Material recall method and device, electronic equipment and storage medium

Abstract

The invention provides a material recall method and device, electronic equipment and a storage medium. The material recall method comprises the following steps: in response to a material recommendation request of a target user, obtaining target user characteristics of the target user and a target recommendation scene indicated by the material recommendation request; inputting the target user features into a trained cross-scene multi-target model to obtain a corresponding target user vector, and determining similar users of the target user based on the target user vector; performing availability filtering on the first historical materials of the similar users to obtain candidate materials; and based on the target user vector and the material vectors of the candidate materials, determining recalled materials of the target user from the candidate materials. According to the invention, the accuracy of recommending materials to the target user can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a material recall method and device, an electronic device, and a storage medium. Background Art

[0002] With the rapid development of computer technology, the use of big data to recommend materials that users may be interested in from the massive amount of materials on the Internet has emerged.

[0003] At present, the commonly used user collaborative recall in related technologies is to directly recommend materials that are liked by users similar to the current user. However, this recall method involves a single scene, has poor learning ability and lacks personalized considerations, resulting in low accuracy in recommending materials for the current user. Summary of the invention

[0004] The embodiments of the present application provide a material recall method and device, an electronic device, and a storage medium, which can effectively improve the accuracy of recommending materials to target users.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] In a first aspect, an embodiment of the present application provides a material recall method, the method comprising:

[0007] In response to a material recommendation request of a target user, obtaining a target user feature of the target user and a target recommendation scenario indicated by the material recommendation request;

[0008] Input the target user features into the trained cross-scenario multi-objective model to obtain a corresponding target user vector, and determine similar users of the target user based on the target user vector; wherein the cross-scenario multi-objective model is obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios;

[0009] Availability filtering is performed on the first historical material of the similar user to obtain candidate materials; wherein the first historical material is a material associated with the similar user in the target recommendation scenario;

[0010] Based on the target user vector and the material vectors of the candidate materials, recall materials for the target user are determined from the candidate materials.

[0011] In a second aspect, an embodiment of the present application provides a material recall device, comprising:

[0012] an acquisition module, configured to, in response to a material recommendation request of a target user, acquire a target user feature of the target user and a target recommendation scenario indicated by the material recommendation request;

[0013] A first determination module is configured to input the target user feature into the trained cross-scenario multi-objective model to obtain a corresponding target user vector, and determine similar users of the target user based on the target user vector; wherein the cross-scenario multi-objective model is obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios;

[0014] A processing module, configured to perform availability filtering on the first historical material of the similar user to obtain a candidate material; wherein the first historical material is a material associated with the similar user in the target recommendation scenario;

[0015] The second determination module is configured to determine the recalled materials of the target user from the candidate materials based on the target user vector and the material vectors of the candidate materials.

[0016] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0017] Memory for storing computer programs or executable instructions;

[0018] The processor is used to implement the method provided in the first aspect of the embodiment of the present application when executing the computer program or executable instructions stored in the memory.

[0019] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having stored thereon a computer program or executable instructions, which, when executed by a processor, implements the method provided in the first aspect of the embodiment of the present application.

[0020] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program or executable instructions. When the computer program or executable instructions are executed by a processor, the method provided in the first aspect of the embodiment of the present application is implemented.

[0021] The embodiments of the present application have the following beneficial effects:

[0022] A material recall method provided in an embodiment of the present application can, when a material recommendation request of a target user is detected, determine similar users of the target user by inputting the target user characteristics of the target user into a target user vector obtained by a cross-scenario multi-objective model, and filter out candidate materials from the first historical materials associated with the similar users in the target recommendation scenario indicated by the material recommendation request, and then determine the recall material of the target user from the candidate materials based on the target user vector and the material vector of the candidate material.

[0023] In this way, unlike the related art of directly recommending materials that are liked by users similar to the current user, the embodiment of the present application can, after determining similar users of the target user, perform multiple filtering from the first historical materials associated with the similar users to obtain the recalled materials for the target user, thereby effectively improving the accuracy of recommending materials to the target user while ensuring personalization; and, by screening out candidate materials from the first historical materials associated with similar users in the target recommendation scenario, the candidate materials can be determined more accurately, thereby improving the matching degree of the recalled materials and the target user. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the architecture of a material recall system provided in an embodiment of the present application.

[0025] Figure 2 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0026] Figure 3 This is a schematic diagram of the process of the material recall method provided in the embodiment of the present application. Figure 1 .

[0027] Figure 4 It is a structural schematic diagram of a preset model in the material recall method provided in an embodiment of the present application.

[0028] Figure 5 It is a schematic diagram of the architecture of a traditional recommendation system provided in an embodiment of the present application.

[0029] Figure 6 This is a schematic diagram of the process of the material recall method provided in the embodiment of the present application. Figure 2 . DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0031] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0032] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0034] At present, the commonly used user collaborative recall in related technologies is to directly recommend materials that are liked by users similar to the current user. However, this recall method involves a single scene, has poor learning ability and lacks personalized considerations, resulting in low accuracy in recommending materials for the current user.

[0035] In response to the above technical problems, the embodiments of the present application provide a material recall method and device, an electronic device and a storage medium, which can effectively improve the accuracy of recommending materials to target users.

[0036] The electronic device provided in the embodiment of the present application for implementing the above-mentioned material recall method is described below. The electronic device provided in the embodiment of the present application can be implemented as various types of terminal devices such as a laptop computer, a desktop computer, a mobile device, etc., and can also be implemented as a background server of a network platform (hereinafter referred to as a server). Below, an exemplary application when the electronic device is implemented as a server of a network platform will be described.

[0037] See also Figure 1 , Figure 1 It is an architectural diagram of a material recall system provided in an embodiment of the present application. To implement the material recall method provided in an embodiment of the present application, a terminal device 101 (the material recall system of the present embodiment is not limited in the number of terminal devices included) is connected to a server 102 via a network 103. The network 103 may be a wide area network or a local area network, or a combination of the two.

[0038] In some possible implementations, a user can log in to an application corresponding to the network platform through a terminal device 101, and the terminal device 101 exchanges information with the server 102 through a network 103; the server 102 obtains the target user and the target user characteristics of the target user through the network 103; the server 102 inputs the target user and the target user characteristics of the target user into a trained cross-scenario multi-target model to obtain a corresponding target user vector, and then determines the recall materials of the target user based on the target user vector. After the materials have been roughly sorted, finely sorted, and followed by rules, the recalled materials can be sent to the terminal device 101 through the network 103, and displayed on the application interface of the terminal device 101.

[0039] In some embodiments, the server 102 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms. The terminal device 101 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart watch, etc., but is not limited thereto. The terminal device and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application.

[0040] See also Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application; wherein, Figure 2 The electronic device 200 shown includes: at least one processor 210, a memory 250, at least one network interface 220 and a user interface 230. The various components in the electronic device 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 240 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 240 is not described in detail. Figure 2 Various buses are labeled as bus system 240 .

