Artificial intelligence-based information recommendation method, apparatus, and electronic device

Abstract

This application provides an information recommendation method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence. The method includes: determining the correlation characteristics between pending location information and all information in an information set; wherein the information set includes at least one of the following: fixed location information; pending location information; the fixed location information is information that has already been assigned display positions in a location sequence, and the pending location information is information about display positions in a location sequence to be assigned; determining a corresponding first click-through rate based on the correlation characteristics of each pending location information; assigning the unassigned and highest-priority display position in the location sequence to the pending location information with the highest first click-through rate and marking it as a new fixed location information; when all display positions in the location sequence have been assigned, performing a recommendation operation based on the priority of each fixed location information and its corresponding assigned display position. This application can improve the accuracy of recommended information.

Description

Technical Field

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

[0002] Cloud computing is a computing model that distributes computing tasks across a large pool of computers, enabling various application systems to access computing power, storage space, and information services as needed. The network providing these resources is called the "cloud." From the user's perspective, resources in the "cloud" appear infinitely scalable, readily available, on-demand, and expandable, with payment based on usage.

[0003] 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.

[0004] Information recommendation is an important application of artificial intelligence. Among related technologies, the reordering module is the final stage of personalized recommendation in the recommendation system. The reordering module disperses the information generated by the sorting module according to preset rules and then presents it to the user to prevent the continuous presentation of highly repetitive information, which would lead to a lack of personalization in the recommendation. Summary of the Invention

[0005] This invention provides an information recommendation method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence, which can improve the accuracy of information recommendation and thus enhance user experience.

[0006] The technical solution of this invention is implemented as follows:

[0007] This invention provides an information recommendation method based on artificial intelligence, comprising:

[0008] Determine the correlation characteristics between the location information to be determined and all information in the information set;

[0009] The information set includes at least one of the following: determined location information; undetermined location information; the determined location information is information in the information set that has been assigned a display location in the location sequence, and the undetermined location information is information in the information set that has been assigned a display location in the location sequence;

[0010] Based on the correlation characteristics of each of the undetermined location information, the corresponding first click rate is determined;

[0011] The highest priority unassigned display position in the position sequence is assigned to the first pending position information with the highest click rate and marked as the new assigned position information;

[0012] When all display positions in the location sequence have been allocated, a recommendation operation is performed based on the priority of each allocated display position and the information of each predetermined position.

[0013] This invention provides an information recommendation device based on artificial intelligence, comprising:

[0014] The feature acquisition module is used to determine the correlation features between the location information to be determined and all information in the information set;

[0015] The information set includes at least one of the following: determined location information; undetermined location information; the determined location information is information in the information set that has been assigned a display location in the location sequence, and the undetermined location information is information in the information set that has been assigned a display location in the location sequence;

[0016] The click-through rate determination module is used to determine the corresponding first click-through rate based on the correlation characteristics of each of the pending location information.

[0017] The location allocation module is used to assign the unallocated and highest priority display location in the location sequence to the first pending location information with the highest click rate, and mark it as the new assigned location information;

[0018] The recommendation module is used to perform a recommendation operation based on the priority of each predetermined location and the corresponding assigned display location when all display locations in the location sequence have been allocated.

[0019] In the above scheme, the feature acquisition module is further configured to, before determining the correlation features between the location information to be determined and all information in the information set,

[0020] Obtain the basic characteristics of each piece of information in the database;

[0021] Based on the general fully connected parameters of the information database, the basic features are processed using a fully connected method to obtain the corresponding second click rate;

[0022] The information database is sorted in descending order based on the second click rate of each piece of information, and the top-ranked pieces of information in the descending order are selected to form the information set.

[0023] In the above scheme, the feature acquisition module is further configured to:

[0024] Obtain the features of each undetermined location information and the features of each determined location information in the information set;

[0025] For each of the aforementioned location information to be determined, the following processing is performed:

[0026] Attention encoding is performed on the features of each piece of information to be determined to obtain the correlation degree between the information to be determined and each piece of information in the information set;

[0027] Based on the correlation degree between the undetermined location information and each piece of information in the information set, the correlation characteristics of the undetermined location information are determined.

[0028] In the above scheme, the feature acquisition module is further configured to:

[0029] Perform linear transformation on the features of each piece of information in the information set to obtain the query vector, key vector, and value vector corresponding to each piece of information;

[0030] The query vector of the undetermined location information is multiplied by the key vector of each piece of information in the information set. The result of the multiplication is then normalized based on the maximum likelihood function to obtain the correlation degree between the undetermined location information and each piece of information in the information set.

[0031] In the above scheme, the feature acquisition module is further configured to:

[0032] The correlation is determined as the attention weight corresponding to the value vector of each piece of information;

[0033] The value vector is weighted based on the attention weights to obtain the association features of the undetermined location information based on attention encoding.

[0034] In the above scheme, the feature acquisition module is further configured to:

[0035] Obtain the basic features of each piece of information in the information set;

[0036] Obtain the location feature of each of the predetermined location information in the information set, wherein the location feature is used to represent the display location of the predetermined location information;

[0037] The basic features of the undetermined location information are used as the features of each undetermined location information.

[0038] The basic features and location features of the predetermined location information are fused together to obtain the features of the predetermined location information.

[0039] In the above scheme, the basic features include at least one of the following:

[0040] Basic attribute features used to characterize the basic information of the user to be recommended; interest tag features used to characterize the interest preferences of the user to be recommended; environmental features used to characterize the recommendation environment for recommending the information to the user to be recommended; category features used to characterize the category of the information; source features used to characterize the source of the information; and content features used to characterize the content of the information.

[0041] In the above scheme, the feature acquisition module is further configured to:

[0042] For each piece of information in the aforementioned information set, the following processing is performed:

[0043] Query multiple feature vectors corresponding to the information from a pre-established feature vector matrix;

[0044] Multiple feature vectors of the information are fused to obtain the basic features corresponding to the information.

[0045] In the above scheme, the click-through rate determination module is further used for:

[0046] Obtain the undetermined position corresponding to each of the undetermined position information, and obtain the fully connected parameters of the corresponding undetermined position;

[0047] Based on the fully connected parameters of each of the undetermined location information, the correlation features of each of the undetermined location information are processed by full connection to obtain the first click-through rate when the undetermined location information is displayed at the corresponding undetermined location.

[0048] In the above scheme, the feature acquisition module is further used to: determine new correlation features between new established location information and new undetermined location information;

[0049] The click-through rate determination module is further configured to: determine a corresponding new first click-through rate based on the correlation characteristics of each new location information to be determined;

[0050] The location allocation module is further configured to: allocate the unallocated and highest priority display location in the location sequence to the new first pending location information with the highest click-through rate, until all display locations in the location sequence are allocated.

[0051] In the above scheme, the first click-through rate of each undetermined location information in the information set is obtained by calling a click-through rate prediction model; the device further includes: a training module, used for:

[0052] Before determining the correlation characteristics between the established location information and the location information to be determined,

[0053] Obtain information sample sequences from the recommendation logs, and the true first click rate of each information sample in the information sample sequence;

[0054] Based on the information sample sequence and the corresponding actual first click rate, the click rate prediction model is trained for the kth display position of the position sequence to update the parameters of the click rate prediction model corresponding to the kth display position, and the parameters of the click rate prediction model corresponding to other display positions are kept unchanged during the update process.

[0055] Where k is an integer greater than or equal to 1, and the other display positions are the display positions in the position sequence other than the first k-1 display positions;

[0056] When the parameters corresponding to the kth display position in the click-through rate prediction model remain unchanged, the click-through rate prediction model is trained for the (k+1)th display position of the position sequence to update the parameters corresponding to the (k+1)th display position in the click-through rate prediction model, and the parameters corresponding to other display positions in the click-through rate prediction model are kept unchanged during the update process.

[0057] When the parameters corresponding to each display position in the click-through rate prediction model are determined, the click-through rate prediction model is considered to have completed training.

[0058] In the above scheme, the training module is further used for:

[0059] The following processing is performed during the training of the click-through rate prediction model for the k-th display position in the location sequence:

[0060] The click-through rate prediction model is used to determine the predicted first click-through rate of each information sample in the information sample sequence, excluding the first k-1 information samples.

[0061] The error between the predicted first click-through rate (CTR) and the actual first CTR for each information sample is determined, and the error is backpropagated in the CTR prediction model according to the loss function corresponding to the k-th display position.

[0062] Determine the parameter change value in the click-through rate prediction model corresponding to the kth display position when the loss function corresponding to the kth display position reaches its minimum value;

[0063] The parameters for the k-th display position in the click-through rate prediction model are updated based on the determined parameter change values.

