Recommended processing method, apparatus, device, storage medium and program product

Abstract

This application provides a recommendation processing method, apparatus, electronic device, computer-readable storage medium, and computer program product based on artificial intelligence. The method includes: acquiring the interaction features and object features of a target account; performing first attention processing on the interaction features based on a first object interest pool to obtain interaction interest features, and performing second attention processing on the object features based on a second object interest pool to obtain object interest features; acquiring the activity features of the target account, and performing selection processing on the interaction interest features and object interest features based on the activity features to obtain recall features of the target account; based on the recall features and the information features of the information to be recommended, obtaining target information from the information to be recommended, and performing a recommendation operation on the target account based on the target information. This application enables accurate personalized recommendations and improves recommendation processing efficiency.

Description

Technical Field

[0001] This application relates to artificial intelligence technology, and more particularly to a recommendation processing method, apparatus, electronic device, computer-readable storage medium, and computer program product based on artificial intelligence. Background Technology

[0002] Artificial intelligence (AI) is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0003] Recommendation processing is a crucial application of artificial intelligence. Recall, as the front end of a recommendation system, determines the upper and lower limits of the entire system. With the development of deep learning, large-scale labeled deep learning networks have been widely adopted and applied at various stages of recommendation systems. Related technologies rely on recalling and recommending to users based on information sequences related to user interactions. However, for users with sparse interactions, it's difficult to accurately characterize their data vectors, making it challenging to effectively learn their interest features. Information recommended in this way cannot effectively match user interests, resulting in a poor user experience. Summary of the Invention

[0004] This application provides an artificial intelligence-based recommendation processing method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can map interaction features and object features to different interest pools, and filter recall features based on activity to more accurately predict target information, thereby achieving precise personalized recommendations and improving recommendation processing efficiency.

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

[0006] This application provides an artificial intelligence-based recommendation processing method, including:

[0007] Obtain at least one interaction characteristic and at least one object characteristic of the target object's account;

[0008] Based on the first object interest pool, the at least one interactive feature is subjected to a first attention processing to obtain at least one interactive interest feature, and based on the second object interest pool, the at least one object feature is subjected to a second attention processing to obtain at least one object interest feature.

[0009] Among them, the activity level of the objects corresponding to the first object interest pool is higher than that of the objects corresponding to the second object interest pool.

[0010] The activity characteristics of the target account are obtained, and the at least one interactive interest characteristic and the at least one object interest characteristic are selected based on the activity characteristics to obtain at least one recall characteristic of the target account.

[0011] Based on the at least one recall feature and the information features of multiple pieces of information to be recommended, target information is obtained from the multiple pieces of information to be recommended, and a recommendation operation is performed on the target object account based on the target information.

[0012] This application provides an artificial intelligence-based recommendation processing device, comprising:

[0013] The acquisition module is used to acquire at least one interaction feature and at least one object feature of the target object's account;

[0014] An attention module is used to perform a first attention process on the at least one interactive feature based on a first object interest pool to obtain at least one interactive interest feature, and to perform a second attention process on the at least one object feature based on a second object interest pool to obtain at least one object interest feature.

[0015] Among them, the activity level of the objects corresponding to the first object interest pool is higher than that of the objects corresponding to the second object interest pool.

[0016] The selection module is used to obtain the activity characteristics of the target object account, and to select the at least one interactive interest characteristic and the at least one object interest characteristic based on the activity characteristics to obtain at least one recall characteristic of the target object account.

[0017] The recommendation module is used to obtain target information from the multiple pieces of information to be recommended based on the at least one recall feature and the information features of multiple pieces of information to be recommended, and to perform a recommendation operation on the target object account based on the target information.

[0018] In the above scheme, the attention module is further configured to: obtain at least one first interest feature from the first object interest pool; for each interaction feature, obtain a target first interest feature corresponding to the interaction feature from the at least one first interest feature; and for each target first interest feature, perform aggregation processing on the target interaction feature corresponding to the target first interest feature to obtain an interaction interest feature corresponding to the target first interest feature.

[0019] In the above scheme, the attention module is further configured to: perform a first mapping process on each first interest feature to obtain a first key feature corresponding to each first interest feature; perform a second mapping process on the interaction feature to obtain a first query feature corresponding to the interaction feature; determine a first weight corresponding to each first interest feature based on the first query feature corresponding to the interaction feature and the first key feature corresponding to each first interest feature; and determine the first interest feature with the highest first weight as the target first interest feature corresponding to the interaction feature.

[0020] In the above scheme, the attention module is further configured to: perform a third mapping process on the target first interest feature to obtain a second query feature corresponding to the target first interest feature; perform a fourth mapping process on each target interaction feature to obtain a second key feature corresponding to each target interaction feature; obtain a first value feature corresponding to each target interaction feature; determine a second weight corresponding to each target interaction feature based on the second query feature corresponding to the target first interest feature and the second key feature corresponding to each target interaction feature; and perform a weighted summation process on the first value feature corresponding to each target interaction feature based on the second weight corresponding to each target interaction feature to obtain an interaction interest feature corresponding to the target first interest feature.

[0021] In the above scheme, the attention module is further configured to: obtain at least one second interest feature from the second object interest pool; for each object feature, obtain a target second interest feature corresponding to the object feature from the at least one second interest feature; and for each target second interest feature, perform aggregation processing on the target object feature corresponding to the target second interest feature to obtain an object interest feature corresponding to the target second interest feature.

[0022] In the above scheme, the attention module is further configured to: perform a fifth mapping process on each second interest feature to obtain a third key feature corresponding to each second interest feature; perform a sixth mapping process on the object feature to obtain a third query feature corresponding to the object feature; determine a third weight corresponding to each second interest feature based on the third query feature corresponding to the object feature and the third key feature corresponding to each second interest feature; and determine the second interest feature with the highest third weight as the target second interest feature corresponding to the object feature.

[0023] In the above scheme, the attention module is further configured to: perform a seventh mapping process on the target second interest feature to obtain a fourth query feature corresponding to the target second interest feature; perform an eighth mapping process on each target object feature to obtain a fourth key feature corresponding to each target object feature; obtain a second value feature corresponding to each target object feature; determine a fourth weight corresponding to each target object feature based on the fourth query feature corresponding to the target second interest feature and the fourth key feature corresponding to each target object feature; and perform a weighted summation process on the second value feature corresponding to each target object feature based on the fourth weight corresponding to each target object feature to obtain an object interest feature corresponding to the target second interest feature.

[0024] In the above scheme, the selection module is further configured to: perform a ninth mapping process on the activity feature to obtain a fifth query feature of the activity feature; perform a tenth mapping process on each interaction interest feature to obtain a fifth key feature corresponding to each interaction interest feature, and perform a tenth mapping process on each object interest feature to obtain a fifth key feature corresponding to each object interest feature; determine a fifth weight corresponding to each interaction interest feature and a fifth weight corresponding to each object interest feature based on the fifth query feature corresponding to the activity feature, the fifth key feature corresponding to each interaction interest feature, and the fifth key feature corresponding to each object interest feature; perform a mixed descending sorting process on the at least one interaction interest feature and the at least one object interest feature based on the fifth weight; and determine the at least one feature with the highest ranking among the at least one interaction interest feature and the at least one object interest feature as the recall feature of the target object account.

[0025] In the above scheme, the information features are obtained through an information tower network, which includes N cascaded fully connected layers, where N is an integer greater than or equal to 2. The recommendation module is further configured to: before obtaining target information from the multiple pieces of information to be recommended based on the at least one recall feature and the information features of the multiple pieces of information to be recommended, perform the following processing for each piece of information to be recommended: perform full connection processing on the input of the nth fully connected layer through the nth fully connected layer in the N cascaded fully connected layers to obtain the nth fully connected result; transmit the nth fully connected result to the (n+1)th fully connected layer for further full connection processing; where n is an integer starting from 1 and increasing in value, and the value range of n satisfies 1≤n≤N-1; when n is 1, the input of the nth fully connected layer is the attribute information of the information to be recommended; when n is 2≤n≤N-1, the input of the nth fully connected layer is the full connection result of the (n-1)th fully connected layer, and the output of the nth fully connected layer is the information feature.

[0026] In the above scheme, the recommendation module is further configured to: perform the following processing for each recall feature: determine the similarity between the recall feature and the information features of each piece of information to be recommended; based on the similarity of each piece of information to be recommended, sort the multiple pieces of information to be recommended in descending order, and determine at least one piece of information to be recommended that ranks first as the target information for the recall feature.

[0027] In the above scheme, the attention processing and selection processing are implemented through an object tower network, the information features of the information to be recommended are obtained through an information tower network, and the processing of acquiring target information is implemented through a similarity network. The multi-interest recall model is composed of an object tower network, an information tower network, and a similarity network. The device also includes a training module, used for: forward propagating at least one interaction feature and at least one object feature of a sample object account in the object tower network to obtain the recall feature of the sample object account; forward propagating multiple sample information to be recommended in the information tower network to obtain the information features of each sample information to be recommended; determining the target sample information to be recommended corresponding to each recall feature through the similarity network; and performing the following processing for each recall feature: determining the loss corresponding to the recall feature based on the predicted similarity and pre-labeled similarity of the target sample information to be recommended corresponding to the recall feature; and updating the parameters of the multi-interest recall model based on the minimum value of the loss corresponding to at least one recall feature.