[0041] The processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0042] The user interface 230 includes one or more output devices 231 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 230 also includes one or more input devices 232, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0043] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices that are physically remote from the processor 210.

[0044] The memory 250 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 250 described in the embodiments of the present application is intended to include any suitable type of memory.

[0045] In some embodiments, memory 250 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0046] Operating system 251, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0047] A network communication module 252, for reaching other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 include: Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.;

[0048] a presentation module 253 for enabling presentation of information via one or more output devices 231 (e.g., display screen, speaker, etc.) associated with the user interface 230 (e.g., a user interface for operating peripherals and displaying content and information);

[0049] The input processing module 254 is used to detect one or more user inputs or interactions from one of the one or more input devices 232 and translate the detected inputs or interactions.

[0050] In some embodiments, the material recall device provided in the embodiments of the present application can be implemented in a software manner. Figure 2 The material recall device 255 stored in the memory 250 is shown, which can be software in the form of a program and a plug-in, etc., including the following software modules: an acquisition module 2551, a first determination module 2552, a processing module 2553 and a second determination module 2554. These modules are logical, so they can be arbitrarily combined or further split according to the functions implemented. The functions of each module will be explained below.

[0051] In other embodiments, the material recall device provided in the embodiments of the present application can be implemented in hardware. As an example, the device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the material recall method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application specific integrated circuits (ASIC), DSP, programmable logic device (PLD), complex programmable logic device (CPLD), field programmable gate array (FPGA) or other electronic components.

[0052] The material recall method provided in the embodiment of the present application will be described below in conjunction with the exemplary application and implementation of the electronic device provided in the embodiment of the present application.

[0053] See also Figure 3 , Figure 3 This is a schematic diagram of the process of the material recall method provided in the embodiment of the present application. Figure 1 ,like Figure 3 As shown, the following steps may be included:

[0054] S310, in response to a material recommendation request of a target user, obtaining a target user feature of the target user and a target recommendation scenario indicated by the material recommendation request;

[0055] S320, inputting the target user features into the trained cross-scenario multi-objective model to obtain a corresponding target user vector, and determining similar users of the target user based on the target user vector; wherein the cross-scenario multi-objective model is obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios;

[0056] S330, filtering the first historical materials of similar users for availability to obtain candidate materials; wherein the first historical materials are materials associated with similar users in a target recommendation scenario;

[0057] S340: Determine a recalled material from the candidate materials based on the target user vector and the material vector of the candidate material.

[0058] In an embodiment of the present application, the above-mentioned material recall method can be applied to material recommendation scenarios that can perform video recommendations, advertising recommendations, or popular Weibo content recommendations, and can be executed by electronic devices related to a network platform; the above-mentioned network platform includes a social media platform, an education platform, or a shopping platform, etc., and the above-mentioned electronic device may include a server of the network platform.

[0059] In step S310, when detecting a material recommendation request from a target user, the server may obtain the target user characteristics of the target user and the target recommendation scenario indicated by the material recommendation request, so as to determine candidate materials according to the target user characteristics and the target recommendation scenario of the target user in a subsequent process.

[0060] Here, the target user may be the object of the server to recommend materials. The material recommendation request may be a material recommendation request sent by the target user to the server by refreshing a recommendation page or clicking a recommendation button on a terminal device.

[0061] It should be noted that the above-mentioned material recommendation request may carry the associated data of the target user; for example, the associated data of the target user may include: the identification information of the target user, historical query information or the recommendation scene where the target user is currently located, etc. Among them, the identification information of the target user may include: the identity document (ID) and geographic location of the target user, etc.; the historical query information of the target user may include: the historical browsing content, browsing time and transaction information of the target user, etc., which are not limited in the embodiments of the present application. In some embodiments, the target user may be a new user of the terminal device or the accessed application, in which case the historical query information may not be included in the associated data of the target user.

[0062] In the embodiment of the present application, the target user features of the target user can be obtained by extracting features from the associated data of the target user. After obtaining the associated data of the target user, the server can perform data preprocessing on the associated data of the target user to obtain user data with a uniform format and without abnormalities, and then extract the target user features of the target user that can reflect the user's characteristics and behavior habits from the preprocessed user data.

[0063] It is understandable that the server can perform dimensionality reduction processing on the preprocessed user data through a hash algorithm to convert the user data into a hash value of a fixed length, and then discretize the reduced dimensionality user data to simplify the user data, thereby facilitating the subsequent extraction of target user features from the simplified user data.

[0064] It should be noted that the target recommendation scenarios indicated by the above material recommendation request may include but are not limited to: Weibo recommendation scenarios and video / blog post recommendation scenarios. Among them, Weibo recommendation scenarios may refer to scenarios where blog posts and / or videos are recommended to users on the Weibo recommendation page; video post recommendation scenarios may refer to scenarios where videos are recommended to users when users continue to scroll down after playing videos from other scenarios; and blog post recommendation scenarios may refer to scenarios where users continue to scroll down after clicking on a blog post and other blog posts are recommended to users.

[0065] In step S320, the server may input the target user feature into the trained cross-scenario multi-target model to obtain the corresponding target user vector, and determine similar users whose similarity to the target user vector is greater than a third preset threshold from the sample user set obtained from the offline task according to the target user vector. Here, the similar user may be a user with similar interests to the target user, that is, the similarity between the user vector of the similar user and the target user vector of the target user is greater than the third preset threshold.

[0066] It should be noted that the third preset threshold can be set according to the actual application scenario, and the embodiment of the present application does not limit it.

[0067] Here, the cross-scenario multi-objective model may include a user tower shared by multiple recommendation scenarios and a material tower corresponding to each recommendation scenario, that is, a material tower may correspond to a material collected in a recommendation scenario. The cross-scenario multi-objective model can be obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios; wherein, the training samples of multiple recommendation scenarios may include the visiting users collected in multiple recommendation scenarios and the materials associated with the visiting users, and the task objectives may include but are not limited to: click-through rate, completion rate or playback time, etc.

[0068] The target user vector may be a user vector obtained by subjecting the target user features of the target user to feature conversion processing performed on the user tower of the cross-scenario multi-target model.

[0069] It is understandable that the server can perform offline tasks according to a preset time period to build a preset list storing user vectors and related feature vectors of the first sample user, so that when the target user vector is obtained in real time, similar users with similar interests to the target user can be found from the preset list.

[0070] Exemplarily, the server may perform an offline task once an hour to collect user features from the first sample users who have visited an application such as Weibo in the past eight days, and construct a preset list based on the user vectors corresponding to the user features of the first sample users. When the target user vector is obtained in real time, the user vector of the first sample user whose similarity with the target user vector is greater than a third preset threshold may be found from the preset list to determine similar users; or, the cluster center vector with the highest similarity with the target user vector may be found from the preset list first, and then the first sample user associated with the cluster center vector may be determined as a similar user, etc.

[0071] In step S330, the server may determine candidate materials from the first historical materials associated with similar users in the above recommendation scenario, so as to determine the recalled materials to be recommended to the target user from the candidate materials in a subsequent process.