[0064] This invention provides an electronic device, comprising:

[0065] Memory, used to store executable instructions;

[0066] The processor, when executing executable instructions stored in the memory, implements the information recommendation method based on artificial intelligence provided in this embodiment of the invention.

[0067] This invention provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the artificial intelligence-based information recommendation method provided in this invention.

[0068] The embodiments of the present invention have the following beneficial effects:

[0069] Based on the correlation features between the undetermined location information and all the determined location information in the information set, as well as the undetermined location information, it is possible not only to model the relationship between multiple pieces of information in the information sequence, but also to model the changes in user interests during the browsing of the information sequence. This results in a click-through rate prediction that is more in line with the actual recommendation scenario. Based on the click-through rate, the positions of multiple pieces of information in the information set are adjusted to maximize the accuracy of the recommendation information and the user experience. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the architecture of an artificial intelligence-based information recommendation system provided in an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of the structure of a server 200 that applies an artificial intelligence-based information recommendation method according to an embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram of model training for the information recommendation method based on artificial intelligence provided in an embodiment of the present invention;

[0073] Figure 4 This is a flowchart illustrating the training process of the information recommendation method based on artificial intelligence provided in this embodiment of the invention.

[0074] Figures 5A-5D This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in an embodiment of the present invention;

[0075] Figure 6 This is a recommendation system architecture diagram of the information recommendation method based on artificial intelligence provided in the embodiments of the present invention;

[0076] Figure 7 This is a schematic diagram illustrating the application of the information recommendation method based on artificial intelligence provided in this embodiment of the invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] 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.

[0079] In the following description, the terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0080] 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.

[0081] 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 invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0082] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be interpreted as follows.

[0083] 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.

[0084] 2) Personalized information recommendation: Based on users' interests and browsing behavior, recommend information that users may be interested in, including products, news, etc.

[0085] 3) Mask Position Awareness: Position awareness explicitly adds the position information corresponding to the sample to the sample's basic features. Through the mask mechanism, the position information is gradually released. For information with a fixed position, the mask of its position feature is 1, adding the position feature to the information's basic features. Based on this, the first click-through rate prediction is performed. For information with an undetermined position, the mask of its position feature is 0, masking the position feature. The first click-through rate prediction is performed only based on the basic features. Mask position awareness can be used to distinguish which information is information with a fixed position, thus revealing the position feature, and which information is information with an undetermined position, masking the position feature. In the process of determining and adjusting the display position of information, the position feature is dynamically added to the information's basic features, effectively taking into account the influence of earlier information on later information in the reordering model (click-through rate prediction model).

[0086] The following technical solutions exist for the reordering problem in related technologies: a shuffling module limits the multiple pieces of information generated by the sorting module according to preset rules and then presents them to the user. Its task is to prevent highly repetitive information from being presented to the user continuously, thereby enhancing the diversity of recommendations and effectively controlling the categories of information obtained in a single refresh. The shuffling module considers specific rules, recurrent neural networks, or pointers to reorder the multiple pieces of information generated by the sorting module. When used online, the reordering module shuffles all the information obtained from the sorting module according to rules and finally recommends the shuffled information to the user.

[0087] The rearrangement modules in related technologies mainly break down information based on a single rule, thus having the following drawbacks:

[0088] 1. Single-rule-based reordering models primarily control the information presented to users through preset strategies, essentially shuffling the information generated by the ranking model. The ranking model mainly learns click-through rates from individual samples and sorts them without considering the impact of relationships between samples. This results in the reordering failing to effectively learn global sequence features and further neglecting the positional relationships between sample sequences. It fails to fully utilize sample and positional features to learn the reordering model, weakening the ranking function of the reordering module. Recurrent neural network (RNN) models input the initial list sequentially; however, RNN-based methods have limited ability to model the interactions between information in the list. The feature information of the previous encoded item decreases with increasing encoding distance, leading to poor interpretability and susceptibility to the initial input sequence and encoding distance, ultimately affecting the ranking effect presented to the user.

[0089] 2. Currently, the rules for the reordering module are statically generated, meaning they are predefined before the user experience. This means they cannot change with the user's interests during browsing and do not explicitly consider the position of each piece of information. Consequently, they cannot fully simulate the changes in user interest after seeing information at the top of the list, which to some extent limits the user's positive experience with the information recommendation product.

[0090] To address the problems of inaccurate recommendations and inability to achieve personalized reordering in the aforementioned methods provided by related technologies, this invention provides an information recommendation method based on artificial intelligence, comprising a device, electronic device, and computer-readable storage medium. This method solves the problems of low recommendation accuracy and lack of personalization. It is a location-aware reordering method based on a mask mechanism, and improvements have been made in the following aspects based on the technical solutions in related technologies:

[0091] 1. While utilizing information features, location information is dynamically added. This not only considers the unique attributes of each piece of information but also effectively utilizes the location semantics of information that the user has already seen, making the recommended information more holistic. By gradually adding location information, the influence of earlier information on later information is explicitly learned. Specifically, the initial features at the bottom layer not only include basic features such as information-side features but also directly add the location features corresponding to the information. The location information is gradually released through the Mask mechanism, effectively taking into account the influence of earlier information on later information in the model, vividly simulating changes in user interests.

[0092] 2. Attention processing is applied to the information learned from the lower level. The feature expression of the current information after being affected by the feature vectors of other information is contained in the feature vector of each information. This fully learns the relationship between the vectors of multiple information and effectively avoids the problem of not being able to fully extract the relationship between information due to the limitation of encoding distance. Specifically, the interaction between the underlying features is learned through the transformer (Trm, Transformer) module to fully express the relationship between the features of multiple information.

[0093] 3. Add a fully connected module to the high level of the rearrangement model (click-through rate prediction model), and add a unique fully connected layer for each position to extract the position semantics learned from the lower level. The parameters of the fully connected layer are matched and adapted to the position information of that position. Since the fully connected layers do not share parameters, the feature vector of this information learns the position semantics at the same time as learning the click-through rate semantics. The position semantics are explicitly perceived at the high level through the fully connected module, effectively avoiding the loss of position information.

[0094] 4. When training the click-through rate prediction model, as location information is gradually added, only the cross-entropy loss caused by the lack of information corresponding to the location in the vectorized representation of the location information is calculated. This allows the testing process to be repeated until the vectorized representation of all location information has been learned. This not only avoids the calculation of invalid loss, but also focuses on the visible part of the vectorized representation of location information. The model trains to learn the click situation of the invisible part of the vectorized representation of location information. This enables the learned model to minimize the click-through rate loss that includes location semantics and maximize the reordering effect.

[0095] The following describes exemplary applications of the electronic device provided in the embodiments of the present invention. The electronic device provided in the embodiments of the present invention can be implemented as a server. Exemplary applications when the electronic device is implemented as a server will be described below.

[0096] A server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0097] Artificial intelligence cloud services are generally also known as AIaaS (AI as a Service). This is currently a mainstream service model for artificial intelligence platforms. Specifically, AIaaS platforms break down several common AI services and provide them as independent or packaged services in the cloud. This service model is similar to opening an AI-themed marketplace: all developers can access and use one or more AI services provided by the platform through API interfaces. Some experienced developers can also use the AI ​​framework and AI infrastructure provided by the platform to deploy and maintain their own dedicated cloud AI services.

[0098] See Figure 1 , Figure 1This is a schematic diagram of the architecture of an artificial intelligence-based information recommendation system provided in an embodiment of the present invention. The information recommendation system can support various information recommendation scenarios, such as news recommendations, product recommendations, video recommendations, etc. Depending on the application scenario, the information can be news, actual products, video information, images and text, etc. In the information recommendation system, terminal 400 connects to server 200 via network 300. Network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of both. Responding to a recommendation information request received from terminal 400, the functions of the information recommendation system are implemented based on various modules within server 200. The second click-through rate (CTR) determination unit 25521 in the click-through rate determination module 2552 in the server determines the second CTR of multiple pieces of information in the information database 500, and sorts the multiple pieces of information in the information database in descending order based on the second CTR to obtain K pieces of information in descending order. Then, the first CTR determination unit 25522 in the click-through rate determination module 2552 determines the first CTR of the first K pieces of information in descending order. Then, the position allocation module 2553 allocates the position of the information with the highest first CTR, and repeats the above process until all information is allocated. Finally, the allocation result is sent to the terminal through the recommendation module 2554 so that the terminal presents the information according to the allocation result.

