim 1, characterized in that The determining, based on the target portrait data, the portrait embedded vector corresponding to the associated account includes: Calling the second vectorization model to process the category labels of the target multimedia data under the associated account to obtain a first portrait embedded vector; Calling the second vectorization model to process the category label of the associated account to obtain a second portrait embedded vector; Calling the second vectorization model to process the pointing group label of the associated account to obtain a third portrait embedded vector; Based on the first portrait embedded vector, the second portrait embedded vector and the third portrait embedded vector, a portrait embedded vector corresponding to the associated account is constructed.

4. The method according to claim 2, characterized in that The method further comprises: Acquire data to be trained, where the data to be trained includes associations between accounts in the account set; Building a reference account relationship graph based on the association relationship between each account in the account set, wherein the reference account relationship graph includes at least one account node; Performing a random walk of nodes in the reference account relationship graph to obtain a reference random walk sequence corresponding to the reference account relationship graph; The initial vectorized model is trained based on the reference random walk sequence to obtain a first vectorized model.

5. The method according to claim 1, wherein The fusing the associated embedded vector and the portrait embedded vector to obtain a target embedded vector includes: Obtaining an association weight corresponding to the association embedded vector and a portrait weight corresponding to the portrait embedded vector; Determining a weighted association embedded vector corresponding to the association embedded vector based on the association weight, and determining a weighted portrait embedded vector corresponding to the portrait embedded vector based on the portrait weight; The weighted association embedded vector and the weighted portrait embedded vector are concatenated to obtain a target embedded vector.

6. The method according to claim 1, characterized in that The filtering the account set to obtain a similar account set based on the target embedded vector includes: Obtaining an embedded vector set corresponding to the account set, the embedded vector set including an embedded vector corresponding to each account in the account set; Determining the similarity between the target embedded vector and each embedded vector in the embedded vector set, and screening out at least one similar embedded vector having a similarity with the target embedded vector greater than a preset similarity; Obtaining an account corresponding to each similar embedded vector in the at least one similar embedded vector to obtain at least one similar account; A similar account set is constructed based on the at least one similar account.

7. The method according to claim 1, characterized in that The acquiring multimedia data of each similar account in the similar account set and recommending data to the target account based on the multimedia data includes: Acquiring multimedia data under each similar account in the similar account set from a content database; wherein the multimedia data includes at least one multimedia file, each multimedia file being uploaded by each similar account through an upstream and downstream content interface service and obtained after undergoing review and duplicate elimination processing; sorting the multimedia files in the multimedia data in descending order according to a preset sorting rule to obtain a sorting result corresponding to each multimedia file; The multimedia files ranked in the top N positions are recommended to the target account, where N is a positive integer.

8. A data recommendation device, characterized in that: include: an acquisition unit, configured to, in response to a data acquisition request sent by a target account, acquire associated account information corresponding to an associated account associated with the target account, the associated account information including a target association relationship and target portrait data, the target association relationship including an association relationship between the associated account and each account in the account set, the target portrait data being used to indicate a category characteristic of the associated account; the associated account being the account corresponding to multimedia data currently being read by the user, the association relationship between the associated account and a first account in the account set indicating that the target account has generated valid consumption behavior with the associated account and the first account, respectively; The target portrait data includes the category label of the target multimedia data under the associated account, the category label of the associated account, and the target group label of the associated account; a first determining unit, configured to determine an associated embedded vector corresponding to the associated account based on the target association relationship, and determine a portrait embedded vector corresponding to the associated account based on the target portrait data; a second determining unit, configured to fuse the associated embedded vector and the portrait embedded vector to obtain a target embedded vector, and filter the account set to obtain a similar account set based on the target embedded vector; The recommendation unit is configured to obtain multimedia data under each similar account in the similar account set, and recommend data to the target account based on the multimedia data.

9. A server, characterized in that: The method comprises a processor, a memory and a network interface, wherein the processor, the memory and the network interface are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that The computer program product comprises computer instructions, and the computer instructions are executed to implement the method according to any one of claims 1 to 7.

Images