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

Abstract

The application provides an information recommendation method and device based on artificial intelligence, an electronic device and a computer readable storage medium; relates to artificial intelligence technology, and the method comprises the following steps: acquiring ranking features of to-be-recommended information; performing mapping processing on the ranking features based on multiple recommendation tasks to obtain first feature mapping results corresponding to each recommendation task; performing mapping processing on the ranking features based on multiple feature dimensions to obtain second feature mapping results corresponding to each feature dimension; taking any one recommendation task as a target recommendation task, performing attention processing on the first feature mapping results and the second feature mapping results corresponding to each feature dimension based on the first feature mapping results and the second feature mapping results corresponding to each feature dimension to obtain a recommendation index of the to-be-recommended information corresponding to the target recommendation task; and performing a recommendation operation based on the recommendation index of the to-be-recommended information corresponding to each recommendation task. Through the application, multiple recommendation tasks can be efficiently implemented.

Description

TECHNICAL FIELD

[0001] The present application relates to artificial intelligence technology, and in particular to an information recommendation method and device based on artificial intelligence, an electronic device, and a computer readable storage medium. BACKGROUND

[0002] Artificial intelligence (AI) is the use of digital computers or digital computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0003] With the development of information technology and the Internet industry, information overload has become a challenge for people to process information, and personalized recommendation systems effectively alleviate such problems by more accurately understanding user intent. However, the way of click rate estimation through linear ranking model or deep ranking model in the related technology still recommends information to users that are ranked in the front but are not really liked by users. SUMMARY

[0004] The embodiments of the present application provide an information recommendation method and device based on artificial intelligence, an electronic device, and a computer readable storage medium, which can improve the recommendation efficiency.

[0005] The technical solutions of the embodiments of the present application are as follows:

[0006] The embodiments of the present application provide an information recommendation method based on artificial intelligence, comprising:

[0007] Obtaining the ranking features of the information to be recommended;

[0008] Mapping processing the ranking features based on multiple recommendation tasks to obtain a first feature mapping result corresponding to each recommendation task;

[0009] Mapping processing the ranking features based on multiple feature dimensions to obtain a second feature mapping result corresponding to each feature dimension;

[0010] Taking any one of the recommendation tasks as a target recommendation task, performing attention processing based on the first feature mapping result and the second feature mapping result corresponding to each feature dimension to obtain a recommendation index of the information to be recommended corresponding to the target recommendation task;

[0011] Based on the recommendation index of the information to be recommended corresponding to each recommendation task, performing a recommendation operation.

[0012] The embodiments of the present application provide an information recommendation device based on artificial intelligence, comprising:

[0013] obtain a sorting feature of the information to be recommended;

[0014] perform mapping processing on the sorting feature based on a plurality of recommendation tasks to obtain a first feature mapping result corresponding to each of the recommendation tasks;

[0015] The mapping module is further configured to perform mapping processing on the sorting feature based on a plurality of feature dimensions to obtain a second feature mapping result corresponding to each of the feature dimensions.

[0016] The attention module is configured to take any one of the recommendation tasks as a target recommendation task, perform attention processing based on the first feature mapping result and the second feature mapping result corresponding to each of the feature dimensions to obtain a recommendation index of the information to be recommended corresponding to the target recommendation task.

[0017] The recommendation module is configured to perform a recommendation operation based on the recommendation index of the information to be recommended corresponding to each of the recommendation tasks.

[0018] In the above scheme, the obtaining module is further configured to: obtain discrete features corresponding to a plurality of feature types of the recommendation information, wherein the plurality of feature types include user features, information features, and environment features; and perform the following processing on the discrete features of any one of the feature types: perform weighted processing on a plurality of discrete values in the discrete features by taking hidden vectors corresponding to the plurality of discrete values as weights to obtain embedding dimension values of the discrete features corresponding to an embedding dimension, and combine the embedding dimension values of the discrete features corresponding to a plurality of embedding dimensions into a dense feature of the feature type; and perform splicing processing on the dense features of the plurality of feature types to obtain the sorting feature corresponding to the information to be recommended.

[0019] In the above scheme, before obtaining the sorting feature of the information to be recommended, the obtaining module is further configured to obtain the recommendation information by at least one of the following methods: obtaining information similar in content to historical browsing information of a to-be-recommended object and having a content similarity degree not less than a content similarity threshold to serve as the information to be recommended; and obtaining a plurality of reference to-be-recommended objects similar in user portrait to the to-be-recommended object and having a portrait similarity degree not less than a portrait similarity threshold, and obtaining historical browsing information corresponding to each of the reference to-be-recommended objects to serve as the information to be recommended.

[0020] In the above scheme, the mapping module is further configured to: perform mapping processing on the sorting feature by a feature full connection layer in an index prediction network corresponding to each recommendation task to obtain a first feature mapping result corresponding to each recommendation task; and map the sorting feature to a feature space corresponding to each expert network by the multiple expert networks to obtain a second feature mapping result of a feature dimension associated with each expert network.

[0021] In the above scheme, the attention module is further configured to: perform attention processing on the first feature mapping result and the second feature mapping result corresponding to each feature dimension to obtain an attention processing result corresponding to the target recommendation task; and perform recommendation index prediction processing on the information to be recommended by a prediction full connection layer of an index prediction network corresponding to the target recommendation task in the multi-recommendation task model in combination with the attention processing result corresponding to the target recommendation task to obtain a recommendation index of the information to be recommended corresponding to the target recommendation task.

[0022] In the above scheme, the attention module is further configured to: multiply the first feature mapping result and a transpose of the second feature mapping result corresponding to each feature dimension to obtain a multiplication result corresponding to each feature dimension; perform linear processing on the linear processing result corresponding to each feature dimension; perform maximum likelihood processing on the linear processing result corresponding to each feature dimension, multiply the maximum likelihood processing result and the second feature mapping result corresponding to each feature dimension to obtain an attention processing result corresponding to each feature dimension; and perform splicing processing on the attention processing result corresponding to each feature dimension to obtain the attention processing result corresponding to the target recommendation task.

[0023] In the above scheme, the attention module is further configured to: perform linear transformation processing on the attention processing results of the multiple recommendation tasks to obtain query vectors, key vectors, and value vectors corresponding to the multiple recommendation tasks after the splicing processing on the attention processing result corresponding to each feature dimension to obtain the attention processing result corresponding to the target recommendation task; determine a co-attention processing result corresponding to the target recommendation task based on the query vectors, the key vectors, and the value vectors corresponding to the multiple recommendation tasks; and update the attention processing result corresponding to the target recommendation task based on the co-attention processing result corresponding to the target recommendation task.