[0099] See Figure 2 , Figure 2 This is a schematic diagram of the structure of a server 200 applying an artificial intelligence-based information recommendation method according to an embodiment of the present invention. The server 200 includes at least one processor 210, a memory 250, and at least one network interface 220. The various components in the server 200 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, ... Figure 2 The general labeled all buses as Bus System 240.

[0100] 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.

[0101] 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.

[0102] 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.

[0103] 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;

[0104] 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).

[0105] In some embodiments, the artificial intelligence-based information recommendation device provided in this invention can be implemented in software. Figure 2 An AI-based information recommendation device 255 stored in memory 250 is shown. It includes multiple modules of the information recommendation system. The modules can be software in the form of programs and plug-ins, including the following software modules: feature acquisition module 2551, click rate determination module 2552, position allocation module 2553, recommendation module 2554, and training module 2555. These modules are logically related and can be arbitrarily combined or further divided according to the functions they implement. The functions of each module will be described below.

[0106] The following will illustrate the information recommendation method based on artificial intelligence provided by the embodiments of the present invention, with reference to exemplary applications and implementations of the information recommendation system provided in the embodiments of the present invention. The information recommendation system includes a training phase and an application phase.

[0107] The following section describes the model used in the AI-based information recommendation method provided in this embodiment of the invention, as well as the training process of the model.

[0108] See Figure 3 , Figure 3This is a schematic diagram of the model training of the information recommendation method based on artificial intelligence provided in this embodiment of the invention. The click-through rate prediction model for reordering includes a feature fusion module, a feature association module, a fully connected module, and an output module. For information i, the feature fusion module is used to obtain the feature vectors of each dimension of the corresponding information i, and to fuse the feature vectors of the corresponding information i in multiple dimensions, i.e., to add them together, to obtain the basic feature F of the corresponding information i. That is, the basic feature is obtained based on at least one of the feature vectors of multiple dimensions. Therefore, the obtained basic feature includes at least one of the feature vectors of multiple dimensions. In addition, the feature fusion module also needs to consider the position information of information i. When information i is determined to be displayed at position j, the position feature P of position j is visible, and the position feature of position j is added to the basic feature F of information i to obtain the feature of information i. When the display position of information i is not determined, the basic feature of information i is directly used as the feature of information i. The feature association module is actually a self-attention network. Through the self-attention network, the association features between information and each piece of information can be obtained. Then, the association features of information are fully connected through the fully connected module to obtain the first click-through rate of the corresponding information.

[0109] In some embodiments, the first click-through rate (CTR) of each pending location information in the information set is obtained by calling a CTR prediction model. Before determining the correlation features between the established location information and the pending location information, the CTR prediction model can be trained using the following method: obtaining the information sample sequence and the true first CTR of each information sample in the information sample sequence from the recommendation log; based on the information sample sequence and the corresponding true first CTR, training the CTR prediction model for the k-th display location in the location sequence to update the parameters of the CTR prediction model corresponding to the k-th display location, and fixing the CTR prediction model during the update process. The parameters corresponding to other display positions remain unchanged; where k is an integer greater than or equal to 1, and other display positions are the display positions in the position sequence excluding the first k-1 display positions; when the parameters corresponding to the k-th display position in the click-through rate prediction model remain unchanged, the click-through rate prediction model continues to be trained for the (k+1)-th display position in the position sequence to update the parameters corresponding to the (k+1)-th display position in the click-through rate prediction model, and the parameters corresponding to other display positions in the click-through rate prediction model remain unchanged during the update process; when the parameters corresponding to each display position in the click-through rate prediction model are determined, the click-through rate prediction model training is considered complete.

[0110] As an example, see Figure 4 , Figure 4This is a training flowchart of the information recommendation method based on artificial intelligence provided in this embodiment of the invention. During the training process, the vectorized representation of each position is trained, that is, the actual training results in the positional features of each position. The positional features of each position obtained by training are applied to the later online prediction stage. The initial input is the first K pieces of information generated by the sorting module. The positional features of these K pieces of information are set to unknown. Then, the first click-through rate prediction is performed through the position-aware reordering algorithm. As the network trains, the positional features of each position are gradually determined. In the process of gradually determining, the sum of the corresponding positional features and the basic features of the information is used as input to the network for repeated training until all positional features are trained, thus completing the training.

[0111] In some embodiments, the above-mentioned training of the click-through rate (CTR) prediction model for the k-th display position of the location sequence to update the parameters of the CTR prediction model corresponding to the k-th display position can be implemented through the following scheme: The following processing is performed during the training of the CTR prediction model for the k-th display position of the location sequence: Using the CTR prediction model, the predicted first CTR of each information sample in the information sample sequence, excluding the first k-1 information samples, is determined; the error between the predicted first CTR and the actual first CTR of each information sample is determined, and the error is backpropagated in the CTR prediction model according to the loss function corresponding to the k-th display position to determine the parameter change value of the CTR prediction model corresponding to the k-th display position when the loss function corresponding to the k-th display position reaches its minimum value; the parameters of the CTR prediction model corresponding to the k-th display position are updated according to the determined parameter change value.

[0112] As an example, see Figure 3First, the model learns the positional features of the first position. A sequence containing K pieces of information is input into the model to obtain the first click-through rate (CTR) for each piece of information at its corresponding position. Specifically, the information sequence is the sequence that is actually presented to the user and either generates actual clicks. Assuming K=3, the actual information sequence is presented such that the first piece of information is presented in the first position, the second piece of information in the second position, and the third piece of information in the third position. The CTR obtained by the model is actually a 3x3 CTR matrix, representing the CTR of each piece of information at each position. However, during training, only the diagonal CTR is used to calculate the loss function because the relationship between information and position in an information sequence with actual clicks is a diagonal relationship (the first piece of information is presented in the first position, the second piece of information in the second position, and the third piece of information in the third position). The CTR in other cases is merely a byproduct of training. For example, the first... The first click-through rate (CTR) of the information at the third position is useless because there is no corresponding real click data. After obtaining the three first CTRs S on the diagonal, the error between each S and the corresponding real click data is calculated. The real click data is set as follows: the real CTR with click behavior is 1, and the real CTR without click behavior is 0. After adding the first CTRs of each piece of information, the loss is calculated to determine the positional features of the first position (the parameters corresponding to the first position). Then, the parameters of the first position are opened. Then, given the parameters of the first position, the first CTR of the second piece of information in the sequence presented in the second position and the first CTR of the third piece of information in the sequence presented in the third position are predicted. Similarly, the loss function is calculated based on the error between the two S and the corresponding real click data to determine the positional features of the second position (the parameters corresponding to the second position). Then, the parameters of the second position are opened. This process continues until the parameters of all positions are trained and determined, thus completing the entire training process.

[0113] The application of the model in the artificial intelligence-based information recommendation method provided in this embodiment of the invention will be described next. See [link to relevant documentation]. Figure 5A , Figure 5A This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in an embodiment of the present invention, which will be combined with... Figure 5A Steps 101-104 shown are explained.

[0114] In step 101, the server determines the correlation characteristics between the location information to be determined and all information in the information set.

[0115] In some embodiments, the information set includes at least one of the following: determined location information; the undetermined location information; the determined location information is information in the information set that has been assigned a display location in a location sequence, and the undetermined location information is information in the information set that has an unassigned display location in a location sequence, wherein the number of display locations is the same as the number of information.

[0116] As an example, all the information in the information set forms an information sequence. In the initial state, each piece of information in the information sequence is a piece of information with a pending position. Each piece of information in the information sequence has a pending position, which is used to determine the first click rate of each piece of information at the pending position. For the information sequence in the initial state, the information in the information sequence is sorted according to the second click rate, which is the click rate output by the fine ranking model. Alternatively, the information sequence is obtained after adjustment based on the second click rate and the scattering strategy. The scattering strategy adjusts the position of the information in the information sequence sorted according to the second click rate so that the position distribution of the adjusted information conforms to the scattering strategy. The scattering strategy can specify the distribution distance between information with the same label.

[0117] In some embodiments, before determining the correlation features between the location information to be determined and all information in the information set, the following technical solution may also be performed: obtain the basic features of each piece of information in the information database; perform full connection processing on the basic features based on the general full connection parameters of the information database to obtain the corresponding second click rate; perform descending sorting processing on the information database based on the second click rate of each piece of information, and select multiple pieces of information ranked first in the descending sorting results to form an information set.

[0118] As an example, the role of the fine-ranking model is to sort the recalled information based on the second click-through rate (CTR), and select the top-ranked information from the descending sort results to form a re-ranked information set. The information in the information set is then sorted in descending order according to the second CTR. The descending sort result determines the pending position of each piece of information in the initial state, so as to output the first CTR of each piece of information at the pending position. For example, if the information set contains information A, information B, and information C, and the second CTR of information A is less than that of information B, and the second CTR of information B is less than that of information C, then the pending position of information C is the highest priority display position in the position sequence, and the pending position of information A is the lowest priority display position in the position sequence. A higher priority position represents the position that users will see first when browsing, such as a relatively high ranking position on the same page (relative to a lower priority position).