[0072] Here, the first historical material may be a material associated with similar users that is collected in advance in the above-mentioned target recommendation scenario; for example, the first historical material may be a material that has been recently clicked or played by similar users in a microblog recommendation scenario or a video recommendation scenario.

[0073] In the embodiment of the present application, the candidate material may be a material obtained by filtering the availability of the first historical material associated with the similar user.

[0074] It can be understood that the above-mentioned availability filtering of the first historical materials of similar users to obtain candidate materials may include but is not limited to: directly determining the first historical materials that are associated with the similar users in the target recommendation scenario and meet the preset playback conditions as candidate materials; or, based on the first sample materials collected in the target recommendation scenario, availability filtering of the first historical materials of similar users to obtain candidate materials; or, determining the materials in the first historical materials that meet the preset playback conditions as second historical materials, and then based on the material vector of the first sample materials collected in the target recommendation scenario, availability filtering of the second historical materials to obtain candidate materials, etc.

[0075] In step S340, the server may determine the target user's recall material from the candidate materials according to the target user vector and the material vector of the candidate materials.

[0076] Here, the material vector of the candidate material can be the material vector obtained after inputting the material features of the candidate material into the material tower of the cross-scenario multi-objective model for feature conversion processing; or, it can also be the material vector determined from the material vector of the first sample material collected in the target recommendation scenario.

[0077] The material features of the candidate materials may be obtained by extracting features from the associated data of the candidate materials. After obtaining the associated data of the candidate materials, the server may perform data preprocessing on the associated data of the candidate materials to obtain material data with a uniform format and without abnormalities, and then extract material features of the candidate materials that can reflect the content or theme of the materials from the preprocessed material data.

[0078] It is understandable that the server can perform dimensionality reduction processing on the preprocessed material data through a hash algorithm to convert the material data into a hash value of a fixed length, and then discretize the reduced dimensionality material data to simplify the material data, so as to facilitate the subsequent extraction of material features of candidate materials from the simplified material data.

[0079] It should be noted that the associated data of the candidate material may be the historical material data stored in the network platform of the server, or it may be the material data collected by the server from the social platform through a web crawler, etc. Among them, the associated data of the candidate material may include: identification information, attribute information or statistical information of the candidate material, etc.; the identification information of the candidate material may include: the ID and label of the candidate material, etc.; the attribute information of the candidate material may include: the release time, release location, content type and content theme of the candidate material, etc.; the statistical information of the candidate material may include: the number of views, comments, likes and shares of the candidate material, etc., which is not limited in the embodiments of the present application.

[0080] In the embodiment of the present application, the recalled materials can be determined from the candidate materials, and returned to the rough sorting, fine sorting and rule-based steps for re-screening before being recommended to the target users.

[0081] It can be understood that the embodiment of the present application can sort the candidate materials according to the similarity between the target user vector and the material vector of the candidate materials, and determine a preset number of candidate materials sorted from high to low as recalled materials.

[0082] It should be noted that recalled materials are the recall results obtained through collaborative recall by users. After the recalled materials are determined, they still need to go through rough sorting, fine sorting and rule-based stages before the materials recommended to target users can be obtained.

[0083] The material recall method provided in the embodiment of the present application can, when a material recommendation request of a target user is detected, determine similar users of the target user by inputting the target user characteristics of the target user into a target user vector obtained by a cross-scenario multi-objective model, and filter out candidate materials from the first historical materials associated with similar users in the target recommendation scenario indicated by the material recommendation request, and then determine the recall material of the target user from the candidate materials based on the target user vector and the material vector of the candidate material.

[0084] In this way, unlike the related art of directly recommending materials that are liked by users similar to the current user, the embodiment of the present application can, after determining similar users of the target user, perform multiple filtering from the first historical materials associated with the similar users to obtain the recalled materials for the target user, thereby effectively improving the accuracy of recommending materials to the target user while ensuring personalization; and, by screening out candidate materials from the first historical materials associated with similar users in the target recommendation scenario, the candidate materials can be determined more accurately, thereby improving the matching degree of the recalled materials and the target user.

[0085] In some embodiments, step S320 may include:

[0086] Based on the similarity between the target user vector and each cluster center vector in the preset list, determine the target cluster center vector from the cluster center vectors; wherein the preset list is constructed by inputting the user features of the first sample user in the first preset period into the cross-scenario multi-objective model to obtain the corresponding user vector;

[0087] The target first sample user associated with the target cluster center vector in the preset list is determined as a similar user; wherein the similarity between the user vector of the target first sample user and the target cluster center vector is greater than a first preset threshold.

[0088] In an embodiment of the present application, the server can determine the target cluster center vector from the cluster center vectors based on the similarity between the target user vector and each cluster center vector in the preset list, and then determine the target first sample user associated with the target cluster center vector in the preset list as a similar user, so that in the subsequent process, the candidate material can be determined based on the first historical material of the similar user.

[0089] Here, the preset list may be constructed by the server executing an offline task; the preset list may store multiple cluster center vectors, each cluster center vector being associated with multiple first sample users. The cluster center vector may represent the cluster vector of the center position of each cluster or cluster in the preset list; the first sample user may be the sample user corresponding to each user vector in the preset list.

[0090] It is understandable that the embodiment of the present application may determine the cluster center vector with the highest similarity with the target user vector in the preset list as the target cluster center vector; or, the cluster center vector with a similarity with the target user vector in the preset list greater than a fourth preset threshold may be determined as the target cluster center vector. In other words, the target cluster center vector may be one or more, and the embodiment of the present application does not limit this.

[0091] In an embodiment of the present application, each cluster center vector in the preset list is associated with at least two first sample users, and the similarity between the user vector of the first sample user and the cluster center vector may be greater than a first preset threshold. Therefore, after determining the target cluster center vector, the target first sample user associated with the target cluster center vector in the preset list may be directly determined as a similar user, that is, a similar user with similar interests to the target user.

[0092] Exemplarily, an inverted file (IVF) inverted index can be used to construct the user vector of the first sample user to obtain a preset list. The basic principle of the IVF inverted index is to divide the user vector of the first sample user into multiple subsets, each subset is called a cluster center or a cluster, and each cluster has a representative vector, called a cluster center vector; by constructing an inverted list, the cluster center vector is associated with the user vector of the first sample user belonging to the cluster. This index structure can effectively reduce the amount of calculation for determining similar users and improve processing efficiency.

[0093] It should be noted that the first preset threshold value may be the same as the third preset threshold value, or may be different from the third preset threshold value, and the embodiment of the present application does not limit this. Similarly, the fourth preset threshold value may be the same as the first preset threshold value, or may be different from the first preset threshold value.

[0094] In an embodiment of the present application, a target cluster center vector can be determined from the cluster center vectors based on the similarity between the target user vector and each cluster center vector in a preset list, and then the target first sample user associated with the target cluster center vector in the preset list can be determined as a similar user. This allows more users to be covered when similar users are queried in real time to obtain similar users who are more similar to the target user, thereby improving the accuracy of subsequent recommended materials.