[0024] In the above scheme, the attention module is further configured to: perform linear transformation processing on the attention processing result of the target recommendation task to obtain a query vector, a key vector and a value vector corresponding to the attention processing result of the target recommendation task; perform linear transformation processing on the attention processing result of at least one other recommendation task to obtain a query vector, a key vector and a value vector corresponding to the attention processing result of each of the other recommendation tasks; and wherein the other recommendation task is a recommendation task other than the target recommendation task among the plurality of recommendation tasks.

[0025] In the above scheme, the attention module is further configured to: for each of the other recommendation tasks, perform the following processing: perform dot product processing on the query vector of the attention processing result of the target recommendation task and the key vector of the attention processing result of the other recommendation task to obtain a relevance between the target recommendation task and the other recommendation task; and perform multiplication processing on the relevance and the value vector of the attention processing result of the other recommendation task to obtain a co-attention processing result of the other recommendation task corresponding to the target recommendation task.

[0026] In the above scheme, the attention module is further configured to: when the number of the other recommendation tasks is one, perform addition processing on the co-attention processing result of the other recommendation task for the target recommendation task and the attention processing result of the target recommendation task, and update the attention processing result corresponding to the target recommendation task based on the obtained addition processing result; and when the number of the other recommendation tasks is a plurality, perform addition processing on the co-attention processing result of each of the other recommendation tasks corresponding to the target recommendation task and the attention processing result of the target recommendation task, and update the attention processing result corresponding to the target recommendation task based on the obtained addition processing result.

[0027] In the above scheme, the recommendation module is further configured to: obtain an aggregation rule corresponding to the plurality of recommendation tasks; based on an operator included in the aggregation rule, perform calculation processing on the recommendation indicators respectively corresponding to the plurality of recommendation tasks to obtain comprehensive recommendation indicators of the plurality of to-be-recommended information respectively corresponding to the plurality of recommendation tasks; determine a recommendation order of the plurality of to-be-recommended information according to the comprehensive recommendation indicators of the plurality of to-be-recommended information respectively corresponding to the plurality of recommendation tasks, and perform a recommendation operation corresponding to a to-be-recommended object based on the recommendation order of the plurality of to-be-recommended information.

[0028] In the foregoing scheme, the apparatus further comprises a training module configured to, before the ranking feature of the information to be recommended is acquired, perform the following processing in each iteration training process of the multi-recommendation task model: performing forward propagation of a sample of information to be recommended in a feature extraction network of the multi-recommendation task model, a plurality of expert networks, and an index prediction network corresponding to the plurality of recommendation tasks one by one, to perform recommendation index prediction on the sample of information to be recommended, to obtain predicted recommendation indexes of the sample of information to be recommended corresponding to each of the recommendation tasks; wherein the index prediction network comprises a feature full connection layer corresponding to the plurality of recommendation tasks one by one, an attention network, and a prediction full connection layer; determining an error between the predicted recommendation indexes of the sample of information to be recommended corresponding to each of the recommendation tasks and prelabeled recommendation indexes corresponding to each of the recommendation tasks; aggregating the errors corresponding to each of the recommendation tasks according to the loss function to obtain an aggregated error, and performing back propagation of the aggregated error in the multi-recommendation task model to determine a parameter change value of the multi-recommendation task model when the loss function reaches a minimum value, and updating parameters of the multi-recommendation task model based on the parameter change value.

[0029] An electronic device is provided in an embodiment of the application, comprising:

[0030] A memory configured to store executable instructions;

[0031] A processor configured to execute the executable instructions stored in the memory to implement the information recommendation method based on artificial intelligence provided in the embodiments of the application.

[0032] A computer readable storage medium is provided in an embodiment of the application, which stores executable instructions, and is configured to be executed by a processor to implement the information recommendation method based on artificial intelligence provided in the embodiments of the application.

[0033] The embodiments of the application have the following beneficial effects:

[0034] By performing mapping processing based on feature dimensions and based on tasks on the ranking features, and performing attention processing on the respective mapping results, the recommendation tasks are directly associated with different feature dimensions, the recommendation indexes corresponding to the plurality of recommendation tasks are more accurately predicted, and thus accurate personalized recommendation is realized for the plurality of recommendation tasks. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 FIG. 1 is a structural schematic diagram of a multi-recommendation task model in the related art;

[0036] Figure 2 FIG. 2 is a structural schematic diagram of an information recommendation system based on artificial intelligence provided in an embodiment of the application;

[0037] Figure 3FIG. 1 is a structural schematic diagram of an electronic device provided by an embodiment of the present application;

[0038] Figures 4A-4D FIG. 2 is a flow schematic diagram of an information recommendation method based on artificial intelligence provided by an embodiment of the present application;

[0039] Figure 5 FIG. 3 is a framework schematic diagram of an information recommendation method based on artificial intelligence provided by an embodiment of the present application;

[0040] Figures 6A-6B FIG. 4 is a structural schematic diagram of a multi-recommendation task model provided by an embodiment of the present application;

[0041] Figure 7 FIG. 5 is a schematic diagram of a feature extraction network of an information recommendation method based on artificial intelligence provided by an embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by those skilled in the art without making creative efforts fall within the scope of protection of the present application.

[0043] In the following description, "some embodiments" are described, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0044] In the following description, the terms "first\second" are only to distinguish similar objects, and do not represent a specific order of the objects, and it can be understood that "first\second" can be interchanged in a specific order or sequence as allowed, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0045] 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 the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0046] The related data collection and processing in the embodiments of the present application 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 authorization of laws and regulations and the personal information subject.

[0047] Before the embodiments of the present application are described in further detail, terms and names involved in the embodiments of the present application are explained, and the terms and names involved in the embodiments of the present application are applicable to the following explanations.

[0048] 1) Recommendation system: The recommendation system is a tool for automatically connecting users and information, which can help users find information of interest in an information overload environment, and can also push information to users interested in them.

[0049] 2) News personalized recommendation: According to the interest characteristics and reading behavior of the user, the user is recommended the news of interest.

[0050] 3) Click rate estimation module: According to the feature list of the item and the ranking model, the click rate of the item in the item candidate set is calculated.

[0051] 4) Multi-task learning (MTL, Multi-task Learning): The process of simultaneously learning multiple related tasks, using useful information contained in multiple learning tasks to help learn more accurate learners for each task.

[0052] 5) Attention mechanism: It is essentially derived from human visual attention mechanism. When people perceive things through vision, they usually do not watch from the beginning to the end every time, but often observe and pay attention to a specific part according to their needs. When people find that the part they want to observe often appears in a certain part of the scene, people will learn to focus their attention on that part in the future when similar scenes appear again, so as to focus more attention on the useful part.