[0119] As an example, the correlation features between information can obtain global information about the entire information sequence, so that the first click rate is not predicted based on independent information, but also takes into account the mutual influence between various pieces of information. This is more in line with the actual browsing and clicking behavior of users. Users are not browsing a single piece of information, but multiple pieces of information are presented at the same time.

[0120] As an example, the correlation features between fixed location information and undetermined location information can dynamically characterize the user's preference changes when the fixed location information is perceived by the user before the undetermined location information. For example, for information A, information B, and information C mentioned above, when predicting the first click-through rate, the position of information A has already been determined as the highest priority position, that is, information A is fixed location information. When information A, which is presented in the highest priority position, is viewed by the user, the user's preference interest will inevitably change, even if it is a very small change. Through the correlation features between fixed location information (information A) and undetermined location information, the first click-through rate of the undetermined location information in the corresponding undetermined position can be predicted more accurately, which approximately simulates the changes in user interest behavior when browsing in reality, so that the information after the layout is adjusted can conform to the dynamic changes in user interests.

[0121] based on Figure 5A See Figure 5B , Figure 5B This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in this embodiment of the invention. In step 101, determining the correlation features between the location information to be determined and all information in the information set can be achieved through steps 1011-1012, combining... Figure 5B Steps 1011-1012 shown are explained.

[0122] In step 1011, the features of each undetermined location information and the features of each determined location information are obtained from the information set.

[0123] In some embodiments, the acquisition of features of each undetermined location information and features of each determined location information in the information set can be achieved through the following technical solutions: acquiring basic features of each information in the information set; acquiring position features of each determined location information in the information set, wherein the position features are used to represent the display position of the determined location information; using the basic features of the undetermined location information as features of each undetermined location information; and fusing the basic features and position features of the determined location information to obtain the features of the determined location information.

[0124] As an example, for the information sequence in the initial state, information A, information B, and information C all belong to information with undetermined positions. The features of information A, information B, and information C are all basic features of information A, information B, and information C. For the information sequence in the process of adjustment, for example, the position of information A has been determined to be the first position with the highest priority, that is, information A is information with a fixed position. The features of information B and information C are still the basic features of information B and information C. The feature of information A is the result of the fusion processing of the basic features of information A and the position feature of the first position with the highest priority, that is, the result of the addition processing. This is so that when determining the association features that represent the correlation between each piece of information with undetermined position and all information, it can approximately simulate the changes in the user's interest behavior when actually browsing, so that the information after the arrangement is adjusted can conform to the dynamic changes in the user's interests.

[0125] In some embodiments, the acquisition of the basic features of each piece of information in the information set can be achieved by the following technical solution: performing the following processing on each piece of information in the information set: querying multiple feature vectors corresponding to the information from a pre-established feature vector matrix; fusing the multiple feature vectors of the information to obtain the basic features of the corresponding information.

[0126] As an example, the feature vector fusion process can be to sum these feature vectors to obtain the feature corresponding to each piece of information.

[0127] In some embodiments, the basic features include at least one of the following: basic attribute features for characterizing the basic information of the user to be recommended; interest tag features for characterizing the interest preferences of the user to be recommended; environmental features for characterizing the recommendation environment for recommending information to the user to be recommended; category features for characterizing the category of information; source features for characterizing the source of information; and content features for characterizing the content of information.

[0128] As an example, the basic features are obtained based on at least one of multiple dimensions of features. That is, the obtained basic features include at least one of multiple dimensions of features. The information recommendation method provided in this embodiment of the invention is a recommendation scheme for the same user. Therefore, it can obtain user-side features, such as the user's gender, age, region, and family structure, information-side features, such as information category and information source, and environment-side features, such as the user's network environment and network device. The feature vectors of at least one of the features in these dimensions are fused to obtain the basic features. That is, by obtaining features in various dimensions, the first click rate of information is modeled to achieve more personalized recommendations.

[0129] In some embodiments, feature engineering can encode the feature data of information to obtain the feature vector of information. Specifically, this can be achieved through the following technical solutions: converting the feature values ​​of feature data into feature indices and hashing the feature indices to obtain feature index codes; hashing the feature names of feature data to obtain feature name codes, and combining the feature name codes and feature index codes to obtain the feature vector of information.

[0130] As an example, during hash encoding, it is necessary to calculate the index corresponding to the feature data and encode the index. In machine learning processing, to facilitate the implementation of related algorithms, it is often necessary to convert the label data (usually a string) into an integer index, or to restore the integer index to the corresponding label after calculation. A converter can encode a series of categorical attribute features (or labels) to make them numerical. The index range starts from 0. This process can make the corresponding features indexed, so that some algorithms that cannot accept categorical features can be used, and improve the efficiency of machine learning algorithms such as decision trees. If the input is numerical data, it can be converted into character data and then encoded. The feature name actually represents the feature category, such as the attribute of the feature. The feature data can be mapped to a 64-bit hash space. The first 16 bits of the 64-bit hash space represent the feature name, and the last 48 bits represent the feature index.

[0131] In step 1012, the following processing is performed for each piece of information to be determined: attention encoding is performed on the features of each piece of information to be determined to obtain the correlation degree between the information to be determined and each piece of information in the information set; based on the correlation degree between the information to be determined and each piece of information in the information set, the associated features of the information to be determined are determined.

[0132] In some embodiments, the attention encoding process performed on the features of each piece of information to be determined in step 1012 to obtain the correlation degree between the information to be determined and each piece of information in the information set can be achieved by the following technical solution: performing linear transformation on the features of each piece of information in the information set to obtain the query vector, key vector and value vector corresponding to each piece of information; performing dot product processing on the query vector of the information to be determined and the key vector of each piece of information in the information set; and performing normalization processing on the dot product processing result based on the maximum likelihood function to obtain the correlation degree between the information to be determined and each piece of information in the information set.

[0133] In some embodiments, the determination of the association features of the undetermined location information based on the association degree between the undetermined location information and each piece of information in the information set in step 1012 can be achieved by the following technical solution: determining the association degree as the attention weight of the value vector corresponding to each piece of information; performing weighted processing on the value vector based on the attention weight to obtain the association features of the undetermined location information based on attention encoding processing.

[0134] As an example, attention encoding is performed on the features of each piece of information in the information set to obtain the attention score of each piece of information for a certain piece of information (the information to be determined). Since the attention score of each piece of information for a certain piece of information (the information to be determined) is obtained, we can obtain the influence of the information at the determined location (position features and basic features) on the information at the determined location, as well as the influence of all the information at the determined location (basic features) on the information at the determined location. Since the location of the information at the determined location is already determined, it is not necessary to obtain the attention score of each piece of information for the determined location. The attention score can be used as the correlation between the information (the information at the determined location) and each piece of information, including the information itself. In this way, we can not only take into account the mutual influence between information, but also approximately characterize the real-time features of the information used by the user to determine the first click rate when browsing information in real time, so that the adjusted information can conform to the actual dynamic changes of the user's interests.

[0135] As an example, calculating the attention score of each piece of information for a given piece of information mainly relies on the dot product of the query vector and the key vector, as well as maximum likelihood processing. After the dot product and before the maximum likelihood processing, the result of the dot product can be divided by the square root of the dimension of the key vector, which makes the gradient more stable. Maximum likelihood processing is implemented using the softmax function, which normalizes the attention scores of all information, resulting in positive scores that sum to 1. This attention score determines the contribution of each piece of information in the information set to encoding that information. The query vector, key vector, and value vector of each piece of information's features are obtained by processing the query vector, key vector, and value vector of the information. The feature vector of the information is obtained by linear transformation, which is to multiply the feature vector of the information by the query parameters, key parameters, and value parameters. The query parameters, key parameters, and value parameters are obtained through model training. To simplify the training process, the query parameters, key parameters, and value parameters can be uniformly set to 1. That is, the linear transformation process is to multiply the feature vector of the information by 1, so that the resulting query vector, key vector, and value vector are the same and are all feature vectors. The attention score of each piece of information relative to a certain piece of information is weighted and summed with the value vector of the corresponding piece of information to obtain the association feature of the information based on attention encoding.

[0136] Attention encoding can learn the information dependencies within a set of information, thereby capturing the internal information relationships within the set. In long-distance dependency problems, since attention encoding requires each piece of information to calculate an attention score with all other pieces of information, the maximum path length is only 1 regardless of the distance between the pieces of information, thus enabling the capture of long-distance dependencies.