[0095] In some embodiments, the material recall method further includes:

[0096] Based on the user features of the acquired plurality of second sample users, filtering the plurality of second sample users to obtain at least two first sample users; the plurality of second sample users are visiting users within a first preset period;

[0097] Inputting the user features of each first sample user into the cross-scenario multi-objective model to obtain a user vector of each first sample user, and determining a cluster center vector based on the user vector of each first sample user;

[0098] Each first sample user is associated with a corresponding cluster center vector to obtain a preset list.

[0099] In an embodiment of the present application, the server can filter the multiple second sample users based on the user features of the acquired multiple second sample users to obtain at least two first sample users, and then input the user features of each first sample user into the cross-scenario multi-target model to obtain the user vector of each first sample user, and determine the cluster center vector based on the user vector of each first sample user, and then associate each first sample user with the corresponding cluster center vector to obtain a preset list, so that similar users of the target user can be determined more quickly in the subsequent process.

[0100] Here, the second sample users may be visiting users collected from the social platform during the first preset period when the server performs offline tasks, and the user features of the second sample users may also be obtained by extracting features from the associated data of the second sample users.

[0101] It should be noted that at least two first sample users may be obtained by performing filtering processing such as removal and screening on a plurality of second sample users.

[0102] It is understandable that the embodiment of the present application can first input the user features of each first sample user into the user tower of the cross-scenario multi-objective model for feature conversion processing to obtain the user vector of each first sample user, and then randomly select k vectors from these user vectors as the initial cluster center, and calculate the distance from each vector to each cluster center, and assign the vector to the cluster center with the closest distance, and then recalculate the center vector of each cluster, that is, the mean of all vectors under the cluster, and finally repeat the above process until the cluster center no longer changes significantly or reaches a preset number of iterations to obtain the cluster center vector in the preset list. After determining the cluster center vector of the preset list, each first sample user is associated with the corresponding cluster center vector, and a complete preset list with an index structure can be obtained.

[0103] In an embodiment of the present application, multiple second sample users can be filtered based on the user features of the acquired multiple second sample users to obtain at least two first sample users, and then the user features of each first sample user are input into the cross-scenario multi-target model to obtain the user vector of each first sample user, and based on the user vector of each first sample user, the cluster center vector is determined, and then each first sample user is respectively associated with the corresponding cluster center vector to obtain a preset list, so that in the subsequent process, similar users of the target user can be determined more quickly through the preset list, thereby improving the efficiency of determining similar users.

[0104] In some embodiments, the filtering process of the plurality of second sample users based on the obtained user features of the plurality of second sample users to obtain at least two first sample users includes:

[0105] Acquire user characteristics of a plurality of second sample users within a first preset period;

[0106] Based on the user characteristics of each second sample user, deduplication processing is performed on the plurality of second sample users to obtain a plurality of third sample users;

[0107] Based on the frequency of occurrence of each third sample user in the second preset period, a plurality of third sample users are screened to obtain at least two first sample users;

[0108] The second preset period is greater than the first preset period.

[0109] In an embodiment of the present application, the server can obtain user characteristics of multiple second sample users within a first preset period, and then deduplicate the multiple second sample users based on the user characteristics of each second sample user to obtain multiple third sample users, and then screen the multiple third sample users based on the frequency of occurrence of each third sample user within the second preset period to obtain at least two first sample users, so that a preset list can be constructed according to the user vectors of the at least two first sample users in a subsequent process.

[0110] Here, the first preset period may be a time period preset by the server, for example, 7 days, 8 days or 10 days, etc. The third sample users may be sample users obtained by deduplicating the second sample users, that is, the number of users of the third sample users is less than that of the second sample users.

[0111] It should be noted that the second preset period may also be a time period preset by the server, and the second preset period is greater than the first preset period, which is not limited in the embodiment of the present application. For example, if the first preset period is 7 days, the second preset period may be 15 days; or, if the first preset period is 8 days, the second preset period may be 30 days, etc.

[0112] It is understandable that the embodiment of the present application can screen a portion of the third sample users in proportion to the frequency of appearance of each third sample user in the second preset period according to the distribution of this frequency, so as to obtain at least two first sample users.

[0113] Exemplarily, the server may execute an offline task once an hour to collect user features from second sample users who have visited an application such as Weibo in the past eight days, and import the collected user features into a distributed task cluster for merging and deduplication to obtain multiple third sample users. The third sample users are then screened in proportion according to the frequency of appearance of each third sample user within 30 days (e.g., the third sample user visits Weibo 7 times, 10 times, 20 times, etc. within 30 days) to obtain at least two first sample users after screening.

[0114] In an embodiment of the present application, user characteristics of multiple second sample users within a first preset period can be obtained, and then based on the user characteristics of each second sample user, the multiple second sample users are deduplicated to obtain multiple third sample users, and then based on the frequency of occurrence of each third sample user within the second preset period, the multiple third sample users are screened to obtain at least two first sample users, so that by filtering the second sample users, the amount of users that need to be processed can be reduced, thereby improving the efficiency of recommending materials to target users.

[0115] In some embodiments, step S330 may include:

[0116] Determine the material in the first historical material that meets the preset playback condition as the second historical material;

[0117] Based on the material vector of the first sample material collected in the target recommendation scenario within a third preset period, the second historical material is filtered for availability to obtain candidate materials; wherein the material vector of the first sample material is obtained by inputting the material characteristics of the first sample material into a cross-scenario multi-objective model.

[0118] In an embodiment of the present application, the server may determine the material in the first historical material that meets the preset playback conditions as the second historical material, and perform availability filtering on the second historical material based on the material vector of the first sample material collected in the target recommendation scenario within a third preset period to obtain candidate materials, so as to determine the recall material of the target user in the subsequent process.

[0119] In some embodiments, the above-mentioned preset playback conditions are satisfied, including: the completion rate is greater than or equal to the preset completion rate threshold, and / or the playback duration is greater than or equal to the preset duration threshold. That is, the material in the first historical material whose completion rate is greater than or equal to the preset completion rate threshold, and / or whose playback duration is greater than or equal to the preset duration threshold is determined as the second historical material.

[0120] Here, the completion rate may refer to the playback time of the material played by the user divided by the video time of the material. The second historical material may be a material in the first historical material that meets the preset playback conditions, that is, for similar users, the completion rate of the second historical material is greater than or equal to the preset completion rate threshold, and / or the playback time is greater than or equal to the preset time threshold.

[0121] It should be noted that the preset completion rate threshold and the preset duration threshold can be set according to the actual application scenario, and the present application embodiment does not limit it. For example, the preset completion rate threshold can be 0.7 or 0.8, and the preset duration threshold can be 8 seconds or 10 seconds.

[0122] It is understandable that the server can eliminate the materials older than 7 days from the second historical materials that meet the preset playback conditions, and then store them as the second historical materials associated with similar users in descending order of timestamps before truncating the preset number, such as 80, so as to be able to determine more accurate candidate materials, thereby improving the accuracy of subsequent material recommendations to target users.