[0053] 6) To-be-recommended object, i.e. the target of information recommendation. Since the medium of information presentation is a terminal, the target of information recommendation is a user operating the corresponding terminal, so in the following, "object" and "user" are described equally. It can be understood that the user here can be a natural person who can operate the terminal, or a robot program running in the terminal that can simulate human behavior.

[0054] 7) Recommended information: i.e. information that can be sent to the terminal for presentation to the user (to-be-recommended object) of the corresponding terminal for recommendation.

[0055] 8) Area under the receiver operating characteristic curve (AUC, Area Under Curve): as a model evaluation standard, the area under the receiver operating characteristic curve is obtained by summing the areas of each part.

[0056] The ranking model in the related art adopts a multi-recommendation task model, see Figure 1 , Figure 1is a structural schematic diagram of a multi-recommendation task model in the related art, the multi-recommendation task model includes a shared bottom structure and a single-door multi-expert network hybrid structure, the single-door multi-expert network hybrid structure includes a gating network corresponding to each recommendation task (for example, a gating network A corresponding to a recommendation task A and a gating network B corresponding to a recommendation task B), a tower network corresponding to each (for example, a tower network A corresponding to the recommendation task A and a tower network B corresponding to the recommendation task B), and an output layer corresponding to each (for example, an output layer A corresponding to the recommendation task A and an output layer B corresponding to the recommendation task B), two recommendation tasks share a plurality of expert networks (expert networks 0-2) and a feature sharing layer, and the correlation between the multi-recommendation tasks is learned through the gating mode of the shared bottom structure and the single-door multi-expert network hybrid structure. The gating network of each recommendation task realizes selective use of the plurality of expert networks through different weights output, the gating networks of different recommendation tasks can learn different modes of combining a plurality of expert networks, but the following problems still exist.

[0057] I. Although the structure of the gating network of the multi-recommendation task model has a certain alleviating effect on the conflict caused by the difference between the recommendation tasks, in actual application, a situation similar to a seesaw still exists, that is, the effect of one recommendation task is improved, and the effect of another recommendation task is reduced, and it is still difficult to achieve an ideal state. The applicant found that the reason for implementing the embodiments of the present application is that the correlation between the multi-recommendation tasks is not sufficiently learned, the focus on and attention to the feature subsets of each recommendation task are not sharp enough, and the sharing granularity between the multi-recommendation tasks is not fine enough.

[0058] II. The gating network in the multi-recommendation task model is essentially outputting different weights to realize selective use of the plurality of expert networks, but the output of the gating network is learned based on the original feature input, and the output of the gating network does not form a direct association with the expert network. The applicant found that during the implementation of the embodiments of the present application that the attention distribution of different recommendation tasks to the expert network can be more directly and effectively learned based on the expert network itself and the recommendation task itself.

[0059] The embodiments of the present application provide an information recommendation method and device based on artificial intelligence, electronic equipment and computer readable storage medium, which can more accurately and directly associate a plurality of recommendation tasks with different feature dimensions to more accurately predict the indicators corresponding to the plurality of recommendation tasks, thereby realizing accurate personalized recommendation and improving information recommendation efficiency. The following describes an exemplary application of the electronic equipment provided by the embodiments of the present application, which can be a server. The following will illustrate an exemplary application when the electronic equipment is implemented as a server.

[0060] Referring to Figure 2 ,Figure 2 is a structural schematic diagram of an information recommendation system based on artificial intelligence provided by an embodiment of the present application, for realizing a news application, a terminal 400 connects an application server 200-2 through a network 300, the network 300 can be a wide area network or a local area network, or a combination of the two. The training server 200-1 pushes the trained multi-recommendation task model to the application server 200-2, and the terminal 400 sends a user request to the application server 200-2, the application server 200-2 obtains the to-be-recommended information and the corresponding sorting features from the recall system 500 through the acquisition module 2551, and maps the sorting features through the mapping module 2552, and obtains the recommendation index of each recommendation task through the attention module 2553 Attention processing of the mapping result, so that the recommendation module 2554 determines the to-be-recommended information with a high ranking according to the recommendation index and returns it to the terminal 400 for presentation.

[0061] In some embodiments, the training server 200-1 and the application server 200-2 can be independent physical servers, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and artificial intelligence platforms. The terminal 400 can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiments of the present application.

[0062] Referring to Figure 3 , Figure 3 is a structural schematic diagram of an electronic device provided by an embodiment of the present application, taking the electronic device as an application server 200-2, Figure 2 The application server 200-2 shown in the figure includes at least one processor 210, a memory 250, at least one network interface 220 and a user interface 230. Each component in the terminal 200 is coupled together through a bus system 220. It can be understood that the bus system 220 is used to realize the connection and communication between the components. In addition to the data bus, the bus system 220 also includes a power bus, a control bus and a state signal bus. However, for the sake of clarity, all kinds of buses are marked as a bus system 240 in Figure 3 .

[0063] The processor 210 can be an integrated circuit chip that has a processing capability of signals, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor.

[0064] The memory 250 can be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical drives, etc. The memory 250 optionally includes one or more storage devices remotely located from the processor 210 in a physical location.

[0065] The memory 250 includes volatile memory or non-volatile memory, and can also include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), and the volatile memory can be random access memory (RAM). The memory 250 described in the embodiments of the present application is intended to include any suitable type of memory.

[0066] In some embodiments, the 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, which are exemplarily illustrated below.

[0067] The operating system 251 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0068] The network communication module 252 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.

[0069] In some embodiments, the artificial intelligence-based information recommendation apparatus provided by the embodiments of the present application can be realized in a software manner, Figure 3 An artificial intelligence-based information recommendation apparatus 255 stored in the memory 250 is shown, which can be software in the form of programs and plug-ins, etc., including the following software modules: an acquisition module 2551, a mapping module 2552, an attention module 2553, a recommendation module 2554, and a training module 2555, these modules are logical, and thus can be combined or further split according to the functions implemented, and the functions of each module will be described below.

[0070] The information recommendation method based on artificial intelligence provided in this application embodiment will be described in conjunction with the exemplary application and implementation of the application server 200-2 provided in the embodiments of this application.