[0137] In step 102, the server determines the corresponding first click rate based on the relevance characteristics of each pending location information.

[0138] based on Figure 5A See Figure 5C , Figure 5C This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in this embodiment of the invention. In step 102, determining the corresponding first click-through rate based on the relevance characteristics of each pending location information can be achieved through steps 1021-1022, combining... Figure 5C Steps 1021-1022 are shown below for explanation.

[0139] In step 1021, the undetermined position corresponding to each undetermined position information is obtained, and the fully connected parameters of the corresponding undetermined position are obtained.

[0140] In step 1022, based on the fully connected parameters of each pending location information, the correlation features of each pending location information are processed by full connection to obtain the first click-through rate when the pending location information is displayed in the corresponding pending location.

[0141] As an example, as the network deepens, the click-through rate prediction model learns increasingly abstract concepts, resulting in poorer interpretability. After the feature association module, the underlying location information (location features) is lost. Therefore, multiple independent fully connected layers are used in the final layer to determine the first click-through rate for different locations, such as... Figure 3 As shown in the fully connected module, each position has a corresponding fully connected layer, and the parameters in the fully connected layer are used to learn the feature information unique to that position.

[0142] In step 103, the server assigns the highest-priority unassigned display position in the position sequence to the first pending position information with the highest click-through rate and marks it as the new assigned position information.

[0143] As an example, the display positions in the position sequence are assigned to the information to be determined in descending order of display position priority. For information A, information B, and information C, in the first click-through rate prediction process (before position allocation), the first click-through rate P of information A, information B, and information C in their corresponding undetermined positions is predicted. A PB and P C The method for determining the pending positions is as described above. If the second click-through rate of information A is less than that of information B, and the second click-through rate of information B is less than that of information C, then the pending position of information C is the highest priority display position in the position sequence, and the pending position of information A is the lowest priority display position in the position sequence. If P A Greater than P B P B Greater than P C If so, the highest-priority unassigned display position in the position sequence will be assigned to information A, and A will be marked as the new assigned position information.

[0144] In some embodiments, after performing step 103, the following technical solution may also be performed: determining new correlation features between new fixed location information and new pending location information; determining the corresponding new first click-through rate based on the correlation features of each new pending location information; allocating the unallocated and highest priority display position in the location sequence to the new pending location information with the highest first click-through rate, until all display positions in the location sequence are allocated.

[0145] As an example, the location allocation in this scheme is actually a process of gradual location allocation. After the first prediction process mentioned above, the unallocated and highest priority display position in the location sequence is reassigned as the pending position for information B, and the lowest priority display position in the location sequence is reassigned as the pending position for information C, for the second click-through rate prediction. During the second click-through rate prediction process, the first click-through rates of information B and information C will be output for the second location allocation. That is, after each location allocation, the pending position information will be reordered in descending order based on the latest first click-through rate prediction result to determine the pending positions of each pending position information in the next first click-through rate prediction. For pending positions, the pending positions with higher priority are still assigned to the pending position information with higher click-through rates.

[0146] In step 104, when all display positions in the location sequence have been allocated, the server performs a recommendation operation based on the information of each assigned location and the priority of the corresponding assigned display position.

[0147] In some embodiments, specifically in a recommendation scenario where multiple pieces of information are requested to be displayed on a single page, the number of pieces of information in the information set is consistent with the number of display positions in the position sequence. That is, the recommendation request carries the number of display positions on the page, the backend sends the corresponding number of pieces of information and the priority of the corresponding display positions, and the SDK displays the information in the display position corresponding to the priority of the information.

[0148] In some embodiments, in response to an information recommendation request, multiple pieces of information are requested at once and displayed on multiple pages. That is, the number of requested pieces of information is greater than the number of display positions on the page. In this case, the software development component displays the multiple pieces of information received and the priority of the corresponding display positions on the multiple pages in sequence, with the priority of the display position on the page that is displayed first being higher than the priority of the display position on the page that is displayed later.

[0149] See Figure 5D , Figure 5D This is a flowchart illustrating an artificial intelligence-based information recommendation method provided in an embodiment of the present invention. In step 200, the server receives an information sequence composed of information from an information set. In step 201, the server determines the correlation characteristics between the pending location information and all information in the information set. In step 202, the server determines the corresponding first click-through rate based on the correlation characteristics of each pending location information. In step 203, the server assigns the unassigned display position with the highest priority in the location sequence to the pending location information with the highest first click-through rate and marks it as a new assigned location information. If there are still unassigned display positions, the server re-executes steps 201-203. If there are no unassigned display positions, step 204 is executed, whereby the server performs a recommendation operation based on the priority of each assigned location information and the corresponding assigned display position.

[0150] The following will describe an exemplary application of the artificial intelligence-based information recommendation method provided in this embodiment of the invention in a real-world application scenario.

[0151] The AI-based information recommendation method provided in this invention can be applied to news recommendation systems. The model used in this AI-based information recommendation method is a Mask-based location-aware reordering model (click-through rate prediction model). This model adjusts the position of information based on the initial sorted list, explicitly using the positional information of the information to improve the initial sorted list. Simultaneously with learning the basic feature vector (e.g., information-side feature vector), it learns a positional feature vector. The learning of positional features is a gradual release process using the Mask method, simulating the real physical process of refreshing news from front to back. Here, the positional feature vector represents the underlying real positional information. After learning through the TRM module (feature association module), the features of each piece of information contain cross-features of other information. To avoid the underlying positional... To address information loss, a fully connected module is introduced at a higher level to extract location information. This module is used to perceive the location information at higher levels, thereby further extracting relevance semantics and location semantics. Here, location semantics is an explicit expression of the semantics of the higher-level information vector. The loss function of this model only considers the loss of samples with unknown location vectors, thus learning more fully the unique location features of each location. The re-ranking model of the information recommendation method based on artificial intelligence provided in this embodiment of the invention effectively learns the relationship features between information and effectively simulates the changes in user interests after seeing the first information. This is of great significance for re-ranking and helps to improve the recommendation effect. In the recommendation system, the recalled articles generate ranked information through the ranking module. The ranked information is then re-ranked by the information recommendation method based on artificial intelligence provided in this embodiment of the invention and finally recommended to the user.

[0152] See Figure 6 , Figure 6 This is a recommendation system architecture diagram of an AI-based information recommendation method provided in this embodiment of the invention. The news recommendation system comprises four modules: a user profile service module, a recall service module, a ranking service module, and a re-ranking service module. The profile module mainly includes information on past user clicks and basic registration information, accumulating and storing users' long-term interests, short-term interests, and basic information to provide a basis for recall and ranking. The recall module is responsible for initially searching for information that users may be interested in from massive amounts of information. The ranking and shuffling modules rank the recalled information and present it to users according to certain rules. The recall algorithm acts as a filter from the information pool to user interest information, providing basic data for subsequent ranking and re-ranking modules. The ranking module scores the recalled information using a corresponding click-through rate prediction model. The re-ranking module, as the connecting module between the ranking module and the final user recommendation, further optimizes the ranked information as a whole. A poorly performing re-ranking module affects the order in which users see information, greatly weakening the effectiveness of the re-ranking service in the recommendation system and thus affecting the user experience.

[0153] The reordering service of related technologies scatters sorted information based on specific rules, dispersing highly similar information to the user's position in the first swipe. However, this service only considers the impact of information repetition and diversity on the final ranking. This only works well on average for the overall user, but is not optimal for each individual user. The reordering module directly scatters the sorted information, so the ranking result only considers the relevance between a single piece of information and the user, without considering the relevance between multiple pieces of information for the same user. Moreover, the user's interests can change dynamically due to the influence of the information at the top. The information ultimately recommended to the user by the reordering module in related technologies cannot dynamically simulate changes in user interests, affecting the accuracy of recommendations and user experience.

[0154] To address these issues, this invention provides an AI-based information recommendation method, a mask-based location-aware reordering method. Located within the reordering module, this method gradually releases the location information of the information through a mask mechanism, allowing for the reordering of the sorted information. When a terminal requests information, the recall module thoroughly mines information that the user might be potentially interested in. The recalled information is then reordered by the sorting module's sorting model and the reordering module, ultimately being recommended to the user. After the user clicks and reads the information according to their preferences, the actual click data is reported to the recommendation system. Different modules iterate and optimize based on the reported user data, continuously improving the user experience.