[0028] This application provides an electronic device, including:

[0029] Memory, used to store executable instructions;

[0030] When the processor executes the executable instructions stored in the memory, it implements the AI-based recommended processing method provided in the embodiments of this application.

[0031] This application provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the AI-based recommendation processing method provided in this application.

[0032] The embodiments of this application have the following beneficial effects:

[0033] The interaction features are processed by first attention based on the first object interest pool to obtain interaction interest features, and the object features are processed by second attention based on the second object interest pool to obtain object interest features. This allows different types of features to be mapped to interest pools of different types of objects. The interaction interest features and object interest features are selected based on activity features to obtain recall features of the target object account. The interest features are then filtered using activity features as a threshold. The selected recall features can be used to represent the interest distribution of the target object account. Subsequently, recommendation processing is performed based on the target information of the corresponding recall features, which can effectively improve recommendation efficiency and accuracy. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the structure of an artificial intelligence-based recommendation system provided in an embodiment of this application;

[0035] Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application;

[0036] Figures 3A-3C This is a flowchart illustrating the AI-based recommendation method provided in an embodiment of this application;

[0037] Figure 4 This is a schematic diagram of the recommended model provided in the embodiments of this application;

[0038] Figure 5 This is an attention diagram of the AI-based recommendation method provided in the embodiments of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is 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.

[0041] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0042] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0044] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0045] 1) Recommendation system: A recommendation system is a tool that automatically connects users with information. It can help users discover information that interests them in an information overload environment, and can also push information to users who are interested in it.

[0046] 2) Recall: Due to limitations in computing power and online system latency (rt), recommendation systems typically employ a funnel-like structure of recall-coarse ranking (optional)-fine ranking-strategy (mixed ranking). Recall is at the front end of the system, responsible for selecting a subset (hundreds, thousands to tens of thousands) from the entire candidate pool (millions to hundreds of millions) that meets the system's objectives and computing power constraints. This ensures the lower limit of the recommendation system and directly affects the upper limit of its performance.

[0047] 3) Attention mechanism: It is essentially derived from the human visual attention mechanism. When people perceive things visually, they do not usually look at the whole thing from beginning to end every time. Instead, they often observe and pay attention to a specific part according to their needs. Moreover, when people find that the things they want to observe often appear in a certain part of a scene, they will learn to focus their attention on that part when similar scenes appear in the future, thus focusing more attention on the useful parts.

[0048] 4) Target account, i.e., the target of recommendation processing. Since the medium for information presentation is the terminal, and the target of recommendation processing is the user operating the corresponding terminal, "object" and "user" are described as equivalent in the following text. Understandably, the user here can be a natural person capable of operating the terminal, or a robot program running on the terminal that can simulate a human.

[0049] 5) Information to be recommended: This refers to information that can be sent to the terminal for presentation and recommendation to the corresponding terminal user (target account), such as videos, products, news, etc.

[0050] In related technologies, the deep networks used in the recall phase of recommendation systems are typically multi-interest recall models trained on user interaction information sequences. These networks are highly dependent on the features of user interaction information sequences and cannot accurately characterize the interest vectors of users with sparse in-domain interaction information, thus failing to accurately characterize the interest vectors of cold-start samples with limited user interaction information. A common recall method for these users is to formulate rules for recall based on cold-start objects (e.g., similar objects, popular items, followed items, etc.) as a supplement. These rules have some effect, but the entire system lacks end-to-end unified optimization, and non-data-driven modeling results in a loss of accuracy, affecting the final recommendation efficiency. In summary, by only considering the interest modeling of users with abundant interaction information, the aforementioned multi-interest recall model degenerates into a model that only processes object information for cold-start users who have just entered the recommendation system and have little interaction information. Furthermore, related technologies mainly rely on individual strategies to process cold-start object accounts rather than using a data-driven approach, leading to poor recommendation performance and results.

[0051] This application provides an artificial intelligence-based recommendation processing method, apparatus, electronic device, computer-readable storage medium, and computer program product. It can map interaction features and object features to different interest pools and filter recall features based on activity levels to more accurately predict target information, thereby achieving precise personalized recommendations and improving recommendation processing efficiency. The following describes exemplary applications of the electronic device provided in this application embodiment, which can be a server. Exemplary applications when the electronic device is implemented as a server will be described below.

[0052] See Figure 1 , Figure 1This is a schematic diagram of the structure of an AI-based recommendation processing system provided in this application embodiment. To support a news application, terminal 400 connects to application server 200-2 via network 300. Network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of both. Training server 200-1 pushes the trained recommendation model to application server 200-2. Terminal 400 sends a user request to application server 200-2. Application server 200-2 obtains the interaction features and object features of the target account. Based on a first object interest pool, it performs first attention processing on the interaction features to obtain interaction interest features, and performs second attention processing on the object features based on a second object interest pool to obtain object interest features. It also obtains the activity features of the target account and performs selection processing on the interaction interest features and object interest features based on the activity features to obtain the recall features of the target account. Based on the recall features and the information features of multiple pieces of information to be recommended, it obtains target information from multiple pieces of information to be recommended, performs fine-grained ranking processing on the target information, and returns the ranking result to the terminal 400 corresponding to the target account for presentation.

[0053] In some embodiments, the training server 200-1 and the application server 200-2 can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal 400 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, and this embodiment does not impose any limitations.

[0054] In some embodiments, the terminal or server can implement the AI-based recommendation processing method provided in this application by running a computer program. For example, the computer program can be a native program or software module in an operating system; it can be a native application (APP), i.e., a program that needs to be installed in the operating system to run, such as a news APP or an e-commerce APP; it can also be a mini-program, i.e., a program that only needs to be downloaded to a browser environment to run; or it can be a mini-program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module, or plugin.

[0055] See Figure 2 , Figure 2This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Taking the electronic device as an application server 200-2 as an example, Figure 2 The application server 200-2 shown includes at least one processor 210, memory 250, at least one network interface 220, and a user interface 230. The various components in terminal 400 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 2 The general labeled all buses as Bus System 240.

[0056] 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. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0057] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0058] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in the embodiments of this application is intended to include any suitable type of memory.

[0059] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0060] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0061] The network communication module 252 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220, such as Bluetooth, WiFi, and Universal Serial Bus (USB).

[0062] In some embodiments, the AI-based recommendation processing apparatus provided in this application can be implemented in software. Figure 2 An AI-based recommendation processing device 255 stored in memory 250 is shown. It can be software in the form of programs and plug-ins, including the following software modules: acquisition module 2551, attention module 2552, selection module 2553, recommendation module 2554, and training module 2555. These modules are logically related and can therefore be arbitrarily combined or further divided according to the functions they implement. The functions of each module will be described below.

[0063] The information recommendation method based on artificial intelligence provided in this application embodiment will be described in conjunction with the exemplary application and implementation of the application server 200-2 provided in the embodiments of this application.

[0064] In some embodiments, see Figure 4 , Figure 4 This is a schematic diagram of the recommendation model provided in the embodiments of this application. The multi-interest recall model includes an object tower network, an information tower network, and a similarity network. The input of the object tower network is interaction information and object information, and the output of the object tower network is at least one recall feature of the object account. The input of the information tower network is relevant data of the information to be recommended (including identifier, type, attribute data, etc.), and the output of the information tower network is the information features of the information to be recommended. The input of the similarity network is at least one recall feature and information features that correspond one-to-one with multiple pieces of information to be recommended, and the output of the similarity network is the target information corresponding to each recall feature among the multiple pieces of information to be recommended.

[0065] See Figure 3A , Figure 3A This is a flowchart illustrating the AI-based recommendation processing method provided in the embodiments of this application, which will be combined with... Figure 3A Steps 101 to 104 are described below.

[0066] In step 101, at least one interaction feature and at least one object feature of the target object account are obtained.

[0067] As an example, the target object account is the account of the recommended user, such as user A's account. The recommendation processing method provided in this application aims to recommend target information that matches user A's interests to user A's account. The input of the object tower network is interaction information and object information. The interaction information includes a sequence of information that has interacted with the user. Taking video as an example, the interaction information includes at least one of the following: videos that the user has watched, videos that the user has clicked, videos that the user has favorited, videos that the user has commented on, videos that the user has forwarded, etc. The object information includes at least one of the following: the user's age group, the user's gender, etc. The interaction feature is a dense vector obtained by embedding and compressing discrete interaction information. For example, if the user has watched video A, then the identifier of video A is embedded and compressed to obtain an interaction feature. If the user's age group is 10 to 20 years old, then the data of the 10 to 20-year-old age group is embedded and compressed to obtain an object feature.

[0068] It is understood that in the embodiments of this application, data related to the user's age group, gender, etc. are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0069] In step 102, at least one interactive feature is subjected to first attention processing based on the first object interest pool to obtain at least one interactive interest feature, and at least one object feature is subjected to second attention processing based on the second object interest pool to obtain at least one object interest feature.