[0071] See Figure 6A , Figure 6A This is a schematic diagram of the structure of the multi-recommendation task model provided in the embodiments of this application. If applied to a news recommendation system, assuming there are two recommendation tasks, there can be multiple recommendation tasks, such as recommendation task A and recommendation task B (recommendation tasks are in...). Figure 6A In this context, we'll refer to them as "tasks" (or simply "tasks"). Recommendation task A is a click-through rate prediction task, and recommendation task B is a reading duration prediction task. Recommendation task A has a corresponding metric prediction network A, and recommendation task B has a corresponding metric prediction network B. Metric prediction network A includes an attention network A, a fully connected feature layer A (which can be replaced with a complete deep neural network), and a fully connected prediction layer A (which can be replaced with a complete deep neural network, such as a tower network A). The fully connected prediction layer A outputs the recommendation metric for recommendation task A. Metric prediction network B includes an attention network B, a fully connected feature layer B (which can be replaced with a complete deep neural network), and a fully connected prediction layer B (which can be replaced with a complete deep neural network, such as a tower network B). The fully connected prediction layer B outputs the recommendation metric for recommendation task B. Metric prediction networks A and B share multiple expert networks (e.g., expert network 0, expert network 1, and expert network 2), each corresponding to a different feature dimension. Metric prediction networks A and B also share a feature input layer (which can be replaced with a feature extraction network), each expert network corresponding to a different feature dimension.

[0072] In some embodiments, the attention network (attention network A and attention network B) includes the following operators: matrix multiplication, linear transformation, and maximum likelihood function. The feature representation of the recommendation task (first mapping result) and the output of an expert network (second mapping result) are multiplied by matrix, and the matrix multiplication result is then processed by linear transformation and processed by the maximum likelihood function. The output of the maximum likelihood function operator is then multiplied by matrix with the output of an expert network. After performing the same processing on multiple expert networks, they are concatenated to obtain the attention processing result Attention for the corresponding recommendation task.

[0073] See Figure 6B , Figure 6B This is a schematic diagram of the structure of the multi-recommendation task model provided in the embodiments of this application. If applied to a news recommendation system, assuming there are two recommendation tasks, there can be multiple recommendation tasks, such as recommendation task A and recommendation task B (recommendation tasks are in...). Figure 6AThe recommendation task A is a click rate prediction task, and the recommendation task B is a reading time prediction task, Figure 6B The model structure of the multi-recommendation task model in Figure 6A The model structure of the multi-recommendation task model in

[0074] In some embodiments, the co-attention network (co-attention network A and co-attention network B) includes an attention processing operator and a residual normalization layer (Add&Norm). Taking the co-attention network A as an example, the attention processing result A (Q A ) of the attention network A and the attention processing result B (K B and V B ) of the attention network B are input into the co-attention network A to obtain the corresponding co-attention processing result A. The attention processing result B (Q B ) of the attention network B and the attention processing result A (K A and Q A ) of the attention network A are input into the co-attention network B to obtain the corresponding co-attention processing result B.

[0075] Referring to Figure 4A , Figure 4A is a flowchart of an information recommendation method based on artificial intelligence provided by an embodiment of the present application, which will be described in combination with the steps shown in Figure 4A .

[0076] In step 101, the ranking features of the information to be recommended are obtained.

[0077] In some embodiments, the ranking features of the information to be recommended in step 101 can be implemented by the following technical solution: obtaining discrete features corresponding to a plurality of feature types respectively for the recommendation information; wherein the plurality of feature types include user features, information features, and environment features; for the discrete features of any one feature type, the following processing is performed: weighting processing is performed on a plurality of discrete values in the discrete features by using the latent vectors corresponding to the plurality of discrete values as weights to obtain an embedding dimension value corresponding to one embedding dimension of the discrete features, and the embedding dimension values corresponding to a plurality of embedding dimensions of the discrete features are combined into a dense feature of the feature type; the dense features of the plurality of feature types are spliced to obtain the ranking features corresponding to the information to be recommended.

[0078] As an example, the user features corresponding to the to-be-recommended object can be acquired, including at least one of: a basic attribute feature for representing basic information of the to-be-recommended object; a social relationship feature for representing a social relationship of the user; an interaction behavior feature for representing an interaction behavior of the user; a reading psychological feature for representing a reading preference of the user; the information features of the to-be-recommended information can be acquired, including at least one of: a category feature for representing a category of the to-be-recommended information; a label feature for representing a content of the to-be-recommended information; a time feature for representing a publishing time of the to-be-recommended information; a publishing feature for representing a source of the to-be-recommended information; and a length feature for representing a length of the to-be-recommended information.

[0079] In some embodiments, the basic information can be basic attributes of the user, such as gender, age, long-term residence, etc., the social relationship can be whether to be married or have a job, etc., the user interaction behavior can be behaviors such as liking, forwarding, or collecting, etc., the reading preference can be a reading interest, the interest point can be entertainment gossip or international news, etc., the category of the to-be-recommended information can be a display carrier category of the information, such as video information, image information, or text information, and the content of the to-be-recommended information can be a content theme, such as an education topic or an entertainment topic, etc.

[0080] In some embodiments, the user features can be combined with the information features of each to-be-recommended information respectively to form a fusion feature corresponding to each to-be-recommended information, and the environment features corresponding to the to-be-recommended object can be acquired, including at least one of: a time feature of pushing the to-be-recommended object; a user location feature of the to-be-recommended object; a device feature of the to-be-recommended object; a network feature in which a device used by the to-be-recommended object is located; and the environment features, the user features, and the information features corresponding to the to-be-recommended object are combined into the fusion feature corresponding to each to-be-recommended information.

[0081] In some embodiments, the environment features also have an impact on the prediction of the recommendation indicators, the time of pushing the to-be-recommended object can affect whether the to-be-recommended object has free time to watch the to-be-recommended information, the location of the to-be-recommended object represents a current life scene of the to-be-recommended object, and different life scenes have an impact on the prediction of the indicators, for example, when the location feature represents that the to-be-recommended object is in a cinema, the prediction results of the two recommendation indicators of the time length and the click rate are quite different from the prediction results of the two recommendation indicators of the time length and the click rate when the location feature represents that the to-be-recommended object is in a self-study room, and the network in which the device used by the to-be-recommended object is located can affect whether the to-be-recommended object wants to receive the to-be-recommended information such as a video which needs to consume a large network resource.