[0155] The training process of the information recommendation method based on artificial intelligence provided in this embodiment of the invention is as follows: Figure 4 As shown, the initial input information sequence to the initialization model is a descending sequence of Top-K information generated by the sorting module, which is actually presented to the user. First, the positions of these K pieces of information are set to unknown, and then they are rearranged multiple times through a position-aware rearrangement algorithm. As the network is trained, the vectorized representation of the position is gradually set to fixed. At this time, the corresponding position vector and the basic feature vector of the information side are added together and input into the rearrangement model for repeated training until the vectorized representation of all positions is determined, thus completing the training.

[0156] Suppose a news client receives K recommended articles in response to a single refresh request. The ranking module's sorting model recommends these K articles to the user based on the second click-through rate prediction. At this point, the position of each article is not globally optimal. We assume the positions of these K articles are adjustable, setting them as unknown, and their corresponding position vectors as empty (i.e., masked). In the network, a masking mechanism is introduced to implement both the fixed and adjustable states of the information presented at a given position. Figure 3As shown, when the initial K pieces of information are obtained for the first time (here, the illustration shows 3 pieces of information, K=3), the position information for all locations is unknown. Therefore, all position information is masked at this point. The underlying input consists only of the basic feature vector F, a combination of information-side features and user-side features. After the network runs once, the information with the highest click-through rate is selected and displayed in the first position (the highest priority position among all positions). At this point, the information for the first position becomes known, and the position vector P of the first position is no longer masked. The position vectors of the other K-1 positions are still masked. This process iterates layer by layer until the information displayed for all positions is known. At this point, the order of the K pieces of information is also adjusted. Let the information be denoted as... 0 represents the first information sorting, i.e., the initial input sorting. At this time, the position of each piece of information is uncertain. A vector can be used to record whether the position of the information is fixed at this time. This refers to whether the information to be displayed at this location has been determined. When the value is 1, it means that the information to be displayed at this position is fixed, and a position feature vector can be introduced for each position. ,like Figure 3 As shown in P, when the information to be displayed at this position is determined, the position feature vector is visible; otherwise, the position feature vector is masked. During the initial input, the information to be displayed at all positions is uncertain. Set all to 0, meaning all position vectors are masked.

[0157] The basic feature vector and the location feature vector are processed by the TRM module (feature association module) to extract the relationship between feature vectors of multiple information. A self-attention network can be used. The input vector is processed by a fully connected layer to obtain three vectors: query vector Q, key vector K, and value vector V. The output is the weighted sum of all value vectors in V, as shown in Equation 1:

[0158] (1);

[0159] The weights are calculated from the query vector and each key vector of the feature vector. The calculation method consists of three steps: 1) Calculate and compare the similarity between Q and K; 2) Normalize the obtained similarity based on the maximum likelihood function; 3) For the calculated weights, perform a weighted summation on all value vectors to obtain the attention vector (association feature). At this time, the attention vector (association feature T) output by the attention network contains the feature expression of the current information feature vector after being affected by the feature vectors of other information, where d is the initial dimension of the input feature vector.

[0160] As the network deepens, the learned information becomes increasingly abstract, leading to poorer interpretability. After passing through the TRM module, significant loss of location information in the underlying design is observed. Therefore, multiple independent fully connected layers are used in the final layer to extract click-through rate and location features from different locations, such as... Figure 3 As shown in the fully connected module, each position has a corresponding fully connected layer. The parameters in the fully connected layer are used to learn the feature information specific to that position. Each fully connected layer learns not only click-through rate semantics but also position semantics. Here, position semantics is an explicit expression of the semantics of the higher-level information vector. The vector composed of the location feature vector and the basic feature vector is transformed by the self-attention network to output the associated feature vector. The parameters for the fully connected layer set here are... Then, the first click-through rate is calculated based on Formula 2:

[0161] (2);

[0162] in, As a bias term, the fully connected layer at each position outputs K click-through rates, and the diagonal portion is taken as the click-through rate corresponding to each piece of information in the final output:

[0163] At the bottom layer, the mask for the position is set based on whether the position of the information is fixed. Information with a fixed position will not participate in the position adjustment of the information. Therefore, the loss calculation for the first click rate of this position is meaningless. Therefore, the loss function for each training is calculated based on formula (3):

[0164] (3);

[0165] Where k is the starting point of the index for information that is not currently fixed. This represents the true value of the user's first click rate.

[0166] In the application phase of the model, see Figure 7 , Figure 7 This is a schematic diagram illustrating the model application of the information recommendation method based on artificial intelligence provided in this embodiment of the invention. The information is sorted according to a first click-through rate, and the information with the highest first click-through rate is selected as the first item. Simultaneously... In Set to 1, at which point the position vector P of the first position is visible, and the information is sorted as follows: 1 represents the result after the first round of sorting, thus completing the first round of sorting. The result of the first round of sorting is then input into the network. At this point, the position vector P of the first position is already visible. Before attention processing, the feature of the information at the first position is the sum of the basic feature vector S and the position vector P. After passing through the feature association module, each piece of information obtains a corresponding association feature T. After passing through the Mask position-aware rearrangement network, the position of the first piece of information is now fixed. Each piece of information other than the information displayed at the first position will again receive a corresponding first click rate S. Therefore, the remaining K-1 pieces of information are sorted to obtain the adjusted information order. ,in The fixed position in the first round, The fixed position in the second round, in the third round of sorting In , The value is set to 1, and so on, until the positions of all K pieces of information are fixed. The fine-tuning algorithm for reordering then ends, and the resulting information order is... .

[0167] The training and testing processes of the model differ as follows: For each training data (information sequence), multiple rounds of training can be conducted based on the information sequence, gradually releasing the Mask position vector to dynamically simulate the change in user interest after seeing the information at the beginning. The loss function for each training data is shown in formula (4):

[0168] (4);

[0169] Each user's data will be trained K times, with the corresponding number of undetermined positions being 0, 1, 2, 3, ..., K-1.

[0170] For each piece of test data, such as Figure 7 As shown, no real data is used in the prediction process; the prediction process will follow... Figure 7 The displayed information at positions 1, 2, and 3 is gradually determined according to the positional order shown, until the information displayed at each position is finalized. After training, the model parameters are stored. When used online, the predicted second click-through rate (CTR) is obtained through the ranking module. Then, the top K CTR results for each request are identified, and the features of these K CTR results are input into the model to simulate... Figure 7 The prediction process shown yields the final adjusted K pieces of information, which are then presented to the user.

[0171] The information recommendation method based on artificial intelligence provided in this invention is applied to a news recommendation system. It optimizes the re-ranking method of the re-ranking module by using a mask mechanism to dynamically incorporate location information into the network structure. This effectively learns the relationship features between information elements while also simulating changes in user interest after seeing top-ranked information. The sum of the feature vectors from the information side and the location feature vectors is fed into the TRM module (feature association module). This facilitates the extraction of feature information between different information vectors and better utilizes the relationships between information elements. Through a unique fully connected layer set for each location, location information can be perceived at a higher level, improving the accuracy of the information recommended by the re-ranking module. The final output is more representative of user interests, thus enhancing the user experience.

[0172] The following description continues to illustrate the exemplary structure of the artificial intelligence-based information recommendation device 255 provided in the embodiments of the present invention as a software module. In some embodiments, such as... Figure 2 As shown, the software modules in the AI-based information recommendation device 255 stored in the memory 250 may include: a feature acquisition module 2551, used to determine the correlation features between the pending location information and all information in the information set; wherein, the information set includes at least one of the following: fixed location information; pending location information; the fixed location information is information in the information set that has been allocated display positions in the location sequence, and the pending location information is information about display positions in the location sequence to be allocated in the information set; a click-through rate determination module 2552, used to determine the corresponding first click-through rate based on the correlation features of each pending location information; a location allocation module 2553, used to allocate the unallocated display position with the highest priority in the location sequence to the pending location information with the highest first click-through rate, and mark it as new fixed location information; and a recommendation module 2554, used to perform a recommendation operation based on the priority of each fixed location information and the corresponding allocated display position when all display positions in the location sequence have been allocated.

[0173] In some embodiments, the click-through rate determination module 2552 is further configured to: obtain the basic features of each piece of information in the information database before determining the correlation features between the determined location information and the location information to be determined; perform full connection processing on the basic features based on the general full connection parameters of the information database to obtain the corresponding second click-through rate; perform descending sorting processing on the information database based on the second click-through rate of each piece of information, and select multiple pieces of information ranked first in the descending sorting results to form an information set.

[0174] In some embodiments, the feature acquisition module 2551 is further configured to: acquire the features of each pending location information in the information set, and the features of each fixed location information; perform the following processing for each pending location information: perform attention encoding processing on the features of each pending location information to obtain the correlation degree between the pending location information and each piece of information in the information set; and determine the associated features of the pending location information based on the correlation degree between the pending location information and each piece of information in the information set.