[0070] As an example, the activity level of objects in the first object interest pool is higher than that of objects in the second object interest pool. The first object interest pool is generated based on objects of the first type, and the second object interest pool is generated based on objects of the second type. Objects of the first type are active users, and objects of the second type are cold-start users. The activity level of active users is higher than that of the first type of active users, and the activity level of cold-start users is lower than that of the second type of active users. The first type of active users is higher than that of the second type of active users. The first interest vectors representing different directions in the first object interest pool are generated through a multi-interest network with dynamic routing (MIND) model. The input of the multi-interest network with dynamic routing model is object data and interaction data of many active users, and the output of the multi-interest network with dynamic routing model is the first interest vector representing different directions. The second interest vectors representing different directions in the second object interest pool are also generated through MIND. The input of the multi-interest network with dynamic routing model is object data and interaction data of many cold-start users, and the output of the multi-interest network with dynamic routing model is the second interest vector representing different directions.

[0071] In step 103, the activity characteristics of the target account are obtained, and at least one interactive interest characteristic and at least one object interest characteristic are selected based on the activity characteristics to obtain at least one recall characteristic of the target account.

[0072] As an example, user activity characteristics can be differentiated according to different scenarios. For example, a user who clicks less than k videos in 7 days is an inactive user for 7 days. This application embodiment adopts a gating mechanism to divide activity into multiple levels. For example, the target account is inactive for 1 day (a user who clicks less than k videos in 1 day is an inactive user for 1 day), the target account is inactive for 3 days, the target account is inactive for 5 days, the target account is inactive for 7 days, and the target account is inactive for 14 days, thus mapping user activity to activity characteristics.

[0073] In step 104, based on at least one recall feature and the information features of multiple pieces of information to be recommended, target information is obtained from the multiple pieces of information to be recommended, and a recommendation operation is performed on the target object account based on the target information.

[0074] As an example, for each recall feature, the target information to be recalled is obtained. For example, for recall feature A, the similarity between the information features of each piece of information to be recommended and recall feature A is calculated. The similarity between features can be measured by the distance between features (in vector form). Based on the similarity, multiple pieces of information to be recommended are sorted in descending order, so that the multiple pieces of information to be recommended at the top of the sort are used as the target information to be recalled for recall feature A.

[0075] In some embodiments, see Figure 3B , Figure 3B This is a flowchart illustrating the AI-based recommendation processing method provided in this application embodiment. In step 102, at least one interactive feature is subjected to first attention processing based on a first object interest pool to obtain at least one interactive interest feature, which can be obtained through... Figure 3B Steps 1021 to 1023 shown are implemented.

[0076] In step 1021, at least one first interest feature is obtained from the first object interest pool.

[0077] As an example, the first object interest pool is generated based on a first type of object, which consists of active users. The activity level of active users is higher than the first activity level. The first interest vectors representing different directions in the first object interest pool are generated using a Multi-Interest Network with Dynamic Routing (MIND) model. The input to the MIND model is object data and interaction data from many active users, and the output is the first interest vectors representing different directions. The vector dimension of the first object interest pool is K*D, representing that the first object interest pool includes K first interest features, each of which represents a certain direction (type) of interest in the first type of object.

[0078] In step 1022, for each interaction feature, the target first interest feature corresponding to the interaction feature is obtained from at least one first interest feature.

[0079] In some embodiments, obtaining the target first interest feature corresponding to the interaction feature from at least one first interest feature in step 1022 can be achieved through the following technical solution: performing a first mapping process on each first interest feature to obtain a first key feature corresponding to each first interest feature; performing a second mapping process on the interaction feature to obtain a first query feature corresponding to the interaction feature; determining a first weight corresponding to each first interest feature based on the first query feature corresponding to the interaction feature and the first key feature corresponding to each first interest feature; and determining the first interest feature with the highest first weight as the target first interest feature of the corresponding interaction feature. By obtaining the first interest feature that best matches each interaction feature through the first weight, the accuracy of interest discovery is improved.

[0080] As an example, each first interest feature is mapped through the first fully connected layer to obtain the first key feature corresponding to each first interest feature. That is, for 10 first interest features, 10 first key features can be obtained. For the interaction feature h1, the second fully connected layer performs a second mapping process on the interaction feature h1 to obtain the first query feature corresponding to the interaction feature h1. Based on the first query feature and 10 first-key features The first weight corresponding to each first interest feature is determined, that is, 10 first weights are determined. The first interest feature with the highest first weight among the 10 first weights is determined as the target first interest feature of the corresponding interaction feature. For example, if the first interest feature k1 has the highest first weight, then the first interest feature k1 is used as the target first interest feature of the interaction feature h1. See [link to relevant documentation]. Figure 5Step 1022 can obtain the target first interest feature k1 of interaction feature h1, the target first interest feature k1 of interaction feature h2, the target first interest feature k1 of interaction feature h3, the target first interest feature k4 of interaction feature h4, the target first interest feature k4 of interaction feature h5, the target first interest feature k6 of interaction feature h6, and the target first interest feature k6 of interaction feature h7.

[0081] As an example, the first weight is determined as follows: first, the first query feature is determined according to formula (1). With each first key feature Perform dot product processing to obtain 10 first dot product results. :

[0082] (1);

[0083] in, It is the first query feature. It is the first key feature corresponding to the i-th first interest feature. It is the first dot product result obtained based on the i-th first interest feature.

[0084] Continue by performing maximum likelihood processing on each first dot product result according to formula (2) to obtain the first weight corresponding to each first interest feature:

[0085] (2);

[0086] in, It is the first weight corresponding to the i-th first interest feature. It is the first dot product result obtained based on the i-th first interest feature.

[0087] In step 1023, for each target first interest feature, the target interaction features corresponding to the target first interest feature are aggregated to obtain the interaction interest features corresponding to the target first interest feature.

[0088] In some embodiments, the aggregation processing of target interaction features corresponding to the target first interest feature in step 1023 to obtain interaction interest features corresponding to the target first interest feature can be implemented through the following technical solution: performing a third mapping processing on the target first interest feature to obtain a second query feature corresponding to the target first interest feature; performing a fourth mapping processing on each target interaction feature to obtain a second key feature corresponding to each target interaction feature; obtaining a first value feature corresponding to each target interaction feature, the first value feature also being obtained by mapping the target interaction feature; determining a second weight corresponding to each target interaction feature based on the second query feature corresponding to the target first interest feature and the second key feature corresponding to each target interaction feature; and performing a weighted summation processing on the first value feature corresponding to each target interaction feature based on the second weight of each target interaction feature to obtain the interaction interest feature corresponding to the target first interest feature. By using an aggregation processing method based on an attention mechanism, interaction interest features representing the interest of interaction features can be obtained, thereby improving the accuracy of interest discovery for active users.

[0089] As an example, for the target first interest feature k1, the target interaction features h1, h2, and h3 mapped to the target first interest feature in step 1022 are aggregated. The aggregation method is as follows: the target first interest feature k1 is mapped through the third fully connected layer to obtain the second query feature Q corresponding to the target first interest feature k1; the target interaction features h1, h2, and h3 are mapped through the fourth mapping layer to obtain the second key feature of the target interaction feature h1, the second key feature of the target interaction feature h2, and the second key feature of the target interaction feature h3, respectively. Based on the second query feature Q and the second key feature of the target interaction feature h1, the second key feature of the target interaction feature h2, and the second key feature of the target interaction feature h3, the second key feature of the target interaction feature h1 is aggregated. Based on the key features, the second key features corresponding to the target interaction feature h2, and the second key features corresponding to the target interaction feature h3, the second weights corresponding to the target interaction feature h1, h2, and h3 are determined. The second weights correspond one-to-one with the target interaction features. Based on the second weights corresponding to the target interaction features h1, h2, and h3, the first value features of the target interaction features h1, h2, and h3 are weighted and summed to obtain the interaction interest features corresponding to the target first interest feature k1.

[0090] As an example, the second weight is determined as follows: first, the second query feature is determined according to formula (3). With each second bond feature Perform dot product processing to obtain 10 second dot product results. :

[0091] (3);

[0092] in, It is the second query feature. It is the second key feature corresponding to the i-th target interaction feature. It is the second dot product result obtained based on the interaction features of the i-th target.

[0093] Continue by performing maximum likelihood processing on each second dot product result according to formula (4) to obtain the second weight corresponding to each target interaction feature:

[0094] (4);

[0095] in, It is the second weight corresponding to the i-th target interaction feature. It is the second dot product result obtained based on the interaction features of the i-th target.

[0096] As an example, the method for performing weighted summation can be found in formula (5):

[0097] (5);

[0098] in, These are the interaction interest features corresponding to the target's first interest feature k1. It is the second weight corresponding to the i-th target interaction feature. It is the first value feature corresponding to the interaction feature of the i-th target.

[0099] In some embodiments, see Figure 3C , Figure 3C This is a flowchart illustrating the recommendation processing method based on artificial intelligence provided in this application embodiment. In step 102, at least one object feature is subjected to second attention processing based on the second object interest pool to obtain at least one object interest feature, which can be achieved through steps 1024 to 1026.

[0100] In step 1024, at least one second interest feature is obtained from the second object interest pool.