[0082] In some embodiments, referring to Figure 7 , Figure 7is a schematic diagram of a feature extraction network of an information recommendation method based on artificial intelligence provided by an embodiment of the present application, the structure of the dense representation layer 601 Figure 7 As shown, although the lengths of the plurality of discrete features input to the sparse feature layer 602 are different (for example, the discrete features of field 1 and the discrete features of field m), the length of the dense feature obtained based on the plurality of discrete features is fixed, assuming that the embedding dimension of the dense feature is 5, for an input information, there can be only one discrete value of 1 in one position (in the same field) in a discrete feature, for the position of the discrete feature, the values of the plurality of embedding dimensions mapped to the dense feature are all 1, then in the process of obtaining the dense feature from the input discrete feature, only one neuron in the sparse feature layer works, and the values corresponding to the five lines connected to the neuron that works are v m1 , v m2 , v m3 , v m4 , v m5 The combination of the five values is the dense feature e m of field m, and for any embedding dimension of the dense feature, the value of the embedding dimension is obtained based on the discrete values of the plurality of positions of the discrete feature, each discrete value contributes to the embedding dimension, and the hidden vector corresponding to the plurality of discrete values of the discrete feature is used as a weight, assuming that the field length of the discrete feature is 3, there are three discrete values, and the corresponding hidden vector is, for example, v 1n , v 2n , v 3n Then the embedding dimension value corresponding to the nth embedding dimension is the weighted processing result of the discrete value of the first position, the discrete value of the second position, and the discrete value of the third position, and the weighted processing is performed for the five embedding dimensions to obtain the dense feature e m The dense feature is spliced to obtain the ranking feature corresponding to the information to be recommended.

[0083] In some embodiments, before obtaining the ranking feature of the information to be recommended in step 101, the following technical solution can also be performed: obtaining the recommendation information by at least one of the following ways: obtaining information similar in content to the historical browsing information of the to-be-recommended object, and the content similarity is not less than a content similarity threshold, as the information to be recommended; obtaining a plurality of reference to-be-recommended objects similar in user portrait to the to-be-recommended object, and the portrait similarity is not less than a portrait similarity threshold, obtaining the historical browsing information corresponding to each reference to-be-recommended object as the information to be recommended.

[0084] As an example, a plurality of to-be-recommended information can be obtained by a recall module in a recommendation system, the plurality of to-be-recommended information is obtained in response to a user request of a to-be-recommended object, the user request here can be a query request carrying a specific target, and can also be a request for initializing an application, the content similarity refers to the similarity between user historical browsing information and to-be-recommended information, and a plurality of reference to-be-recommended objects having a similar user portrait as the to-be-recommended object can also be obtained, and the historical browsing information of the reference to-be-recommended object is taken as to-be-recommended information.

[0085] In step 102, the ranking feature is mapped based on a plurality of recommendation tasks to obtain a first feature mapping result corresponding to each recommendation task.

[0086] In some embodiments, the mapping of the ranking feature based on a plurality of recommendation tasks in step 102 to obtain a first feature mapping result corresponding to each recommendation task can be implemented by the following technical solution: the ranking feature is mapped by a feature full connection layer in an index prediction network corresponding to each recommendation task in a multi-recommendation task model to obtain a first feature mapping result corresponding to each recommendation task.

[0087] As an example, referring to Figure 6A or Figure 6B , the ranking feature is mapped by a feature full connection layer A in an index prediction network A to obtain a first feature mapping result corresponding to a recommendation task A, and the ranking feature is mapped by a feature full connection layer B in an index prediction network B to obtain a first feature mapping result corresponding to a recommendation task B, the number of recommendation tasks is not limited to two, the index prediction network is one-to-one corresponding to the recommendation task, and the feature full connection layer can be replaced by other deep neural networks.

[0088] In step 103, the ranking feature is mapped based on a plurality of feature dimensions to obtain a second feature mapping result corresponding to each feature dimension.

[0089] In some embodiments, the mapping of the ranking feature based on a plurality of feature dimensions in step 103 to obtain a second feature mapping result corresponding to each feature dimension can be implemented by the following technical solution: the ranking feature is mapped to a feature space corresponding to each expert network by a plurality of expert networks in the multi-recommendation task model to obtain a second feature mapping result of a feature dimension associated with each expert network.

[0090] As an example, referring to Figure 6A or Figure 6BThe plurality of expert networks shared by the index prediction network A and the index prediction network B perform mapping processing on the ranking features corresponding to each feature dimension to obtain second feature mapping results corresponding to the feature dimension of each expert network. The number of recommendation tasks is not limited to two, there is no corresponding relationship between the number of expert networks and the number of recommendation tasks, there is no parameter sharing between the plurality of expert networks, each expert network is usually a small-scale fully connected layer with multiple layers. Assuming that there are three expert networks, there are three second feature mapping results corresponding to different feature dimensions.

[0091] In step 104, taking any one recommendation task as a target recommendation task, performing attention processing based on the first feature mapping result and the second feature mapping result corresponding to each feature dimension to obtain a recommendation index of the to-be-recommended information corresponding to the target recommendation task.

[0092] In some embodiments, referring to Figure 4B , Figure 4B is a flowchart of the information recommendation method based on artificial intelligence provided by the embodiments of the present application. In step 104, the first feature mapping result and the second feature mapping result corresponding to each feature dimension are used to perform attention processing to obtain a recommendation index of the to-be-recommended information corresponding to the target recommendation task, which can be realized by steps 1041-1042.

[0093] In step 1041, the first feature mapping result and the second feature mapping result corresponding to each feature dimension are used to perform attention processing to obtain an attention processing result corresponding to the target recommendation task.

[0094] In some embodiments, referring to Figure 4C , Figure 4C is a flowchart of the information recommendation method based on artificial intelligence provided by the embodiments of the present application. In step 1041, the first feature mapping result and the second feature mapping result corresponding to each feature dimension are used to perform attention processing to obtain an attention processing result corresponding to the target recommendation task, which can be realized by steps 10411-10413.

[0095] In step 10411, the first feature mapping result is multiplied by the transpose of the second feature mapping result corresponding to each feature dimension, and the obtained multiplication result corresponding to each feature dimension is linearly processed.

[0096] In step 10412, the linearly processed result corresponding to each feature dimension is maximum likelihood processed, the obtained maximum likelihood processing result is multiplied by the second feature mapping result corresponding to each feature dimension, and an attention processing result corresponding to each feature dimension is obtained.

[0097] In step 10413, the attention processing results corresponding to each feature dimension are spliced to obtain an attention processing result corresponding to the target recommendation task.

[0098] As an example, the above embodiment is taken as the recommendation task A as the target recommendation task, the first mapping result is subjected to attention processing with the second mapping result corresponding to the plurality of expert networks. If there are three expert networks, each expert network corresponds to a feature dimension, each expert network outputs the second mapping result corresponding to the respective feature dimension, see formula (1):

[0099] (1);

[0100] Wherein, softmax is the maximum likelihood function, E matrix is the second mapping result output by a certain expert network, t k is the first mapping result of the kth recommendation task, for example, the first mapping result of the recommendation task A, d is the dimension of the E matrix, is the attention processing result of the corresponding kth recommendation task for a certain expert network (feature dimension).