[0123] In the embodiment of the present application, the first sample material may be the latest sample material collected by the server in the target recommendation scenario through an offline task within the third preset period. The material vector of the first sample material may be obtained by inputting the material feature of the first sample material into the material tower corresponding to the target recommendation scenario in the cross-scenario multi-objective model and performing feature conversion processing.

[0124] It should be noted that the third preset period may also be a time period preset by the server, and the third preset period may be the same as or different from the first preset period, which is not limited in the embodiment of the present application. For example, the third preset period may be 20 minutes, 1 hour, or 2 hours.

[0125] It is understandable that the server can perform an offline task every third preset period (such as every 20 minutes) to collect the first sample material in the target recommendation scenario, and input the material characteristics of the first sample material into the cross-scenario multi-objective model to obtain the material vector of the first sample material, and then determine whether the second historical material can be found in the first sample material through the material vector of the first sample material and the associated data of the first sample material, such as the ID of the first sample material, that is, perform availability filtering on the second historical material to obtain candidate materials, which are materials that can be found in the first sample material.

[0126] In an embodiment of the present application, the material in the first historical material that meets the preset playback conditions can be determined as the second historical material, and the second historical material can be filtered for availability based on the material vector of the first sample material collected in the target recommendation scenario within a third preset period to obtain candidate materials. In this way, by filtering the first historical materials of similar users in the target recommendation scenario multiple times, more accurate candidate materials can be obtained to improve the accuracy of subsequent material recommendations to the target user.

[0127] In some embodiments, step S340 may include:

[0128] Inputting the material features of the first sample material collected in the target recommendation scenario into the cross-scenario multi-objective model to obtain the material vector of the first sample material;

[0129] Determine a material vector of a candidate material from the material vector of the first sample material;

[0130] Among the material vectors of the candidate materials, the material corresponding to the material vector whose similarity with the target user vector is greater than a second preset threshold is determined as the recalled material.

[0131] In an embodiment of the present application, the server can input the material characteristics of the first sample material collected in the target recommendation scenario into the cross-scenario multi-target model to obtain the material vector of the first sample material, and then determine the material vector of the candidate material from the material vector of the first sample material, and determine the material corresponding to the material vector in the material vector of the candidate material whose similarity with the target user vector is greater than a second preset threshold as a recalled material, so as to be able to recommend more accurate materials to the target user in the subsequent process.

[0132] Here, the material vector of the first sample material may be a material vector obtained by subjecting the material features of the first sample material to feature conversion processing of the material tower corresponding to the target recommendation scene in the cross-scenario multi-objective model. The material vector of the candidate material may be a material vector determined from the material vector of the first sample material according to the ID of the first sample material.

[0133] It is understandable that the embodiment of the present application can first determine the similarity between the material vector of each candidate material and the target user vector, and then determine the candidate material whose similarity with the target user vector is greater than a second preset threshold as the recalled material of the target user.

[0134] Exemplarily, the inner product, i.e., dot product, of the material vector of each candidate material can be calculated with the target user vector, and the obtained inner product value can be used as the scoring result. The candidate materials can be sorted from high to low according to the scoring result, and the top 1,000 candidate materials can be cut off as recalled materials, i.e., recall results, which are returned to the rough sorting stage, the fine sorting stage, and the rule stage, and then recommended to the target user.

[0135] It should be noted that the second preset threshold may be the same as the first preset threshold, or may be different from the first preset threshold, and this embodiment of the present application does not limit this.

[0136] In an embodiment of the present application, the material characteristics of the first sample material collected in the target recommendation scenario can be input into the cross-scenario multi-target model to obtain the material vector of the first sample material, and then the material vector of the candidate material is determined from the material vector of the first sample material, and the material corresponding to the material vector in the material vector of the candidate material whose similarity with the target user vector is greater than a second preset threshold is determined as a recalled material, thereby effectively improving the accuracy of recommending materials to target users while ensuring personalization.

[0137] In some embodiments, the cross-scenario multi-objective model includes a user tower shared by multiple recommendation scenarios and a material tower corresponding to each recommendation scenario; the material recall method further includes:

[0138] Obtaining user features of multiple fourth sample users and material features of multiple second sample materials; wherein the multiple fourth sample users are visiting users collected in multiple recommendation scenarios, and the multiple second sample materials are materials respectively associated with the multiple fourth sample users in the multiple recommendation scenarios;

[0139] Inputting user features of the plurality of fourth sample users into the user tower in the cross-scenario multi-objective model to obtain user vectors of the plurality of fourth sample users, and inputting material features of the plurality of second sample materials into the material towers corresponding to the recommendation scenarios in the cross-scenario multi-objective model to obtain material vectors of the plurality of second sample materials;

[0140] Based on the similarity between the user vectors of the multiple fourth sample users and the material vectors of the multiple second sample materials, the cross-scene multi-objective model is iteratively trained until the cross-scene multi-objective model converges to obtain a trained cross-scene multi-objective model.

[0141] In an embodiment of the present application, the server can obtain user characteristics of multiple fourth sample users and material characteristics of multiple second sample materials, and then input the user characteristics of the multiple fourth sample users into the user tower in the cross-scene multi-objective model to obtain user vectors of the multiple fourth sample users, and input the material characteristics of the multiple second sample materials into the material towers corresponding to the recommended scenes in the cross-scene multi-objective model to obtain material vectors of the multiple second sample materials, and then iteratively train the cross-scene multi-objective model according to the similarity between the user vectors of the multiple fourth sample users and the material vectors of the multiple second sample materials until the cross-scene multi-objective model converges to obtain the trained cross-scene multi-objective model.

[0142] Here, the plurality of fourth sample users may be sample users collected during the training of the cross-scenario multi-objective model, i.e., visiting users within the training time period of the cross-scenario multi-objective model. The plurality of second sample materials may be materials respectively associated with the plurality of fourth sample users in multiple recommendation scenarios during the training of the cross-scenario multi-objective model; the plurality of second sample materials may include but are not limited to: materials collected in the microblog recommendation scenario and materials collected in the video post-recommendation scenario.

[0143] The user vectors of the plurality of fourth sample users may be user vectors obtained by subjecting the user features of the plurality of fourth sample users to feature conversion processing by the user tower of the cross-scenario multi-objective model. The material vectors of the plurality of second sample materials may be material vectors obtained by subjecting the material features of the plurality of second sample materials to feature conversion processing by the material tower of the corresponding recommendation scene in the cross-scenario multi-objective model.

[0144] It is understandable that the embodiment of the present application can determine the loss value based on the similarity between the user vectors of multiple fourth sample users and the material vectors of multiple second sample materials, and iteratively train the cross-scene multi-objective model based on the loss value until the cross-scene multi-objective model converges to obtain the trained cross-scene multi-objective model. In other words, the loss value is reversely recursively derived from the output layer of the cross-scene multi-objective model to the input layer of the cross-scene multi-objective model to determine the gradient value corresponding to the model parameter; wherein the gradient value is used to characterize the changing relationship between the target loss value and the model parameter; based on the gradient value, the model parameters are updated until the model convergence condition is reached to obtain the trained cross-scene multi-objective model.