[0101] As an example, the second object interest pool is generated based on a second type of object, which consists of cold-start users whose activity level is lower than that of the second type of user. The second interest vectors representing different directions within the second object interest pool are generated using a Multi-Interest Network with Dynamic Routing (MIND) model. The input to the MIND model is object data and interaction data from many cold-start users, and the output is the second interest vectors representing different directions. The vector dimension of the second object interest pool is K*D, representing that the pool includes K second interest features, each of which represents a specific direction (type) of interest for the second type of object.

[0102] In step 1025, for each object feature, a target second interest feature corresponding to the object feature is obtained from at least one second interest feature.

[0103] In some embodiments, obtaining the target second interest feature corresponding to the object feature from at least one second interest feature in step 1025 can be achieved through the following technical solution: performing a fifth mapping process on each second interest feature to obtain a third key feature corresponding to each second interest feature; performing a sixth mapping process on the object feature to obtain a third query feature corresponding to the object feature; determining a third weight corresponding to each second interest feature based on the third query feature corresponding to the object feature and the third key feature corresponding to each second interest feature; and determining the second interest feature with the highest third weight as the target second interest feature of the corresponding object feature. By obtaining the second interest feature that best matches each object feature through the third weight, the accuracy of interest discovery is improved.

[0104] As an example, the fifth fully connected layer performs a fifth mapping process on each second interest feature to obtain the third key feature corresponding to each second interest feature. That is, for 10 second interest features, 10 third key features can be obtained. For the object feature (age range), the sixth fully connected layer performs a sixth mapping process on the object feature (age range) to obtain the third query feature corresponding to the object feature (age range). Based on third query features And 10 third-bond features The third weight is determined for each second interest feature, resulting in 10 third weights. The second interest feature with the highest third weight among these 10 is then identified as the target second interest feature for the corresponding object feature (age range). For example, if the second interest feature m2 has the highest third weight, then second interest feature m2 is selected as the target second interest feature for the object feature (age range). See [link to relevant documentation]. Figure 5Through step 1032, the target second interest feature m2 of object features (age range), target second interest feature m2 of object features (gender), target second interest feature m2 of object features (city), target second interest feature m6 of object features (spending power), and target second interest feature m4 of object features (city interaction data) can be obtained.

[0105] As an example, the third weight is determined as follows: first, the third query feature is determined according to formula (6). With each third bond feature Performing dot product processing yields 10 third dot product results. :

[0106] (6);

[0107] in, It is the third query feature. It is the third key feature corresponding to the i-th second interest feature. It is the third dot product result obtained based on the i-th second interest feature.

[0108] Continue by performing maximum likelihood processing on each third dot product result according to formula (7) to obtain the third weight corresponding to each second interest feature:

[0109] (7);

[0110] in, It is the third weight corresponding to the i-th second interest feature. It is the third dot product result obtained based on the i-th second interest feature.

[0111] In step 1026, for each target second interest feature, the target object features corresponding to the target second interest feature are aggregated to obtain the object interest features corresponding to the target second interest feature.

[0112] In some embodiments, the aggregation processing of object features corresponding to the target second interest feature in step 1026 to obtain object interest features corresponding to the target second interest feature can be implemented through the following technical solution: performing a seventh mapping processing on the target second interest feature to obtain a fourth query feature corresponding to the target second interest feature; performing an eighth mapping processing on each target object feature to obtain a fourth key feature corresponding to each target object feature; obtaining a second value feature corresponding to each target object feature; determining a fourth weight corresponding to each target object feature based on the fourth query feature corresponding to the target second interest feature and the fourth key feature corresponding to each target object feature; and performing a weighted summation processing on the second value feature corresponding to each target object feature based on the fourth weight to obtain the object interest feature corresponding to the target second interest feature. By using an aggregation processing method based on an attention mechanism, object interest features representing the interest of object features can be obtained, thereby improving the accuracy of interest discovery for cold-start users.

[0113] As an example, for the target second interest feature m2, the target object features (age range), target object features (gender), and target object features (city) mapped to the target second interest feature in step 1024 are aggregated. The aggregation method is as follows: the target second interest feature m2 is mapped through the seventh fully connected layer to obtain the fourth query feature q corresponding to the target second interest feature m2; the target object features (age range), target object features (gender), and target object features (city) are mapped through the eighth layer to obtain the fourth key features of the target object features (age range), the fourth key features of the target object features (gender), and the fourth key features of the target object features (city). Based on the second query feature Q and the fourth key features of the target object features (age range), the aggregation is performed. The fourth key feature, the fourth key feature corresponding to the target object feature (gender), and the fourth key feature corresponding to the target object feature (city) are used to determine the fourth weight, the fourth weight, and the fourth weight of the target object feature (age range), where the fourth weight corresponds one-to-one with the target object feature. Based on the fourth weight, the second value feature, the second value feature, and the second value feature of the target object feature (city) are weighted and summed to obtain the interaction object feature corresponding to the target second interest feature m2.

[0114] As an example, the fourth weight is determined as follows: first, the fourth query feature is determined according to formula (8). With each fourth bond feature Perform dot product processing to obtain 10 second dot product results. :

[0115] (8);

[0116] in, It is the fourth query feature. It is the fourth key feature corresponding to the feature of the i-th target object. It is the fourth dot product result obtained based on the features of the i-th target object.

[0117] Continue by performing maximum likelihood processing on each fourth dot product result according to formula (9) to obtain the fourth weight corresponding to each target object feature:

[0118] (9);

[0119] in, It is the fourth weight corresponding to the feature of the i-th target object. It is the fourth dot product result obtained based on the features of the i-th target object.

[0120] As an example, the weighted summation process can be performed using formula (10):

[0121] (10);

[0122] in, It is the object interest feature corresponding to the second interest feature m2 of the target. It is the fourth weight corresponding to the feature of the i-th target object. It is the second value feature corresponding to the feature of the i-th target object.

[0123] In some embodiments, the selection process in step 103 of at least one interactive interest feature and at least one object interest feature based on activity features to obtain at least one recall feature of the target object account can be implemented through the following technical solution: Perform a ninth mapping process on the activity features to obtain a fifth query feature of the activity features; perform a tenth mapping process on each interactive interest feature to obtain a fifth key feature corresponding to each interactive interest feature, and perform a tenth mapping process on each object interest feature to obtain a fifth key feature corresponding to each object interest feature; determine a fifth weight for each interactive interest feature and a fifth weight for each object interest feature based on the fifth query feature corresponding to the activity features, the fifth key feature corresponding to each interactive interest feature, and the fifth key feature corresponding to each object interest feature; perform a mixed descending sorting process on at least one interactive interest feature and at least one object interest feature based on the fifth weight; and determine at least one feature with the highest ranking among the at least one interactive interest feature and at least one object interest feature as the recall feature of the target object account. By selecting interest features through activity features, the interest features that best represent the adaptation to user activity can be obtained.

[0124] As an example, user activity characteristics can be differentiated according to different scenarios. For example, a user who clicks less than k videos in 7 days is an inactive user for 7 days. This application embodiment adopts a gating mechanism to divide activity into multiple levels. For example, the target account is inactive for 1 day (a user who clicks less than k videos in 1 day is an inactive user for 1 day), the target account is inactive for 3 days, the target account is inactive for 5 days, the target account is inactive for 7 days, and the target account is inactive for 14 days, thus mapping the user's activity to activity characteristics.

[0125] As an example, the ninth fully connected layer performs the ninth mapping process on the activity feature to obtain the fifth query feature corresponding to the activity feature. The tenth fully connected layer performs the tenth mapping process on object interest features A1, A2, A3, and A4 respectively to obtain the fifth key features of object interest features A1, A2, A3, A4, and A5. The tenth fully connected layer performs the tenth mapping process on interaction interest features B1, B2, and B3 to obtain the fifth key features of interaction interest features B1, B2, and B3 respectively. Based on the fifth query feature, the fifth key feature of object interest feature A1, and the corresponding... The fifth key features of interest feature A2, the fifth key features of corresponding object interest features A3, A4, and A5, the fifth key features of corresponding interaction interest features B1, B2, and B3 are used to determine the fifth weight of each object interest feature and each interaction interest feature. That is, if there are 3 object interest features and 4 interaction interest features, then 7 fifth weights will be obtained. Based on the obtained multiple fifth weights, the interaction interest features and object interest features corresponding to the fifth weights are sorted in descending order. The top 3 interaction interest features and object interest features among the 7 fifth weights are determined as the recall features of the target object account. For example, the fifth weights of object interest features A1, A2, and B3 are the top 3 interest features.

[0126] As an example, the fifth weight is determined as follows: first, the fifth query feature is compared with each fifth key feature according to formula (11). Perform dot product processing to obtain multiple fifth dot product results. :

[0127] (11);

[0128] in, It is the fifth query feature. It is the fifth key feature corresponding to the i-th interest feature (a certain interaction interest feature or a certain object interest feature). It is the fifth dot product result obtained based on the i-th interest feature.

[0129] Continue by performing maximum likelihood processing on each fifth dot product result according to formula (12) to obtain the fifth weight corresponding to each interest feature (a certain interaction interest feature or a certain object interest feature):

[0130] (12);

[0131] in, It is the fifth weight corresponding to the i-th interest feature. It is the fifth dot product result obtained based on the i-th interest feature.