[0101] As an example, after splicing the plurality of attention processing results of the recommendation task A on the plurality of expert networks , the attention processing result A of the index prediction network A corresponding to the recommendation task A is obtained. The attention distribution of each recommendation task to the expert network plays a role of selective focusing on the information output by the expert network to focus attention on the information that needs to be paid attention to.

[0102] In some embodiments, after the above splicing processing of the attention processing results corresponding to each feature dimension to obtain the attention processing result corresponding to the target recommendation task, the following technical solutions can be further executed: performing linear transformation processing on the attention processing results of the plurality of recommendation tasks to obtain query vectors, key vectors and value vectors corresponding to the plurality of recommendation tasks; determining a co-attention processing result for the target recommendation task based on the query vectors, key vectors and value vectors corresponding to the plurality of recommendation tasks; and updating the attention processing result corresponding to the target recommendation task based on the co-attention processing result for the target recommendation task.

[0103] In some embodiments, the linear transformation processing is performed on the attention processing result of each of the plurality of recommendation tasks to obtain the query vector, the key vector and the value vector corresponding to the attention processing result of each of the plurality of recommendation tasks.

[0104] As an example, the linear transformation processing is actually multiplying the attention processing result by three parameter matrices to obtain the query vector Q, the key vector K and the value vector V corresponding to the attention processing result. In the above embodiment, for example, taking recommendation task A as the target recommendation task, the query vector Q A , the key vector K A and the value vector V A corresponding to the attention processing result A of the recommendation task A are obtained. If the recommendation task A is taken as the target task, the recommendation task B other than the recommendation task A in the plurality of recommendation tasks is taken as the other recommendation task. For example, the linear transformation processing is performed on the attention processing result B of the recommendation task B to obtain the query vector Q B , the key vector K B and the value vector V B corresponding to the attention processing result B of the recommendation task B. If there is a recommendation task C other than the recommendation task B, the recommendation task C also belongs to the other recommendation task relative to the recommendation task A. The linear transformation processing is performed on the attention processing result C of the recommendation task C to obtain the query vector Q C , the key vector K C and the value vector V C corresponding to the attention processing result C of the recommendation task C.

[0105] In some embodiments, the co-attention processing result of the target recommendation task is determined based on the query vector, the key vector and the value vector corresponding to the plurality of recommendation tasks. The following processing is performed for each of the other recommendation tasks: the dot product processing is performed on the query vector of the attention processing result of the target recommendation task and the key vector of the attention processing result of the other recommendation task to obtain the relevance between the target recommendation task and the other recommendation task; and the multiplication processing is performed on the relevance and the value vector of the attention processing result of the other recommendation task to obtain the co-attention processing result of the other recommendation task for the target recommendation task.

[0106] As an example, taking the above-mentioned embodiments as an example, taking recommendation task A as the target recommendation task, the relevance between the attention processing result A of the recommendation task A ( ) and the attention processing result B of the recommendation task B ( ) can be measured in the form of dot product, see formula (2):

[0107] (2);

[0108] Among them, is the relevance between the attention processing result Attention A of the recommendation task A and the attention processing result Attention B of the recommendation task B, is the dot product operator, is the attention processing result A of the recommendation task A, is the attention processing result B of the recommendation task B, is an activation function.

[0109] As an example, the calculation of the prediction full connection layer A of the index prediction network A of the recommendation task A is as follows, see formula (3):

[0110] (3);

[0111] Among them, is the recommendation index of the recommendation task A, is the prediction full connection layer of the index prediction network of the recommendation task A, is essentially to multiply the relevance and the value vector V B of the attention processing result of the other recommendation task (recommendation task B), and the multiplication processing result is the co-attention processing result of the recommendation task A and the recommendation task B.

[0112] As an example, if there is a recommendation task C in addition to the recommendation task B, and the recommendation task C also belongs to other recommendation tasks relative to the recommendation task A, the relevance between the attention processing result Attention A of the recommendation task A and the attention processing result Attention C of the recommendation task C is determined, and the relevance is multiplied with the value vector V C of the attention processing result of the other recommendation task (recommendation task C), and the multiplication processing result is the co-attention processing result of the recommendation task A and the recommendation task C.

[0113] As an example, there is a correlation between the target recommendation task A and other recommendation task B, and the upper layer representation information of the target recommendation task A can be used to guide the attention distribution of the other recommendation task B, and vice versa when the recommendation task B is the target recommendation task, the upper layer representation information of the target recommendation task B can be used to guide the attention distribution of the other recommendation task A.

[0114] In some embodiments, the above updating the attention processing result of the corresponding target recommendation task based on the co-attention processing result of the target recommendation task can be implemented by the following technical solutions: when the number of other recommendation tasks is one, the co-attention processing result of the other recommendation task for the target recommendation task is added to the attention processing result of the target recommendation task, and the attention processing result of the corresponding target recommendation task is updated based on the obtained addition processing result; when the number of other recommendation tasks is multiple, the co-attention processing result of each other recommendation task for the target recommendation task is added to the attention processing result of the target recommendation task, and the attention processing result of the corresponding target recommendation task is updated based on the obtained addition processing result.

[0115] As an example, when the number of other recommendation tasks is one, that is, there is only other recommendation task B in addition to the target recommendation task A, the co-attention processing result of the other recommendation task B for the target recommendation task A is added to the attention processing result A of the target recommendation task A, and the attention processing result of the corresponding target recommendation task is updated based on the obtained addition processing result.

[0116] As an example, when the number of other recommendation tasks is multiple, that is, in addition to the target recommendation task A, there are other recommendation task B and other recommendation task C, the co-attention processing result of each other recommendation task for the target recommendation task is added to the attention processing result of the target recommendation task, that is, the co-attention processing result of the other recommendation task B for the target recommendation task A, the co-attention processing result of the other recommendation task C for the target recommendation task A and the attention processing result A of the target recommendation task A are added, and the attention processing result of the corresponding target recommendation task is updated based on the obtained addition processing result.

[0117] In step 1042, the recommendation index prediction processing of the to-be-recommended information is performed through the prediction full connection layer of the index prediction network corresponding to the target recommendation task in the multi-recommendation task model, and the attention processing result of the corresponding target recommendation task is combined to obtain the recommendation index of the target recommendation task corresponding to the to-be-recommended information.

[0118] As an example, referring to Figure 6A Or Figure 6BThe attention processing result of the target recommendation task can be an output result of an attention network or a co-attention network of the target recommendation task. The prediction full connection layer A in the index prediction network A performs recommendation index prediction on the output result of the attention network A or the co-attention network A, to obtain a recommendation index corresponding to the recommendation task A. The prediction full connection layer B in the index prediction network B performs recommendation index prediction on the output result of the attention network B or the co-attention network B, to obtain a recommendation index corresponding to the recommendation task B. The number of recommendation tasks is not limited to two. The index prediction network corresponds to the recommendation task in a one-to-one manner, and the prediction full connection layer can be replaced by other deep neural networks.