[0145] Here, the model convergence conditions may include: the loss value reaches a preset value, or the number of iterations reaches a set number, etc., which is not limited in the embodiments of the present application.

[0146] It should be noted that the gradient value can be the partial derivative of the loss function corresponding to the loss value with respect to the model parameters of the cross-scenario multi-objective model, and the gradient value can represent the rate of change and direction of the loss value in the parameter space. The gradient value can include positive gradient values ​​and negative gradient values; a positive gradient value can represent that increasing the corresponding model parameter value will lead to an increase in the loss value, while a negative gradient value can represent that increasing the corresponding model parameter value will lead to a decrease in the loss value.

[0147] It can be understood that by determining the sign and size of the gradient value, the server can determine how to adjust the model parameters of the cross-scene multi-objective model to reduce the loss value. When the loss value is reduced to a preset value or the number of iterations reaches a set number, the iterative training of the cross-scene multi-objective model can be stopped to improve the accuracy of the trained cross-scene multi-objective model in recommending materials of interest to target users.

[0148] For example, Figure 4 As shown, 41 indicates the user tower shared by multiple recommendation scenarios in the cross-scenario multi-objective model, 42 indicates the first material tower of the cross-scenario multi-objective model, and 43 indicates the second material tower of the cross-scenario multi-objective model. Since most users will not involve only one recommendation scenario, using samples of at least two recommendation scenarios (such as Weibo recommendation scenarios and post-video recommendation scenarios) for training and learning will help to characterize users more three-dimensionally and obtain more comprehensive and complete user training. Among them, the user tower of the cross-scenario multi-objective model is a shared tower, and the materials collected in multiple recommendation scenarios are input into the cross-scenario multi-objective model in real time, so that the cross-scenario multi-objective model can take into account the learning of multiple recommendation scenarios in real time, which can enable the cross-scenario multi-objective model to have a more comprehensive and detailed user characterization capability.

[0149] In an embodiment of the present application, user characteristics of multiple fourth sample users and material characteristics of multiple second sample materials can be obtained, and then the user characteristics of the multiple fourth sample users are input into the user tower in the cross-scene multi-objective model to obtain user vectors of the multiple fourth sample users, and the material characteristics of the multiple second sample materials are respectively input into the material tower corresponding to the recommended scenes in the cross-scene multi-objective model to obtain material vectors of the multiple second sample materials, and then the cross-scene multi-objective model is iteratively trained according to the similarity between the user vectors of the multiple fourth sample users and the material vectors of the multiple second sample materials until the cross-scene multi-objective model converges to obtain a trained cross-scene multi-objective model. Therefore, unlike the related art that only involves a single-scene model, the embodiment of the present application can enhance the ability to learn user vectors through the cross-scene multi-objective model to effectively improve the accuracy of recommending materials to target users.

[0150] In the related technologies, in order to provide users with real-time personalized recommendation information flow services, the traditional recommendation system architecture widely used in the industry is currently adopted, such as Figure 5 As shown, 51 indicates the material warehouse, 52 indicates the recall link, 53 indicates the rough sorting link, 54 indicates the fine sorting link, and 55 indicates the rule link.

[0151] According to the general situation in the industry, the material library 51 used by the traditional recommendation system is dynamically updated, and generally reaches the million level, approximately between 1 million and 5 million; after receiving the user's request, it will go through the recall stage 52 in succession, and obtain nearly 10,000 materials from millions of materials as the recall result. This process will screen out 99% of the material library; then, after the rough sorting stage 53 is screened, 1,000 materials are retained, and then the fine sorting stage 54 scores and sorts the 1,000 materials, and finally the rule stage 55 selects the 10 highest-ranked and strategy-compliant materials from the sorting.

[0152] Among them, multi-channel parallel recall is adopted in the recall stage, and the merged information is returned to the rough sorting stage. After the rough sorting stage, the information is truncated from high to low according to the sorting score and returned to the fine sorting stage. Therefore, the accuracy and diversity of the recall method determine its contribution to the traditional recommendation system. Among the many recall methods, user collaborative recall is usually used to directly recommend materials that similar users like to the current user.

[0153] The user collaborative recall in the related technology is composed of three parts: model, offline task and online service. The model adopts random walk algorithm, and obtains user vector by learning the interaction behavior graph of user ID and material ID in the recommendation scene after video. The offline task uses user vector as input, uses faiss tool to build user index, and then uses user vector to query user index to obtain similar users of the user. The user ID and the user's similar user ID are used as key vector and value vector respectively, and written into remote dictionary server (Remote Dictionary Server, redis) for use by online service. When the online service receives the user request, it uses user ID to query redis to obtain similar user ID, and then uses similar user ID to query user history storage to obtain behavior history of similar users, and uses these behavior histories as recall candidates, and then sorts them from high to low according to the material playback volume and truncates them and returns them to the rough sorting link.

[0154] However, the key to user collaborative recall in related technologies is to find users similar to the current user. Since it is impossible to know in advance which users will visit the recommendation system today, the current method of obtaining similar users based on pre-calculation cannot cover the users who visit today. Moreover, the random walk algorithm involved in this recall method only uses user ID and material ID to learn user vectors. The features are single and not rich, and it is a single-scene model with poor learning ability, and it is impossible to obtain more accurate user vectors. In addition, this recall method uses the material playback volume to sort the recall candidate set, which lacks personalized considerations, resulting in low accuracy in recommending materials for the current user.

[0155] Based on this, the implementation process of the material recall method of the above embodiment in an actual application scenario is specifically described below. Figure 6 This is a schematic diagram of a material recall method provided in an embodiment of the present application. Figure 2 ,like Figure 6 As shown, the material recall method provided in the embodiment of the present application can be applied to electronic devices such as servers, and may include the following steps:

[0156] S601: In response to a material recommendation request from a target user, obtain target user characteristics of the target user and a target recommendation scenario indicated by the material recommendation request.

[0157] Here, the target user may be the object of the server to recommend materials. The material recommendation request may be a material recommendation request sent by the target user to the server by refreshing a recommendation page or clicking a recommendation button on a terminal device.

[0158] It should be noted that the target recommendation scenarios indicated by the above material recommendation request may include but are not limited to: Weibo recommendation scenarios and video / blog post recommendation scenarios. Among them, Weibo recommendation scenarios may refer to scenarios where blog posts and / or videos are recommended to users on the Weibo recommendation page; video post recommendation scenarios may refer to scenarios where videos are recommended to users when users continue to scroll down after playing videos from other scenarios; and blog post recommendation scenarios may refer to scenarios where users continue to scroll down after clicking on a blog post and other blog posts are recommended to users.

[0159] S602: Input the target user features into the trained cross-scenario multi-target model to obtain the corresponding target user vector.