[0132] In some embodiments, the information features are obtained through an information tower network, which includes N cascaded fully connected layers, where N is an integer greater than or equal to 2. In step 104, before obtaining target information from multiple pieces of information to be recommended based on at least one recall feature and the information features of multiple pieces of information to be recommended, the following processing is performed for each piece of information to be recommended: the input of the nth fully connected layer is fully connected through the nth fully connected layer in the N cascaded fully connected layers to obtain the nth fully connected result; the nth fully connected result is transmitted to the (n+1)th fully connected layer for further full connected processing; where n is an integer starting from 1 and increasing in value, and the value range of n satisfies 1≤n≤N-1; when n is 1, the input of the nth fully connected layer is the attribute information of the information to be recommended; when n is 2≤n≤N-1, the input of the nth fully connected layer is the fully connected result of the (n-1)th fully connected layer, and the output of the nth fully connected layer is the information feature.

[0133] As an example, the input of the Information Tower Network includes attribute information such as the identifier, category, and attributes of the information to be recommended. After processing the input through 2-3 cascaded fully connected layers, information features are obtained. For example, the Information Tower Network includes 3 cascaded fully connected layers. The attribute information is fully connected through the first fully connected layer to obtain the first fully connected result. The first fully connected result is transmitted to the second fully connected layer, which fully connects the first fully connected result to obtain the second fully connected result. The second fully connected result is transmitted to the third fully connected layer, which fully connects the second fully connected result to obtain the third fully connected result. The obtained third fully connected result is used as the information feature.

[0134] In some embodiments, the step 104, which involves obtaining target information from multiple pieces of information to be recommended based on at least one recall feature and the information features of multiple pieces of information to be recommended, can be achieved through the following technical solution: For each recall feature, perform the following processing: determine the similarity between the recall feature and the information features of each piece of information to be recommended; based on the similarity of each piece of information to be recommended, sort the multiple pieces of information to be recommended in descending order, and determine at least one piece of information to be recommended that ranks first as the target information for the recall feature.

[0135] As an example, for recall feature A1 and recall feature A2, the similarity between the information features of each of the 1000 pieces of information to be recommended and recall feature A1 is determined, resulting in 1000 similarity scores. Based on these similarity scores, the 1000 pieces of information to be recommended are sorted in descending order, and the top 100 pieces of information are selected as the target information for recall feature A1. The similarity between the information features of each of the 1000 pieces of information to be recommended and recall feature A2 is determined, resulting in 1000 similarity scores. Based on these similarity scores, the 1000 pieces of information to be recommended are sorted in descending order, and... The top 100 recommended items are used as target information for recall feature A2. If there is no overlap between the 100 target items obtained from recall feature A1 and the 100 target items obtained from recall feature A2, a total of 200 target items can be recalled, and recommendation operations are performed based on the 200 target items. If there is overlap between the 100 target items obtained from recall feature A1 and the 100 target items obtained from recall feature A2, for example, if one target item is duplicated, a total of 199 target items can be recalled, and recommendation operations are performed based on the 199 target items.

[0136] In some embodiments, attention processing and selection processing are implemented through an object tower network, the information features of the information to be recommended are obtained through an information tower network, and the processing of acquiring target information is implemented through a similarity network. The multi-interest recall model is composed of an object tower network, an information tower network, and a similarity network. At least one interaction feature and at least one object feature of the sample object account are forward propagated through the object tower network to obtain the recall features of the sample object account; multiple sample information to be recommended are forward propagated through the information tower network to obtain the information features of each sample information to be recommended; the target sample information to be recommended corresponding to each recall feature is determined through the similarity network; the following processing is performed for each recall feature: the loss of the corresponding recall feature is determined based on the predicted similarity and pre-labeled similarity of the target sample information to be recommended corresponding to the recall feature; the parameters of the multi-interest recall model are updated based on the minimum value of the loss corresponding to at least one recall feature.

[0137] As an example, see Figure 4The multi-interest recall model includes an object pyramid network, an information pyramid network, and a similarity network. It can also include two interest pool generation networks, such as two MINDs. The object pyramid network takes interaction information and object information as input and outputs at least one recall feature of the object account. The information pyramid network takes relevant data (including identifiers, types, attribute data, etc.) as input and outputs information features of the information to be recommended. The similarity network takes at least one recall feature and information features corresponding one-to-one with multiple pieces of information to be recommended and outputs target information corresponding to each recall feature among the multiple pieces of information to be recommended. During the training phase, at least one interaction feature and at least one object feature of the sample account are forward propagated through the object tower network to obtain the recall features of the sample account. Specifically, the interaction features of the sample account are subjected to first attention processing based on a first object interest pool to obtain the interaction interest features of the sample account, and the object features of the sample account are subjected to second attention processing based on a second object interest pool to obtain the object interest features of the sample account. The activity features of the sample account are then obtained, and the interaction interest features and object interest features of the sample account are selected based on the activity features to obtain the recall features of the sample account. Multiple sample recommendation information is forward propagated through the information tower network to obtain the information features of each sample recommendation information. Finally, the target sample recommendation information for each recall feature of the corresponding sample account is determined through a similarity network. For a specific recall feature A of a sample account, the following processing is performed: Based on the predicted similarity and pre-labeled similarity of the target sample to be recommended information corresponding to the recall feature, the loss of the corresponding recall feature is determined, that is, the difference between the predicted similarity and pre-labeled similarity of the to-be-recommended information of each sample is determined, and the difference corresponding to the to-be-recommended information of multiple samples is averaged to obtain the loss of the corresponding recall feature A. Based on the minimum value of the loss corresponding to at least one recall feature, the parameters of the multi-interest recall model are updated, or based on the minimum value of the loss corresponding to at least one recall feature, the regularized loss of each interest feature in the first object interest pool, and the regularized loss of each interest feature in the second object interest pool, the parameters of the multi-interest recall model are updated. For example, the regularized loss is the norm of the interest pool vector H, i.e., ||H*HT-Id||, where I is a diagonal matrix, thereby avoiding training collapse caused by the difference between each interest vector being too small.

[0138] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0139] In some embodiments, the terminal sends a recommendation request to the application server. The application server obtains the interaction features and object features of the target object account, performs first attention processing on the interaction features based on a first object interest pool to obtain interaction interest features, and performs second attention processing on the object features based on a second object interest pool to obtain object interest features. The application server also obtains the activity features of the target object account, and performs selection processing on the interaction interest features and object interest features based on the activity features to obtain recall features of the target object account. Based on the recall features and the information features of multiple pieces of information to be recommended, the application server obtains target information from the multiple pieces of information to be recommended, performs fine ranking processing on the target information, and returns the ranking result to the terminal corresponding to the target object account for presentation.

[0140] The AI-based recommendation processing method provided in this application integrates cold-start information and interaction information, and performs multi-interest recall modeling based on cold-start information and interaction information. It effectively improves the situation where recommendation systems rely too much on interaction information and lack the ability to capture interests of cold-start objects. It can be widely applied to the recall algorithm framework of industry recommendation systems and can effectively improve the recommendation effect.

[0141] The multi-interest recall model, based on a dual-tower deep neural network, extracts k interest expression vectors from the user interaction information sequence at the object tower network. These k interest expression vectors are then compared with the information vector output by the information tower network to calculate similarity for recall. This multi-interest recall model achieves a transition from single-interest vector recall to multi-interest vector recall. While the multi-interest recall model fully utilizes the interaction information of objects and extracts multiple interest vectors of different categories, it primarily focuses on extraction and modeling based on the user interaction information sequence, lacking modeling for object information. This application's embodiments address this by constructing a multi-interest recall model that considers both the user interaction information sequence and the user's object information. This allows the multi-interest recall model provided by this application to completely extract both the user's object information and user interaction information, and dynamically fuses interest vectors of different categories based on the user's interaction information and object information, thereby improving the accuracy of the recall module in modeling user expression vectors in the recommendation system.

[0142] The following details the multi-interest recall model, which includes an object tower network and an information tower network. The information tower network's input includes the identifier, category, and attributes of the information to be recommended. After processing through 2-3 cascaded fully connected layers, information features are obtained. The object tower network's input includes user interaction information (e.g., the k most recently clicked videos) and user object information (e.g., age group, gender, city, followed accounts, etc.). The object tower network's input also includes user activity features, which can be differentiated according to different scenarios. For example, a user who has clicked fewer than k videos in 7 days is considered an inactive user for 7 days. This application's embodiment uses a gating mechanism to classify activity into multiple levels, such as: 1 day inactive (a user who has clicked fewer than k videos in 1 day is considered an inactive user for 1 day), 3 days inactive, 5 days inactive, 7 days inactive, and 14 days inactive. User activity is mapped to activity features, which are then input into the object tower network for processing.

[0143] In this embodiment of the application, two interest pools need to be constructed, including an active user interest pool and a cold-start user interest pool. The active user interest pool is matched and expressed according to the user interaction information sequence, and the cold-start user interest pool is matched and expressed according to the user's object information. Both interest pools are K*D dimensional vectors, representing that each interest pool can hold up to K interest vectors.