[0119] In step 105, a recommendation operation is performed based on the recommendation index corresponding to each recommendation task of the information to be recommended.

[0120] In some embodiments, the recommendation operation in step 105 is performed based on the recommendation index corresponding to each recommendation task of the information to be recommended, which can be implemented by the following technical solution: obtaining an aggregation rule corresponding to the multiple recommendation tasks; based on the operators included in the aggregation rule, calculating and processing the recommendation indexes corresponding to the multiple recommendation tasks respectively, to obtain comprehensive recommendation indexes of the multiple pieces of information to be recommended corresponding to the multiple recommendation tasks respectively; determining a recommendation order of the multiple pieces of information to be recommended according to the comprehensive recommendation indexes of the multiple pieces of information to be recommended corresponding to the multiple recommendation tasks respectively, and performing a recommendation operation corresponding to the to-be-recommended object based on the recommendation order of the multiple pieces of information to be recommended.

[0121] As an example, step 104 is performed for each recommendation task as a target recommendation task to obtain a recommendation index corresponding to each recommendation task.

[0122] As an example, the aggregation rule of the multiple recommendation tasks can correspond to different multiple recommendation index aggregation processing manners. Here, the multiple recommendation tasks can be at least two of the recommendation tasks for predicting reading time, predicting click rate, predicting collection number, predicting forwarding number, etc. of the information to be recommended. The aggregation rule can be a parameterized addition, multiplication or other operation rule. Here, the multiplication of the prediction reading time and the prediction click rate can be taken as the aggregation processing manner of the multiple recommendation tasks.

[0123] As an example, the recommendation order of the multiple pieces of information to be recommended is determined according to the comprehensive recommendation indexes of the information to be recommended, and a recommendation operation corresponding to the to-be-recommended object is performed based on the recommendation order of the multiple pieces of information to be recommended, that is, the information to be recommended is sorted in descending order according to the comprehensive recommendation indexes of the information to be recommended, at least one piece of information to be recommended with a high ranking is selected from the descending order sorting result, and a recommendation operation corresponding to the to-be-recommended object is performed based on the selected information to be recommended.

[0124] As an example, the top N recommended information in the descending order of the sorting result is selected as the recommended information to be recommended to the to-be-recommended object, and a recommendation operation corresponding to the to-be-recommended object is performed based on the selected recommended information. The recommendation operation can be directly pushed to the to-be-recommended object, or the selected recommended information can be reordered based on the selected recommended information. The reordering refers to reordering the selected recommended information in different ways of aggregation processing of multiple recommendation tasks, so as to obtain recommended information that is more likely to be liked by the to-be-recommended object from a more comprehensive perspective.

[0125] As an example, the recommendation operation corresponding to the to-be-recommended object is performed based on the selected recommended information. The recommended information generated as a recommendation result is based on the prediction of multiple recommendation tasks, rather than the prediction of a single recommendation task, so the recommendation accuracy is high, and the to-be-recommended object is more likely to be interested in the recommended information. The to-be-recommended object will read, click, collect, and forward the recommended information as the recommendation result, thereby improving the proportion of interactive behaviors. The operation of reading, clicking, collecting, and forwarding by the to-be-recommended object will be archived as original information of the user portrait, and used to train the model at time intervals to continuously improve the prediction performance of the model.

[0126] In some embodiments, referring to Figure 4D , Figure 4D is a flowchart of an information recommendation method based on artificial intelligence provided by the embodiments of the present application. Before the step 101 of obtaining the sorting features of the to-be-recommended information, the following steps 106-108 can be performed in each iteration training process of the multi-recommendation task model.

[0127] In step 106, the to-be-recommended information sample is forward propagated in the feature extraction network of the multi-recommendation task model, the multiple expert networks, and the index prediction network corresponding to the multiple recommendation tasks, to perform recommendation task prediction on the to-be-recommended information sample, and obtain the predicted recommendation index of the to-be-recommended information sample corresponding to each recommendation task.

[0128] As an example, the index prediction network includes a feature fully connected layer, an attention network, and a prediction fully connected layer corresponding to the multiple recommendation tasks. The attention network can be replaced by a combination of the attention network and the co-attention network.

[0129] As an example, the training server preprocesses the logs of the recommendation system to construct a training sample set including information samples to be recommended, the training server is a server for training the multi-recommendation task model, and the logs are user behavior data in the recommendation system. The three elements of the user behavior event include: operation, defining an operation action (such as clicking, dragging); parameter / attribute, the parameter can be any attribute related to the event, including (people, time, place, device, business information of operation) triggering the event; attribute value, value of parameter / attribute.

[0130] As an example, first, the discrete features of the information samples to be recommended are transformed into ranking features (dense features) by the feature extraction network of the multi-recommendation task model, the ranking features are mapped by the feature full connection layer A of the index prediction network A of the recommendation task A to obtain a first mapping result, the ranking features are mapped by the feature full connection layer B of the index prediction network B of the recommendation task B to obtain a first mapping result, the ranking features are mapped by multiple expert networks to obtain multiple second mapping results, the first mapping result and the multiple second mapping results corresponding to the recommendation task A are processed by the attention network of the index prediction network A of the recommendation task A (or on the basis of attention processing, the co-attention processing is performed), to obtain the attention processing result A, the first mapping result and the multiple second mapping results corresponding to the recommendation task B are processed by the attention network of the index prediction network B of the recommendation task B (or on the basis of attention processing, the co-attention processing is performed), to obtain the attention processing result B, the prediction full connection layer of the index prediction network A of the recommendation task A performs recommendation index prediction on the attention processing result A to obtain the predicted recommendation index of the recommendation task A, and the prediction full connection layer of the index prediction network B of the recommendation task B performs recommendation index prediction on the attention processing result B to obtain the predicted recommendation index of the recommendation task B.

[0131] In step 107, the error between the predicted recommendation index of the information sample to be recommended corresponding to each recommendation task and the pre-labeled recommendation index corresponding to each recommendation task is determined.

[0132] In step 108, the errors corresponding to each recommendation task are aggregated according to the loss function to obtain an aggregated error, and the aggregated error is back propagated in the multi-recommendation task model to determine the parameter change value of the multi-recommendation task model when the loss function reaches the minimum value, and the parameters of the multi-recommendation task model are updated based on the parameter change value.