[0160] Here, the cross-scenario multi-objective model may include a user tower shared by multiple recommendation scenarios and a material tower corresponding to each recommendation scenario, that is, a material tower may correspond to a material collected in a recommendation scenario. The cross-scenario multi-objective model can be obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios; wherein, the training samples of multiple recommendation scenarios may include the visiting users collected in multiple recommendation scenarios and the materials associated with the visiting users, and the task objectives may include but are not limited to: click-through rate, completion rate or playback time, etc. The target user vector may be a user vector obtained after the target user features of the target user are converted through the user tower of the cross-scenario multi-objective model.

[0161] S603: Determine a target cluster center vector from the cluster center vectors based on the similarity between the target user vector and each cluster center vector in the preset list.

[0162] Here, the cluster center vector with the highest similarity to the target user vector in the preset list can be determined as the target cluster center vector; or, the cluster center vector with a similarity to the target user vector in the preset list greater than a fourth preset threshold can be determined as the target cluster center vector.

[0163] S604: Determine the target first sample user associated with the target cluster center vector in the preset list as a similar user.

[0164] Here, the similarity between the user training of similar users and the target user vector is greater than the third preset threshold; the similarity between the user vector of the target first sample user and the target cluster center vector is greater than the first preset threshold; the third preset threshold can be the same as or different from the first preset threshold.

[0165] S605: Determine, as the second historical material, materials whose completion rate is greater than or equal to a preset completion rate threshold and / or whose playing time is greater than or equal to a preset time threshold in the first historical materials of the similar user.

[0166] Here, the first historical material is the material associated with similar users in the target recommendation scenario indicated by the material recommendation request; the second historical material can be the material in the first historical material that meets the preset playback conditions, that is, for similar users, the completion rate of the second historical material is greater than or equal to the preset completion rate threshold, and / or the playback duration is greater than or equal to the preset duration threshold.

[0167] S606: Based on the material vector of the first sample material collected in the target recommendation scenario within the third preset period, the second historical material is filtered for availability to obtain a candidate material.

[0168] Here, the server can execute an offline task every third preset period to collect the first sample material in the target recommendation scenario, and input the material characteristics of the first sample material into the cross-scenario multi-objective model to obtain the material vector of the first sample material, and then determine whether the second historical material can be found in the first sample material through the material vector of the first sample material and the associated data of the first sample material, such as the ID of the first sample material, that is, perform availability filtering on the second historical material to obtain a candidate material, which is a material that can be found in the first sample material.

[0169] S607: Determine the material vector of the candidate material from the material vector of the first sample material.

[0170] Here, the material vector of the first sample material may be a material vector obtained by subjecting the material features of the first sample material to feature conversion processing of the material tower corresponding to the target recommendation scene in the cross-scenario multi-objective model. The material vector of the candidate material may be a material vector determined from the material vector of the first sample material according to the ID of the first sample material.

[0171] S608: Determine, as a recalled material, a material corresponding to a material vector whose similarity with the target user vector is greater than a second preset threshold among the material vectors of the candidate materials.

[0172] Here, the similarity between the material vector of each candidate material and the target user vector may be determined first, and then the candidate material having a similarity with the target user vector greater than a second preset threshold may be determined as the recalled material of the target user.

[0173] In some embodiments, the recalled materials are the recall results obtained by collaborative recall by users. After the recalled materials are determined, they still need to go through a rough sorting stage, a fine sorting stage and a rule stage to obtain the materials recommended to the target users.

[0174] The material recall method provided in the embodiment of the present application can, when detecting the material recommendation request of the target user, determine the target user vector obtained by inputting the cross-scenario multi-target model through the target user feature of the target user, and filter out the candidate materials from the first historical materials associated with the similar users in the target recommendation scenario indicated by the material recommendation request, and then determine the recall material of the target user from the candidate materials according to the target user vector and the material vector of the candidate material. In this way, unlike the related art of directly recommending materials liked by users similar to the current user, the embodiment of the present application can obtain the recall material of the target user by filtering multiple times from the first historical materials associated with the similar users after determining the similar users of the target user, so as to effectively improve the accuracy of recommending materials to the target user while ensuring personalization; and by screening out the candidate materials from the first historical materials associated with the similar users in the recommendation scenario, the candidate materials can be determined more accurately, thereby improving the matching degree between the recall material and the target user.

[0175] The following is a description of an exemplary structure of the material recall device 255 provided in the application embodiment implemented as a software module. In some embodiments, for example, Figure 2 As shown, the software modules stored in the material recall device 255 of the memory 250 may include:

[0176] An acquisition module 2551 is configured to, in response to a material recommendation request of a target user, acquire a target user feature of the target user and a target recommendation scenario indicated by the material recommendation request;

[0177] The first determination module 2552 is configured to input the target user feature into the trained cross-scenario multi-objective model to obtain a corresponding target user vector, and determine similar users of the target user based on the target user vector; wherein the cross-scenario multi-objective model is obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios;

[0178] The processing module 2553 is configured to perform availability filtering on the first historical materials of the similar users to obtain candidate materials; wherein the first historical materials are materials associated with the similar users in the target recommendation scenario;

[0179] The second determination module 2554 is configured to determine the recall material of the target user from the candidate materials based on the target user vector and the material vector of the candidate materials.

[0180] In some embodiments, the first determination module 2552 is specifically configured as follows: based on the similarity between the target user vector and each cluster center vector in a preset list, determining the target cluster center vector from the cluster center vectors; wherein the preset list is constructed by inputting the user characteristics of the first sample user within the first preset period into the cross-scenario multi-target model to obtain the corresponding user vector; the target first sample user associated with the target cluster center vector in the preset list is determined as a similar user; wherein the similarity between the user vector of the target first sample user and the target cluster center vector is greater than a first preset threshold.

[0181] In some embodiments, the processing module 2553 is also configured to: filter the multiple second sample users based on the user features of the acquired multiple second sample users to obtain at least two first sample users; the multiple second sample users are visiting users within a first preset period; input the user features of each first sample user into the cross-scenario multi-objective model to obtain the user vector of each first sample user, and determine the cluster center vector based on the user vector of each first sample user; and associate each first sample user with the corresponding cluster center vector to obtain a preset list.

[0182] In some embodiments, the processing module 2553 is also configured to: obtain user characteristics of multiple second sample users within a first preset period; deduplicate the multiple second sample users based on the user characteristics of each second sample user to obtain multiple third sample users; screen the multiple third sample users based on the frequency of occurrence of each third sample user within the second preset period to obtain at least two first sample users; wherein the second preset period is greater than the first preset period.

[0183] In some embodiments, the processing module 2553 is specifically configured as follows: determining the material in the first historical material that meets the preset playback conditions as the second historical material; based on the material vector of the first sample material collected in the target recommendation scenario within a third preset period, filtering the second historical material for availability to obtain candidate materials; wherein the material vector of the first sample material is obtained by inputting the material characteristics of the first sample material into a cross-scenario multi-objective model.

[0184] In some embodiments, satisfying a preset playback condition includes: a completion rate being greater than or equal to a preset completion rate threshold, and / or a playback duration being greater than or equal to a preset duration threshold.