[0144] See Figure 5 , Figure 5 This is a schematic diagram of the attention mechanism in the AI-based recommendation method provided in this application embodiment. First, the interaction information sequence is extracted and processed, that is, the interaction information sequence is merged into the corresponding interest vector. Then, an attention mechanism maps the interaction features of the interaction information sequence to the active user interest pool (first object interest pool). The interaction features are dense features (embeddings) that are embedded and compressed from discrete interaction information. The interaction features of the interaction information sequence are h*D dimensional vectors, representing that each interaction feature has a dimension of D, and there are a total of h interaction features (each interaction information corresponds to one interaction feature). Using the attention mechanism, the h interaction features are respectively processed with the active user interest pool. For each h, the interest feature K with the largest weight is selected as the interest expressed by this interaction information. A fully connected layer is used to map the interaction feature h to obtain the corresponding vector (first query feature Q). A fully connected layer is then used to map each interest feature in the active user interest pool to obtain the vector corresponding to each interest feature (first key feature k). i In this embodiment, the first query feature and the first key feature are actually processed using a dot product method to obtain the first processing result. (For example, the dot product result), but not limited to the dot product processing method, see formulas (13) to (15):

[0145] (13);

[0146] (14);

[0147] (15);

[0148] Continue by performing maximum likelihood processing on each first dot product result according to formula (16) to obtain the first weight corresponding to each interest feature K:

[0149] (16);

[0150] in, It is the weight corresponding to the i-th interest feature. It is the dot product result obtained based on the i-th interest feature.

[0151] See Figure 5 It can obtain the target first interest feature k1 of interaction feature h1, the target first interest feature k1 of interaction feature h2, the target first interest feature k1 of interaction feature h3, the target first interest feature k4 of interaction feature h4, the target first interest feature k4 of interaction feature h5, the target first interest feature k6 of interaction feature h6, and the target first interest feature k6 of interaction feature h7. Finally, based on the interaction information sequence, three interaction multi-interest expression vectors (first interest features) can be extracted from the active user interest pool.

[0152] After extracting three interactive multi-interest expression vectors, interest aggregation can be performed on the interactive information sequence. For the i-th interest feature, the interactive information mapped to it is selected for attention aggregation processing. In the attention aggregation processing, the interest feature is mapped through a fully connected layer to obtain the corresponding vector (second query feature Q). The j target interactive features mapped to the i-th interest feature are mapped through a fully connected layer to obtain the second key feature K of each target interactive feature. The j target interactive features mapped to the i-th interest feature are mapped through a fully connected layer to obtain the first value feature V of each target interactive feature. Based on the second key feature K and the second query feature Q of each target interactive feature, the weight of each target interactive feature is determined. Based on the weight, the first value feature V of each target interactive feature is weighted and summed to obtain the user's interactive interest expression vector mi (interactive interest feature) under the i-th interest feature.

[0153] Similar to extracting the first object interest pool based on the interactive information sequence, the user's object information is constructed as a sequence input similar to interactive information to obtain s object features (in vector form) such as user age, gender, city, consumption level, and followed accounts. The object features based on the cold start user interest pool are determined through the aforementioned similar interest extraction method, and then the user's object interest expression vector (object interest feature) is obtained through object feature aggregation processing.

[0154] After obtaining the interaction interest expression vector and the object interest expression vector, the interest features are selected by extracting user activity features. The dynamic gating selection structure of attention is also adopted. The user activity features are used as Q in formulas (13) to (15), and the interaction interest expression vector and the object interest expression vector are used as K in formulas (13) to (15). The weights of each interaction interest expression vector and each object interest expression vector are calculated, and the interest vectors corresponding to the top k weights after descending sorting are used as the final output user vectors. In online application, the multiple pieces of information with the highest similarity are determined for each of the k user vectors (recall features). The determined multiple pieces of information to be recommended are merged and used as the recall result to be sent to the subsequent links of the recommendation system.

[0155] During the training phase, the loss value of multiple target information recalled by each interest vector can be determined. The minimum loss value among the loss values ​​of k interest vectors can be selected as the basis for updating parameters. At the same time, the regularization loss between the vectors in the interest pool is increased. The regularization loss is the norm of the interest pool vector H, i.e., ||H*HT-Id||, where I is a diagonal matrix, thereby avoiding training collapse caused by the difference between the interest vectors being too small.

[0156] This application proposes a multi-interest recall model that comprehensively considers cold start information and interaction information. This allows the model to prioritize different aspects of object information and interaction information based on user activity, better capturing user interests and improving the efficiency of the recommendation system's distribution, exploration, and utilization.

[0157] This application proposes a multi-interest recall model that integrates cold start information and user interaction information. It comprehensively considers the different interests expressed by users' object information and interaction information, and uses an activity gating mechanism to dynamically select different types of interests in the interest pool. This model can make more complete use of user information input, fully utilize user interaction information within the domain, and also has sufficient interest capture capability for cold start users, effectively improving the conversion rate of cold start users. It can take into account both cold start users and active users.

[0158] It is understood that in the embodiments of this application, data such as user information are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0159] The following continues to describe the exemplary structure of the AI-based recommendation processing device 255 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the AI-based recommendation processing device 255 in the memory 250 may include: an acquisition module 2551, used to acquire at least one interaction feature and at least one object feature of the target object account; an attention module 2552, used to perform a first attention processing on the at least one interaction feature based on a first object interest pool to obtain at least one interaction interest feature, and to perform a second attention processing on the at least one object feature based on a second object interest pool to obtain at least one object interest feature; wherein the object activity corresponding to the first object interest pool is higher than the object activity corresponding to the second object interest pool; a selection module 2553, used to acquire the activity feature of the target object account, and to perform selection processing on the at least one interaction interest feature and at least one object interest feature based on the activity feature to obtain at least one recall feature of the target object account; and a recommendation module 2554, used to acquire target information from multiple pieces of information to be recommended based on at least one recall feature and information features of multiple pieces of information to be recommended, and to perform a recommendation operation on the target object account based on the target information.

[0160] In some embodiments, the attention module 2552 is further configured to: obtain at least one first interest feature from a first object interest pool; for each interaction feature, obtain a target first interest feature corresponding to the interaction feature from the at least one first interest feature; and for each target first interest feature, perform aggregation processing on the target interaction feature corresponding to the target first interest feature to obtain an interaction interest feature corresponding to the target first interest feature.

[0161] In some embodiments, the attention module 2552 is further configured to: perform a first mapping process on each first interest feature to obtain a first key feature corresponding to each first interest feature; perform a second mapping process on the interaction feature to obtain a first query feature corresponding to the interaction feature; determine a first weight corresponding to each first interest feature based on the first query feature corresponding to the interaction feature and the first key feature corresponding to each first interest feature; and determine the first interest feature with the highest first weight as the target first interest feature of the corresponding interaction feature.

[0162] In some embodiments, the attention module 2552 is further configured to: perform a third mapping process on the target first interest feature to obtain a second query feature corresponding to the target first interest feature; perform a fourth mapping process on each target interaction feature to obtain a second key feature corresponding to each target interaction feature; obtain a first value feature corresponding to each target interaction feature; determine a second weight corresponding to each target interaction feature based on the second query feature corresponding to the target first interest feature and the second key feature corresponding to each target interaction feature; and perform a weighted summation process on the first value feature corresponding to each target interaction feature based on the second weight corresponding to each target interaction feature to obtain an interaction interest feature corresponding to the target first interest feature.

[0163] In some embodiments, the attention module 2552 is further configured to: obtain at least one second interest feature from the second object interest pool; for each object feature, obtain a target second interest feature corresponding to the object feature from the at least one second interest feature; and for each target second interest feature, perform aggregation processing on the target object feature corresponding to the target second interest feature to obtain an object interest feature corresponding to the target second interest feature.

[0164] In some embodiments, the attention module 2552 is further configured to: perform a fifth mapping process on each second interest feature to obtain a third key feature corresponding to each second interest feature; perform a sixth mapping process on the object feature to obtain a third query feature corresponding to the object feature; determine a third weight corresponding to each second interest feature based on the third query feature corresponding to the object feature and the third key feature corresponding to each second interest feature; and determine the second interest feature with the highest third weight as the target second interest feature of the corresponding object feature.

[0165] In some embodiments, the attention module 2552 is further configured to: perform a seventh mapping process on the target second interest feature to obtain a fourth query feature corresponding to the target second interest feature; perform an eighth mapping process on each target object feature to obtain a fourth key feature corresponding to each target object feature; obtain a second value feature corresponding to each target object feature; determine a fourth weight corresponding to each target object feature based on the fourth query feature corresponding to the target second interest feature and the fourth key feature corresponding to each target object feature; and perform a weighted summation process on the second value feature corresponding to each target object feature based on the fourth weight corresponding to each target object feature to obtain an object interest feature corresponding to the target second interest feature.

[0166] In some embodiments, the selection module 2553 is further configured to: perform a ninth mapping process on the activity feature to obtain a fifth query feature of the activity feature; perform a tenth mapping process on each interaction interest feature to obtain a fifth key feature corresponding to each interaction interest feature, and perform a tenth mapping process on each object interest feature to obtain a fifth key feature corresponding to each object interest feature; determine a fifth weight corresponding to each interaction interest feature and a fifth weight corresponding to each object interest feature based on the fifth query feature of the corresponding activity feature, the fifth key feature corresponding to each interaction interest feature, and the fifth key feature corresponding to each object interest feature; and determine the feature with the highest fifth weight among at least one interaction interest feature and at least one object interest feature as the recall feature of the target object account.