[0175] In some embodiments, the feature acquisition module 2551 is further configured to: perform linear transformation processing on the features of each piece of information in the information set to obtain the query vector, key vector and value vector corresponding to each piece of information; perform dot product processing on the query vector of the pending location information and the key vector of each piece of information in the information set, and perform normalization processing on the dot product processing result based on the maximum likelihood function to obtain the correlation degree between the pending location information and each piece of information in the information set.

[0176] In some embodiments, the feature acquisition module 2551 is further configured to: determine the correlation degree as the attention weight of the value vector corresponding to each piece of information; and perform weighted processing on the value vector based on the attention weight to obtain the correlation features of the undetermined position information based on attention encoding processing.

[0177] In some embodiments, the feature acquisition module 2551 is further configured to: acquire the basic features of each piece of information in the information set; acquire the position features of each piece of information with a fixed position in the information set, wherein the position features are used to represent the display position of the information with a fixed position; use the basic features of the information with a pending position as the features of each piece of information with a pending position; and fuse the basic features and position features of the information with a fixed position to obtain the features of the information with a fixed position.

[0178] In some embodiments, the basic features include at least one of the following: basic attribute features for characterizing the basic information of the user to be recommended; interest tag features for characterizing the interest preferences of the user to be recommended; environmental features for characterizing the recommendation environment for recommending information to the user to be recommended; category features for characterizing the category of information; source features for characterizing the source of information; and content features for characterizing the content of information.

[0179] In some embodiments, the feature acquisition module 2551 is further configured to: perform the following processing for each piece of information in the information set: query multiple feature vectors corresponding to the information from a pre-established feature vector matrix; and perform fusion processing on the multiple feature vectors of the information to obtain the basic features of the corresponding information.

[0180] In some embodiments, the click-through rate determination module 2552 is further configured to: obtain the pending location corresponding to each pending location information, and obtain the fully connected parameters of the corresponding pending location; based on the fully connected parameters of each pending location information, perform fully connected processing on the correlation features of each pending location information to obtain the first click-through rate when the pending location information is displayed in the corresponding pending location.

[0181] In some embodiments, the feature acquisition module 2551 is further configured to: determine new correlation features between new fixed location information and new undetermined location information; the click-through rate determination module 2552 is further configured to: determine a corresponding new first click-through rate based on the correlation features of each new undetermined location information; and the location allocation module 2553 is further configured to: allocate the unallocated and highest priority display position in the location sequence to the new undetermined location information with the highest first click-through rate, until all display positions in the location sequence are allocated.

[0182] In some embodiments, the first click-through rate of each undetermined location information in the information set is obtained by calling a click-through rate prediction model; the apparatus 255 further includes: a training module 2555, configured to: before determining the correlation features between the determined location information and the undetermined location information, Obtain information sample sequences and the true first click-through rate (CTR) of each information sample in the information sample sequence from the recommendation logs. Based on the information sample sequences and the corresponding true first CTRs, train the CTR prediction model for the k-th display position in the location sequence to update the parameters of the CTR prediction model for the k-th display position, while keeping the parameters of the CTR prediction model for other display positions unchanged during the update process. Here, k is an integer greater than or equal to 1, and other display positions are the display positions in the location sequence excluding the first k-1 display positions. When the parameters of the CTR prediction model for the k-th display position remain unchanged, continue training the CTR prediction model for the (k+1)-th display position in the location sequence to update the parameters of the CTR prediction model for the (k+1)-th display position, while keeping the parameters of the CTR prediction model for other display positions unchanged during the update process. When the parameters of the CTR prediction model for each display position are determined, the CTR prediction model training is considered complete.

[0183] In some embodiments, the training module 2555 is further configured to: perform the following processing during the training of the click-through rate prediction model for the k-th display position of the position sequence: using the click-through rate prediction model, determine the predicted first click-through rate for each information sample in the information sample sequence excluding the first k-1 information samples; determine the error between the predicted first click-through rate and the actual first click-through rate for each information sample, and backpropagate the error in the click-through rate prediction model according to the loss function corresponding to the k-th display position, so as to determine the parameter change value in the click-through rate prediction model corresponding to the k-th display position when the loss function corresponding to the k-th display position reaches its minimum value; update the parameters in the click-through rate prediction model corresponding to the k-th display position according to the determined parameter change value.

[0184] This invention 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 artificial intelligence-based information recommendation method provided in this invention. For example, ... Figures 5A-5D The example shown is an information recommendation method based on artificial intelligence.

[0185] 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.

[0186] 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.

[0187] 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).

[0188] 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.

[0189] In summary, through the embodiments of the present invention, based on the correlation features between the undetermined location information and all determined location information in the information set, as well as the undetermined location information, it is possible not only to model the relationship between multiple pieces of information in the information sequence, but also to model the changes in user interests during the browsing of the information sequence. This results in a click-through rate prediction result that is more in line with the actual recommendation scenario, and the position of multiple pieces of information in the information set is adjusted based on the click-through rate, thereby maximizing the accuracy of the recommendation information and the user experience.

[0190] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

Claims

1. An information recommendation method based on artificial intelligence, characterized in that, include: Determine the correlation characteristics between the location information to be determined and all information in the information set; The information set includes at least one of the following: determined location information; undetermined location information; the determined location information is information in the information set that has been assigned a display location in the location sequence, and the undetermined location information is information in the information set that has been assigned a display location in the location sequence; Obtain the undetermined position corresponding to each of the undetermined position information, and obtain the fully connected parameters of the corresponding undetermined position; based on the fully connected parameters of each of the undetermined position information, perform fully connected processing on the correlation features of each of the undetermined position information to obtain the first click-through rate when the undetermined position information is displayed in the corresponding undetermined position; The highest priority unassigned display position in the position sequence is assigned to the first pending position information with the highest click rate and marked as the new assigned position information; When all display positions in the location sequence have been allocated, a recommendation operation is performed based on the priority of each allocated display position and the information of each predetermined position.

2. The method according to claim 1, characterized in that, Before determining the correlation characteristics between the location information to be determined and all information in the information set, the method further includes: Obtain the basic characteristics of each piece of information in the database; Based on the general fully connected parameters of the information database, the basic features are processed using a fully connected method to obtain the corresponding second click rate; The information database is sorted in descending order based on the second click rate of each piece of information, and the top-ranked pieces of information in the descending order are selected to form the information set.

3. The method according to claim 1, characterized in that, The determination of the correlation features between the location information to be determined and all information in the information set includes: Obtain the features of each undetermined location information and the features of each determined location information in the information set; For each of the aforementioned location information to be determined, the following processing is performed: Attention encoding is performed on the features of each piece of information to be determined to obtain the correlation degree between the information to be determined and each piece of information in the information set; Based on the correlation degree between the undetermined location information and each piece of information in the information set, the correlation characteristics of the undetermined location information are determined.

4. The method according to claim 3, characterized in that, The step of performing attention encoding on the features of each piece of information to be determined, to obtain the correlation between the information to be determined and each piece of information in the information set, includes: Perform linear transformation on the features of each piece of information in the information set to obtain the query vector, key vector, and value vector corresponding to each piece of information; The query vector of the undetermined location information is multiplied by the key vector of each piece of information in the information set. The result of the multiplication is then normalized based on the maximum likelihood function to obtain the correlation degree between the undetermined location information and each piece of information in the information set.

5. The method according to claim 3, characterized in that, The step of determining the association features of the undetermined location information based on the association degree between the undetermined location information and each piece of information in the information set includes: The correlation is determined as the attention weight corresponding to the value vector of each piece of information; The value vector is weighted based on the attention weights to obtain the association features of the undetermined location information based on attention encoding.

6. The method according to claim 3, characterized in that, The process of obtaining the features of each undetermined location information and the features of each determined location information in the information set includes: Obtain the basic features of each piece of information in the information set; Obtain the location feature of each of the predetermined location information in the information set, wherein the location feature is used to represent the display location of the predetermined location information; The basic features of the undetermined location information are used as the features of each undetermined location information. The basic features and location features of the predetermined location information are fused together to obtain the features of the predetermined location information.

7. The method according to claim 6, characterized in that, The basic features include at least one of the following: Basic attribute features used to characterize the basic information of the user to be recommended; interest tag features used to characterize the interest preferences of the user to be recommended; environmental features used to characterize the recommendation environment for recommending the information to the user to be recommended; category features used to characterize the category of the information; source features used to characterize the source of the information; and content features used to characterize the content of the information.