[0185] In some embodiments, the second determination module 2554 is specifically configured as follows: inputting the material characteristics of the first sample material collected in the target recommendation scenario into the cross-scenario multi-target model to obtain the material vector of the first sample material; determining the material vector of the candidate material from the material vector of the first sample material; and determining the material corresponding to the material vector of the candidate material whose similarity with the target user vector is greater than a second preset threshold as a recalled material.

[0186] In some embodiments, the processing module 2553 is further configured to: obtain user characteristics of multiple fourth sample users and material characteristics of multiple second sample materials; wherein the multiple fourth sample users are visiting users collected in multiple recommendation scenarios, and the multiple second sample materials are materials respectively associated with the multiple fourth sample users in multiple recommendation scenarios; input the user characteristics of the multiple fourth sample users into the user tower in the cross-scenario multi-objective model to obtain user vectors of the multiple fourth sample users, and input the material characteristics of the multiple second sample materials into the material towers corresponding to the recommendation scenarios in the cross-scenario multi-objective model to obtain material vectors of the multiple second sample materials; based on the similarity between the user vectors of the multiple fourth sample users and the material vectors of the multiple second sample materials, iteratively train the cross-scenario multi-objective model until the cross-scenario multi-objective model converges to obtain the trained cross-scenario multi-objective model.

[0187] It should be noted that the description of the material recall device in the embodiment of the present application is similar to the description of the above-mentioned material recall method embodiment, and has similar beneficial effects as the method embodiment, so it will not be repeated.

[0188] The embodiment of the present application provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the material recall method described in the embodiment of the present application.

[0189] The present application embodiment provides a computer-readable storage medium storing executable instructions, wherein a computer program or executable instructions are stored. When the computer program or executable instructions are executed by a processor, the processor will be caused to execute the material recall method provided in the present application embodiment, for example, Figure 3 Material recall method shown.

[0190] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0191] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0192] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0193] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0194] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A material recall method, characterized in that: The method comprises: In response to a material recommendation request of a target user, obtaining a target user feature of the target user and a target recommendation scenario indicated by the material recommendation request; Input the target user features into the trained cross-scenario multi-objective model to obtain a corresponding target user vector, and determine similar users of the target user based on the target user vector; wherein the cross-scenario multi-objective model is obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios; Availability filtering is performed on the first historical material of the similar user to obtain candidate materials; wherein the first historical material is a material associated with the similar user in the target recommendation scenario; Based on the target user vector and the material vectors of the candidate materials, the recalled materials for the target user are determined from the candidate materials.

2. The method according to claim 1, characterized in that The determining similar users of the target user based on the target user vector includes: Based on the similarity between the target user vector and each cluster center vector in the preset list, a target cluster center vector is determined from the cluster center vectors; wherein the preset list is constructed by inputting the user features of the first sample user in the first preset period into the cross-scenario multi-objective model to obtain the corresponding user vector; The target first sample user associated with the target cluster center vector in the preset list is determined as the similar user; wherein the similarity between the user vector of the target first sample user and the target cluster center vector is greater than a first preset threshold.

3. The method according to claim 2, characterized in that The method further comprises: Based on the user features of the acquired multiple second sample users, filtering the multiple second sample users to obtain at least two of the first sample users; the multiple second sample users are visiting users within the first preset period; Inputting the user features of each first sample user into the cross-scenario multi-objective model to obtain a user vector of each first sample user, and determining the cluster center vector based on the user vector of each first sample user; Each first sample user is associated with a corresponding cluster center vector to obtain the preset list.

4. The method according to claim 3, characterized in that The filtering the plurality of second sample users based on the acquired user features to obtain at least two of the first sample users includes: Acquire user characteristics of a plurality of second sample users within the first preset period; Based on the user characteristics of each second sample user, performing deduplication processing on the plurality of second sample users to obtain a plurality of third sample users; Based on the frequency of occurrence of each third sample user within the second preset period, the plurality of third sample users are screened to obtain at least two of the first sample users; Wherein, the second preset period is greater than the first preset period.

5. The method according to any one of claims 1 to 4, characterized in that: The filtering of the first historical materials of the similar user to obtain candidate materials includes: Determine the material in the first historical material that meets the preset playback condition as the second historical material; Based on the material vector of the first sample material collected in the target recommendation scenario within a third preset period, the second historical material is filtered for availability to obtain the candidate material; wherein the material vector of the first sample material is obtained by inputting the material characteristics of the first sample material into the cross-scenario multi-objective model.

6. The method according to claim 5, characterized in that The preset playback condition is met, including: The completion rate is greater than or equal to a preset completion rate threshold, and / or the playback duration is greater than or equal to a preset duration threshold.

7. The method according to any one of claims 1 to 4, characterized in that: The determining the recalled material of the target user from the candidate materials based on the target user vector and the material vector of the candidate materials includes: Inputting the material characteristics of the first sample material collected in the target recommendation scenario into the cross-scenario multi-objective model to obtain the material vector of the first sample material; Determine the material vector of the candidate material from the material vector of the first sample material; The material corresponding to the material vector whose similarity with the target user vector among the material vectors of the candidate materials is greater than a second preset threshold is determined as the recalled material.

8. The method according to any one of claims 1 to 4, characterized in that: The cross-scenario multi-objective model includes a user tower shared by multiple recommendation scenarios and a material tower corresponding to each recommendation scenario; the method further includes: Obtaining user features of a plurality of fourth sample users and material features of a plurality of second sample materials; wherein the plurality of fourth sample users are visiting users collected under the plurality of recommendation scenarios, and the plurality of second sample materials are materials respectively associated with the plurality of fourth sample users under the plurality of recommendation scenarios; Inputting user features of the plurality of fourth sample users into the user tower in the cross-scenario multi-objective model to obtain user vectors of the plurality of fourth sample users, and inputting material features of the plurality of second sample materials into material towers corresponding to the recommended scenarios in the cross-scenario multi-objective model to obtain material vectors of the plurality of second sample materials; Based on the similarity between the user vectors of the multiple fourth sample users and the material vectors of the multiple second sample materials, the cross-scene multi-objective model is iteratively trained until the cross-scene multi-objective model converges to obtain a trained cross-scene multi-objective model.

9. A material recall device, characterized in that: include: an acquisition module, configured to, in response to a material recommendation request of a target user, acquire a target user feature of the target user and a target recommendation scenario indicated by the material recommendation request; A first determination module is configured to input the target user feature into the trained cross-scenario multi-objective model to obtain a corresponding target user vector, and determine similar users of the target user based on the target user vector; wherein the cross-scenario multi-objective model is obtained by learning and training multiple task objectives through training samples of multiple recommendation scenarios; A processing module, configured to perform availability filtering on the first historical material of the similar user to obtain a candidate material; wherein the first historical material is a material associated with the similar user in the target recommendation scenario; The second determination module is configured to determine the recalled material of the target user from the candidate materials based on the target user vector and the material vector of the candidate material.

10. An electronic device, characterized in that: include: Memory for storing computer programs or executable instructions; A processor, configured to implement the method according to any one of claims 1 to 8 when executing the computer program or the executable instructions stored in the memory.

Images