[0167] In some embodiments, the information features are obtained through an information tower network, which includes N cascaded fully connected layers, where N is an integer greater than or equal to 2. The recommendation module 2554 is further configured to: before obtaining target information from multiple pieces of information to be recommended based on at least one recall feature and the information features of multiple pieces of information to be recommended, perform the following processing for each piece of information to be recommended: perform fully connected processing on the input of the nth fully connected layer through the nth fully connected layer in the N cascaded fully connected layers to obtain the nth fully connected result; transmit the nth fully connected result to the (n+1)th fully connected layer for further fully connected processing; where n is an integer starting from 1 and increasing in value, and the value range of n satisfies 1≤n≤N-1; when n is 1, the input of the nth fully connected layer is the attribute information of the information to be recommended; when n is 2≤n≤N-1, the input of the nth fully connected layer is the fully connected result of the (n-1)th fully connected layer, and the output of the nth fully connected layer is the information feature.

[0168] In some embodiments, the recommendation module 2554 is further configured to: for each recall feature, perform the following processing: determine the similarity between the recall feature and the information features of each piece of information to be recommended; based on the similarity of each piece of information to be recommended, sort the multiple pieces of information to be recommended in descending order, and determine at least one piece of information to be recommended that ranks first as the target information for the recall feature.

[0169] In some embodiments, attention processing and selection processing are implemented through an object tower network, the information features of the information to be recommended are obtained through an information tower network, and the processing of acquiring target information is implemented through a similarity network. The multi-interest recall model is composed of an object tower network, an information tower network, and a similarity network. The device also includes a training module 2555, used for: forward propagating at least one interaction feature and at least one object feature of a sample object account through the object tower network to obtain the recall feature of the sample object account; forward propagating multiple sample information to be recommended through the information tower network to obtain the information features of each sample information to be recommended; determining the target sample information to be recommended corresponding to each recall feature through the similarity network; and performing the following processing for each recall feature: determining the loss of the corresponding recall feature based on the predicted similarity and pre-labeled similarity of the target sample information to be recommended corresponding to the recall feature; and updating the parameters of the multi-interest recall model based on the minimum value of the loss corresponding to at least one recall feature.

[0170] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the artificial intelligence-based recommended processing method described above in this application.

[0171] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to execute the AI-based recommended processing method provided in this application. For example... Figures 3A-3C The example shown is an AI-based recommendation processing method.

[0172] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0173] In some embodiments, executable instructions may take 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 standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0174] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0175] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0176] In summary, this application embodiment performs first attention processing on interaction features based on a first object interest pool to obtain interaction interest features, and performs second attention processing on object features based on a second object interest pool to obtain object interest features. This allows different types of features to be mapped to interest pools of different types of objects. Based on activity features, interaction interest features and object interest features are selected to obtain recall features of the target object account. By using activity features as a threshold to filter interest features, the selected recall features can be used to characterize the interest distribution of the target object account. Subsequently, recommendation processing is performed based on the target information of the corresponding recall features, which can effectively improve recommendation efficiency and accuracy.

[0177] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A recommendation processing method based on artificial intelligence, characterized in that, The method includes: Obtain at least one interaction characteristic and at least one object characteristic of the target object's account; From at least one first interest feature in the first object interest pool, obtain a target first interest feature corresponding to the interaction feature; perform aggregation processing on the target interaction feature corresponding to the target first interest feature to obtain the interaction interest feature corresponding to the target first interest feature; From at least one second interest feature in the second object interest pool, a target second interest feature corresponding to the object feature is obtained; the target object feature corresponding to the target second interest feature is aggregated to obtain the object interest feature corresponding to the target second interest feature. Among them, the activity level of the objects corresponding to the first object interest pool is higher than that of the objects corresponding to the second object interest pool. The activity characteristics of the target account are obtained, and the activity characteristics, the at least one interactive interest characteristics, and the at least one object interest characteristics are mapped respectively. Based on the obtained weights of each interactive interest characteristic and each object interest characteristic, the at least one interactive interest characteristic and the at least one object interest characteristic are sorted in a mixed descending order, and the at least one characteristic with the highest ranking is determined as the recall characteristic of the target account. Based on the at least one recall feature and the information features of multiple pieces of information to be recommended, target information is obtained from the multiple pieces of information to be recommended, and a recommendation operation is performed on the target object account based on the target information.

2. The method according to claim 1, characterized in that, The step of obtaining the target first interest feature corresponding to the interaction feature from at least one first interest feature in the first object interest pool includes: Obtain at least one first interest feature from the first object interest pool; For each of the interaction features, a target first interest feature corresponding to the interaction feature is obtained from the at least one first interest feature.

3. The method according to claim 2, characterized in that, The step of obtaining the target first interest feature corresponding to the interaction feature from the at least one first interest feature includes: Perform a first mapping process on each of the first interest features to obtain a first key feature corresponding to each of the first interest features; The interaction features are subjected to a second mapping process to obtain a first query feature corresponding to the interaction features; Based on the first query feature corresponding to the interaction feature and the first key feature corresponding to each first interest feature, a first weight corresponding to each first interest feature is determined. The first interest feature with the highest weight is determined as the target first interest feature corresponding to the interaction feature.

4. The method according to claim 2, characterized in that, The aggregation process of the target interaction features corresponding to the target first interest feature to obtain the interaction interest features corresponding to the target first interest feature includes: A third mapping process is performed on the target's first interest feature to obtain a second query feature corresponding to the target's first interest feature; A fourth mapping process is performed on each of the target interaction features to obtain the second key feature corresponding to each of the target interaction features; Obtain the first value feature corresponding to each of the target interaction features; Based on the second query feature corresponding to the first interest feature of the target and the second key feature corresponding to each of the target interaction features, a second weight corresponding to each of the target interaction features is determined; Based on the second weight corresponding to each of the target interaction features, the first value features corresponding to each of the target interaction features are weighted and summed to obtain the interaction interest features corresponding to the target first interest features.

5. The method according to claim 1, characterized in that, The step of obtaining a target second interest feature corresponding to the object feature from at least one second interest feature in the second object interest pool includes: Obtain at least one second interest feature from the second object interest pool; For each of the object features, a target second interest feature corresponding to the object feature is obtained from the at least one second interest feature.

6. The method according to claim 5, characterized in that, The step of obtaining the target second interest feature corresponding to the object feature from the at least one second interest feature includes: Perform a fifth mapping process on each of the second interest features to obtain the third key feature corresponding to each of the second interest features; The object features are subjected to a sixth mapping process to obtain the third query feature corresponding to the object features; Based on the third query feature corresponding to the object feature and the third key feature corresponding to each second interest feature, a third weight corresponding to each second interest feature is determined. The second interest feature with the highest third weight is determined as the target second interest feature corresponding to the object feature.

7. The method according to claim 5, characterized in that, The aggregation process of the target object features corresponding to the target second interest feature to obtain the object interest features corresponding to the target second interest feature includes: A seventh mapping process is performed on the target second interest feature to obtain a fourth query feature corresponding to the target second interest feature; The eighth mapping process is performed on each of the target object features to obtain the fourth key feature corresponding to each of the target object features; Obtain the second value feature corresponding to each of the target object features; Based on the fourth query feature corresponding to the second interest feature of the target and the fourth key feature corresponding to each target object feature, the fourth weight corresponding to each target object feature is determined; Based on the fourth weight corresponding to each target object feature, the second value feature corresponding to each target object feature is weighted and summed to obtain the object interest feature corresponding to the target second interest feature.

8. The method according to claim 1, characterized in that, The process of mapping the activity feature, the at least one interaction interest feature, and the at least one object interest feature, and then sorting the at least one interaction interest feature and the at least one object interest feature in a mixed descending order based on the obtained weights corresponding to each interaction interest feature and each object interest feature, includes: The activity feature is subjected to a ninth mapping process to obtain the fifth query feature of the activity feature; The tenth mapping process is performed on each of the interactive interest features to obtain the fifth key feature corresponding to each of the interactive interest features, and the tenth mapping process is performed on each of the object interest features to obtain the fifth key feature corresponding to each of the object interest features. Based on the fifth query feature corresponding to the activity feature, the fifth key feature corresponding to each interaction interest feature, and the fifth key feature corresponding to each object interest feature, the fifth weight corresponding to each interaction interest feature and the fifth weight corresponding to each object interest feature are determined. Based on the fifth weight, the at least one interactive interest feature and the at least one object interest feature are subjected to a mixed descending sorting process.

9. The method according to claim 1, characterized in that, The information features are obtained through an information tower network, which includes N cascaded fully connected layers, where N is an integer greater than or equal to 2. Before obtaining target information from the multiple pieces of information to be recommended based on the at least one recall feature and the information features of the multiple pieces of information to be recommended, the method further includes: For each piece of information to be recommended, the following processing is performed: The input of the nth fully connected layer is processed by the nth fully connected layer in the N cascaded fully connected layers to obtain the nth fully connected result. The nth fully connected layer result is transmitted to the (n+1)th fully connected layer to continue the fully connected processing; Wherein, n is an integer starting from 1 and increasing in increments, and the range of n satisfies 1≤n≤N-1; when n is 1, the input of the nth fully connected layer is the attribute information of the information to be recommended; when n is 2≤n≤N-1, the input of the nth fully connected layer is the fully connected result of the (n-1)th fully connected layer; and the output of the nth fully connected layer is the information feature.