8. The method according to claim 6, characterized in that, The acquisition of the basic features of each piece of information in the information set includes: For each piece of information in the aforementioned information set, the following processing is performed: Query multiple feature vectors corresponding to the information from a pre-established feature vector matrix; Multiple feature vectors of the information are fused to obtain the basic features corresponding to the information.

9. The method according to claim 1, characterized in that, The method further includes: Determine new correlation characteristics between new established location information and new undetermined location information; Based on the correlation characteristics of each new location information to be determined, a corresponding new first click rate is determined; The highest priority unassigned display position in the position sequence is assigned to the new pending position information with the highest first click rate, until all display positions in the position sequence are assigned.

10. The method according to claim 1, characterized in that, The first click-through rate of each undetermined location information in the information set is obtained by calling a click-through rate prediction model; Before determining the correlation characteristics between the known location information and the location information to be determined, the method further includes: Obtain information sample sequences from the recommendation logs, and the true first click rate of each information sample in the information sample sequence; Based on the information sample sequence and the corresponding actual first click rate, the click rate prediction model is trained for the kth display position of the position sequence to update the parameters of the click rate prediction model corresponding to the kth display position, and the parameters of the click rate prediction model corresponding to other display positions are kept unchanged during the update process. Where k is an integer greater than or equal to 1, and the other display positions are the display positions in the position sequence other than the first k-1 display positions; When the parameters corresponding to the kth display position in the click-through rate prediction model remain unchanged, the click-through rate prediction model is trained for the (k+1)th display position of the position sequence to update the parameters corresponding to the (k+1)th display position in the click-through rate prediction model, and the parameters corresponding to other display positions in the click-through rate prediction model are kept unchanged during the update process. When the parameters corresponding to each display position in the click-through rate prediction model are determined, the click-through rate prediction model is considered to have completed training.

11. The method according to claim 10, characterized in that, The step of training the click-through rate (CTR) prediction model for the k-th display position in the location sequence to update the parameters of the CTR prediction model corresponding to the k-th display position includes: The following processing is performed during the training of the click-through rate prediction model for the k-th display position in the location sequence: The click-through rate prediction model is used to determine the predicted first click-through rate of each information sample in the information sample sequence, excluding the first k-1 information samples. The error between the predicted first click-through rate (CTR) and the actual first CTR for each information sample is determined, and the error is backpropagated in the CTR prediction model according to the loss function corresponding to the k-th display position. Determine the parameter change value in the click-through rate prediction model corresponding to the kth display position when the loss function corresponding to the kth display position reaches its minimum value; The parameters for the k-th display position in the click-through rate prediction model are updated based on the determined parameter change values.

12. An information recommendation device based on artificial intelligence, characterized in that, include: The feature acquisition module is used to determine the correlation features between the location information to be determined and all information in the information set; The information set includes at least one of the following: determined location information; undetermined location information; the determined location information is information in the information set that has been assigned a display location in the location sequence, and the undetermined location information is information in the information set that has been assigned a display location in the location sequence; The click-through rate determination module is used to obtain the pending positions corresponding to each of the pending position information, and to obtain the fully connected parameters of the corresponding pending positions; based on the fully connected parameters of each of the pending position information, the module performs fully connected processing on the correlation features of each of the pending position information to obtain the first click-through rate when the pending position information is displayed in the corresponding pending position; The location allocation module is used to assign the unallocated and highest priority display location in the location sequence to the first pending location information with the highest click rate, and mark it as the new assigned location information; The recommendation module is used to perform a recommendation operation based on the priority of each predetermined location and the corresponding assigned display location when all display locations in the location sequence have been allocated.

13. The apparatus according to claim 12, characterized in that, The feature acquisition module is also used to acquire the basic features of each piece of information in the information database before determining the correlation features between the location information to be determined and all information in the information set; Based on the general fully connected parameters of the information database, the basic features are processed by full connection to obtain the corresponding second click rate; based on the second click rate of each piece of information, the information database is sorted in descending order, and multiple pieces of information ranked first in the descending order are selected to form the information set.

14. The apparatus according to claim 12, characterized in that, The feature acquisition module is further configured to acquire the features of each pending location information and the features of each determined location information in the information set; and to perform the following processing on each pending location information: perform attention encoding processing on the features of each pending location information to obtain the correlation degree between the pending location information and each piece of information in the information set; Based on the correlation degree between the undetermined location information and each piece of information in the information set, the correlation characteristics of the undetermined location information are determined.

15. The apparatus according to claim 14, characterized in that, The feature acquisition module is further configured to: perform linear transformation processing on the features of each piece of information in the information set to obtain the query vector, key vector and value vector corresponding to each piece of information; perform dot product processing on the query vector of the undetermined location information and the key vector of each piece of information in the information set, and perform normalization processing on the dot product processing result based on the maximum likelihood function to obtain the correlation degree between the undetermined location information and each piece of information in the information set.

16. The apparatus according to claim 14, characterized in that, The feature acquisition module is further configured to determine the correlation degree as the attention weight of the value vector corresponding to each piece of information; and to perform weighted processing on the value vector based on the attention weight to obtain the correlation features of the undetermined position information based on attention encoding processing.

17. The apparatus according to claim 14, characterized in that, The feature acquisition module is further configured to acquire the basic features of each piece of information in the information set; acquire the position features of each piece of information with a fixed position in the information set, wherein the position features are used to represent the display position of the information with a fixed position; use the basic features of the information with a pending position as the features of each piece of information with a pending position; and fuse the basic features and position features of the information with a fixed position to obtain the features of the information with a fixed position.

18. The apparatus according to claim 17, characterized in that, The basic features include at least one of the following: Basic attribute features used to characterize the basic information of the user to be recommended; interest tag features used to characterize the interest preferences of the user to be recommended; environmental features used to characterize the recommendation environment for recommending the information to the user to be recommended; category features used to characterize the category of the information; source features used to characterize the source of the information; and content features used to characterize the content of the information.

19. The apparatus according to claim 17, characterized in that, The feature acquisition module is further configured to perform the following processing for each piece of information in the information set: query multiple feature vectors corresponding to the information from a pre-established feature vector matrix; and perform fusion processing on the multiple feature vectors of the information to obtain the basic features corresponding to the information.

20. The apparatus according to claim 12, characterized in that, The feature acquisition module is also used to determine new correlation features between new established location information and new undetermined location information; The click-through rate determination module is further configured to determine a new first click-through rate based on the correlation characteristics of each new pending location information. The location allocation module is further configured to allocate the unallocated and highest priority display location in the location sequence to the new first pending location information with the highest click-through rate, until all display locations in the location sequence have been allocated.

21. The apparatus according to claim 12, characterized in that, The first click-through rate of each undetermined location information in the information set is obtained by calling a click-through rate prediction model, and the device further includes: The training module is used to obtain an information sample sequence and the true first click-through rate (CTR) of each information sample in the recommendation log before determining the correlation features between the known and unknown location information. Based on the information sample sequence and the corresponding true CTR, the CTR prediction model is trained for the k-th display position of the location sequence to update the parameters of the CTR prediction model corresponding to the k-th display position, while keeping the parameters of the CTR prediction model corresponding to other display positions unchanged during the update process; where k is an integer greater than or equal to 1. The other display positions are the display positions in the position sequence other than the first k-1 display positions; when the parameters corresponding to the k-th display position in the click-through rate prediction model remain unchanged, the click-through rate prediction model is trained again for the k+1-th display position in the position sequence to update the parameters corresponding to the k+1-th display position in the click-through rate prediction model, and the parameters corresponding to the other display positions in the click-through rate prediction model remain unchanged during the update process; when the parameters corresponding to each display position in the click-through rate prediction model are determined, the click-through rate prediction model is considered to have completed training.

22. The apparatus according to claim 21, characterized in that, The training module is further configured to perform the following processing during the training of the click-through rate prediction model for the k-th display position of the position sequence: using the click-through rate prediction model, determining the predicted first click-through rate for each information sample in the information sample sequence excluding the first k-1 information samples; determining the error between the predicted first click-through rate and the actual first click-through rate for each information sample, and backpropagating the error in the click-through rate prediction model according to the loss function corresponding to the k-th display position, to determine the parameter change value in the click-through rate prediction model corresponding to the k-th display position when the loss function corresponding to the k-th display position reaches its minimum value; and updating the parameters in the click-through rate prediction model corresponding to the k-th display position according to the determined parameter change value.

23. An electronic device, characterized in that, include: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the information recommendation method based on artificial intelligence as described in any one of claims 1 to 11.

24. A computer-readable storage medium, characterized in that, It stores executable instructions for implementing the artificial intelligence-based information recommendation method according to any one of claims 1 to 11 when executed by a processor.

Images