10. The method according to claim 1, characterized in that, The step of obtaining target information from the multiple pieces of information to be recommended based on the at least one recall feature and the information features of the multiple pieces of information to be recommended includes: For each of the aforementioned recall features, the following processing is performed: Determine the similarity between the recall features and the information features of each piece of information to be recommended; Based on the similarity of each piece of information to be recommended, the multiple pieces of information to be recommended are sorted in descending order, and at least one piece of information to be recommended that ranks first is determined as the target information for the recall feature.

11. The method according to claim 1, characterized in that, The aggregation process of target object features corresponding to the target second interest feature to obtain object interest features corresponding to the target second interest feature, and the aggregation process of target object features corresponding to the target second interest feature to obtain object interest features corresponding to the target second interest feature, are implemented through an object pyramid network. The information features of the information to be recommended are obtained through an information pyramid network. The processing of obtaining target information is implemented through a similarity network. The multi-interest recall model is composed of an object pyramid network, an information pyramid network, and a similarity network. The method further includes: At least one interaction feature and at least one object feature of the sample object account are forward propagated in the object tower network to obtain the recall feature of the sample object account. Multiple sample recommendation information is forward propagated in the information tower network to obtain the information features of each sample recommendation information; The similarity network is used to determine the target sample information to be recommended for each of the recall features; For each of the recall features, the following processing is performed: based on the predicted similarity and pre-labeled similarity of the target sample to be recommended information corresponding to the recall feature, the loss corresponding to the recall feature is determined; The parameters of the multi-interest recall model are updated based on the minimum value of the loss corresponding to at least one of the recall features.

12. A recommendation processing device based on artificial intelligence, characterized in that, The device includes: The acquisition module is used to acquire at least one interaction feature and at least one object feature of the target object's account; An attention module is configured to: obtain a target first interest feature corresponding to the interaction feature from at least one first interest feature in a first object interest pool; perform aggregation processing on the target interaction feature corresponding to the target first interest feature to obtain an interaction interest feature corresponding to the target first interest feature; obtain a target second interest feature corresponding to the object feature from at least one second interest feature in a second object interest pool; and perform aggregation processing on the target object feature corresponding to the target second interest feature to obtain an object interest feature corresponding to the target second interest feature. Among them, the activity level of the objects corresponding to the first object interest pool is higher than that of the objects corresponding to the second object interest pool. The selection module is used to obtain the activity characteristics of the target account, and to map the activity characteristics, the at least one interactive interest characteristics, and the at least one object interest characteristics respectively. Based on the obtained weights of each interactive interest characteristic and each object interest characteristic, the at least one interactive interest characteristic and the at least one object interest characteristic are sorted in a mixed descending order, and the at least one characteristic with the highest ranking is determined as the recall characteristic of the target account. The recommendation module is used to obtain target information from the multiple pieces of information to be recommended based on the at least one recall feature and the information features of multiple pieces of information to be recommended, and to perform a recommendation operation on the target object account based on the target information.

13. The apparatus according to claim 12, characterized in that, The attention module is further configured to obtain at least one first interest feature from the first object interest pool; and for each interaction feature, to obtain a target first interest feature corresponding to the interaction feature from the at least one first interest feature.

14. The apparatus according to claim 13, characterized in that, The attention module is further configured to perform a first mapping process on each of the first interest features to obtain a first key feature corresponding to each of the first interest features; The interaction features are subjected to a second mapping process to obtain a first query feature corresponding to the interaction features; Based on the first query feature corresponding to the interaction feature and the first key feature corresponding to each first interest feature, a first weight corresponding to each first interest feature is determined; the first interest feature with the highest first weight is determined as the target first interest feature corresponding to the interaction feature.

15. The apparatus according to claim 13, characterized in that, The attention module is further configured to perform a third mapping process on the target first interest feature to obtain a second query feature corresponding to the target first interest feature; and to perform a fourth mapping process on each target interaction feature to obtain a second key feature corresponding to each target interaction feature; Obtain the first value feature corresponding to each of the target interaction features; Based on the second query feature corresponding to the first interest feature of the target and the second key feature corresponding to each of the target interaction features, a second weight corresponding to each of the target interaction features is determined; Based on the second weight corresponding to each of the target interaction features, the first value features corresponding to each of the target interaction features are weighted and summed to obtain the interaction interest features corresponding to the target first interest features.

16. The apparatus according to claim 12, characterized in that, The attention module is further configured to obtain at least one second interest feature from the second object interest pool; and for each object feature, to obtain a target second interest feature corresponding to the object feature from the at least one second interest feature.

17. The apparatus according to claim 16, characterized in that, The attention module is further configured to perform a fifth mapping process on each second interest feature to obtain a third key feature corresponding to each second interest feature; and to perform a sixth mapping process on the object feature to obtain a third query feature corresponding to the object feature. Based on the third query feature corresponding to the object feature and the third key feature corresponding to each second interest feature, a third weight corresponding to each second interest feature is determined; the second interest feature with the highest third weight is determined as the target second interest feature corresponding to the object feature.

18. The apparatus according to claim 16, characterized in that, The attention module is further configured to perform a seventh mapping process on the target second interest feature to obtain a fourth query feature corresponding to the target second interest feature; and to perform an eighth mapping process on each target object feature to obtain a fourth key feature corresponding to each target object feature. Obtain the second value feature corresponding to each of the target object features; Based on the fourth query feature corresponding to the second interest feature of the target and the fourth key feature corresponding to each target object feature, the fourth weight corresponding to each target object feature is determined; Based on the fourth weight corresponding to each target object feature, the second value feature corresponding to each target object feature is weighted and summed to obtain the object interest feature corresponding to the target second interest feature.

19. The apparatus according to claim 12, characterized in that, The selection module is also used to perform a ninth mapping process on the activity feature to obtain a fifth query feature of the activity feature; A tenth mapping process is performed on each of the interactive interest features to obtain a fifth key feature corresponding to each of the interactive interest features, and a tenth mapping process is performed on each of the object interest features to obtain a fifth key feature corresponding to each of the object interest features; based on the fifth query feature corresponding to the activity feature, the fifth key feature corresponding to each of the interactive interest features, and the fifth key feature corresponding to each of the object interest features, a fifth weight corresponding to each of the interactive interest features and a fifth weight corresponding to each of the object interest features are determined; Based on the fifth weight, the at least one interactive interest feature and the at least one object interest feature are subjected to a mixed descending sorting process.

20. The apparatus according to claim 12, characterized in that, The information features are obtained through an information tower network, which includes N cascaded fully connected layers, where N is an integer greater than or equal to 2. The recommendation module is further configured to perform the following processing on each piece of information to be recommended before obtaining target information from the multiple pieces of information to be recommended based on the at least one recall feature and the information features of the multiple pieces of information to be recommended: The input of the nth fully connected layer is processed by the nth fully connected layer in N cascaded fully connected layers to obtain the nth fully connected result; the nth fully connected result is then transmitted to the (n+1)th fully connected layer for further fully connected processing; where n is an integer starting from 1 and the value of n satisfies 1≤n≤N-1; when n is 1, the input of the nth fully connected layer is the attribute information of the information to be recommended; when n is 2≤n≤N-1, the input of the nth fully connected layer is the fully connected result of the (n-1)th fully connected layer; and the output of the nth fully connected layer is the information feature.

21. The apparatus according to claim 12, characterized in that, The recommendation module is further configured to perform the following processing for each of the recall features: determine the similarity between the recall feature and the information features of each of the information to be recommended; based on the similarity of each of the information to be recommended, sort the information in descending order, and determine at least one of the information to be recommended that ranks first as the target information for the recall feature.

22. The apparatus according to claim 12, characterized in that, The aggregation processing of target object features corresponding to the target second interest feature to obtain object interest features corresponding to the target second interest feature is implemented through an object tower network. The information features of the information to be recommended are obtained through an information tower network. The processing of obtaining target information is implemented through a similarity network. The multi-interest recall model is composed of an object tower network, an information tower network, and a similarity network. The device further includes: The training module is used to forward propagate at least one interaction feature and at least one object feature of the sample object account in the object tower network to obtain the recall feature of the sample object account; forward propagate multiple sample recommendation information in the information tower network to obtain the information feature of each sample recommendation information; determine the target sample recommendation information corresponding to each recall feature through the similarity network; and perform the following processing for each recall feature: determine the loss corresponding to the recall feature based on the predicted similarity and pre-labeled similarity of the target sample recommendation information corresponding to the recall feature; and update the parameters of the multi-interest recall model based on the minimum value of the loss corresponding to at least one recall feature.

23. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the AI-based recommendation processing method according to any one of claims 1 to 11.

24. A computer-readable storage medium storing executable instructions, characterized in that, When the executable instructions are executed by the processor, they implement the AI-based recommendation processing method according to any one of claims 1 to 11.

25. A computer program product comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the AI-based recommendation processing method according to any one of claims 1 to 11.

Images