[0133] In some embodiments, the multi-recommendation task model is initialized, and a plurality of sub-loss functions corresponding to the plurality of recommendation tasks are initialized, the sub-loss functions including discrete features of the information samples to be recommended and corresponding pre-labeled recommendation indicators, the predicted recommendation indicators corresponding to the discrete features of the information samples to be recommended are substituted into the sub-loss functions corresponding to the recommendation tasks, and the errors (obtained based on the sub-loss functions) corresponding to the respective recommendation tasks are aggregated according to the loss function to obtain an aggregated error, specifically, the respective sub-loss functions are weighted and summed by combining the loss weights corresponding to the respective sub-loss functions to obtain a loss function corresponding to the multi-recommendation task model; the loss function is minimized to obtain parameters of the multi-recommendation task model when the loss function reaches a minimum value; wherein the parameters of the multi-recommendation task model include structure parameters of the multi-recommendation task model and loss weights of the loss function.

[0134] In some embodiments, the training server converts the parameters of the multi-recommendation task model into constants and fixes them in the multi-recommendation task model to generate a fixed binary model file, and the training server pushes the binary model file to the recommendation system so that the multi-recommendation task model used in the recommendation system is consistent with the structure of the multi-recommendation task model obtained through training.

[0135] In some embodiments, the network structure and the model training parameters are fused, the model parameters are converted into constants, and the parameters are fixed in the model network structure to ensure the consistency of the network structure of the offline training model and the online prediction model, the online recommendation system loads the fixed model file to obtain the network structure and the model training parameters, thereby ensuring consistency, the final binary model file is pushed to the online recommendation system according to a preset frequency, day by day or hour by hour, and the file content is encoded by a message digest algorithm for subsequent model verification, thereby completing the entire offline training process.

[0136] In the following, an exemplary application of the information recommendation method based on artificial intelligence provided by the embodiments of the present application in an actual application scenario will be described.

[0137] In the related art, the multi-recommendation task model based on a deep neural network is widely applied in business applications such as news recommendation and video recommendation. For example, in a news recommendation system, when recommending information to a user, not only information that the user is interested in is recommended, but also reading completion conversion is promoted as much as possible, and therefore, both click rate and reading time are modeled. For a multi-recommendation task model, focusing on a feature subset related to the recommendation task itself and sharing useful information learned by other recommendation tasks is a key to determining the prediction accuracy of the multi-recommendation task model.

[0138] In order to improve the focusing ability of different recommendation tasks on feature subsets and more finely control the sharing granularity between multiple recommendation tasks, the application provides an information recommendation method based on artificial intelligence, which is implemented based on a multi-recommendation task model. The multi-recommendation task model is a model for learning the correlation relationship of multiple recommendation tasks based on an attention mechanism. In the multi-recommendation task model, first, the expression vectors (first mapping results corresponding to each recommendation task) related to each recommendation task are learned from the input feature information (ranking features), and then the expression vectors (first mapping results) of different recommendation tasks are respectively subjected to attention processing with the expression vectors (second mapping results) of each expert network to obtain the attention distribution of different recommendation tasks on the expert network. The focusing result (attention processing result corresponding to the recommendation task) of the attention processing is taken as the input of the tower network (prediction full connection layer corresponding to the recommendation task) of the respective recommendation task to replace the gating mode of the gating network in the multi-recommendation task model, so as to realize the focusing of the feature subset of the respective recommendation task, that is, the focusing of the feature subset of the feature dimension corresponding to the recommendation task. Finally, the collaborative attention processing is performed for different recommendation tasks to realize the sharing of the upper expression of the multi-recommendation task model, so as to capture the correlation relationship between the recommendation tasks.

[0139] Compared with the gating network in the multi-recommendation task model, the attention processing mode is more direct, more fine, and more interpretable for focusing on the feature subset. Experimental results show that the multi-recommendation task model provided by the application can more accurately distinguish the features in the feature subset that are related to the respective recommendation task and the features that are not related to the respective recommendation task. The attention of the multi-recommendation task model is concentrated on the features that actually have an impact on the recommendation task. The focusing on the feature subset is more accurate, and the sharing granularity between multiple recommendation tasks is better controlled.

[0140] The information recommendation method based on artificial intelligence provided by the application can be applied to a news recommendation system. Referring to Figure 5 , Figure 5 is a framework schematic diagram of the information recommendation method based on artificial intelligence provided by the application. The news recommendation system usually includes four modules: a user portrait module, a recall module, a click rate estimation module, and a rearrangement module. The recall module performs rough selection on the information in the content pool. The recall module outputs the rough selected information to the click rate estimation module. The click rate estimation module performs fine sorting on the recalled information. The rearrangement module performs reordering (diversity and style mixing) on the sorted results. Finally, the information is recommended to the user. The recall module, the click rate estimation module, and the rearrangement module are all run based on the user portrait data output by the user portrait module. In the entire recommendation process, the click rate estimation module plays a decisive role. The click rate estimation module depends on the sorting model, and the advantages and disadvantages of the sorting model determine the accuracy of the recommendation result.

[0141] The applicant finds that when the multi-recommendation task model is used as a ranking model, the feature subset related to the recommendation task itself can be focused on, and useful information learned by other recommendation tasks can be shared. The attention mechanism can determine the feature subset that needs to be focused on, and allocate limited processing resources to important information. For the learning process of multiple recommendation tasks, the focused learning of the feature subset is an effective technical means.

[0142] Referring to Figure 6A , Figure 6A is a structural diagram of the multi-recommendation task model provided by the embodiments of the present application. First, the input features (ranking features) are learned by the feature full connection layer of the index prediction network corresponding to each recommendation task to obtain the vector expression of the ranking features for different recommendation tasks (the first mapping result corresponding to the kth recommendation task), see formula (4):

[0143] (4);

[0144] wherein x is the input feature, is the feature full connection layer of the index prediction network corresponding to the kth recommendation task, is the output of the feature full connection layer of the index prediction network of the kth recommendation task (i.e., the vector expression of different recommendation tasks).

[0145] In some embodiments, the vector expressions of different recommendation tasks are respectively subjected to attention processing with the vector expression of the expert network to obtain the attention distribution of different recommendation tasks on the expert network. The focused result of the attention processing is used as the input of the prediction full connection layer corresponding to each recommendation task to replace the gating mode in the multi-recommendation task model, see formula (5):

[0146] (5);

[0147] wherein softmax is the maximum likelihood function, E is the vector expression of the expert network, t k is the vector expression of the kth recommendation task, d is the dimension of E, is the output of the attention network of the index prediction network corresponding to the kth recommendation task.

[0148] Finally, the attention focused results of the kth recommendation task on multiple expert networks are spliced and input into the prediction full connection layer (the corresponding tower network) of the index prediction network corresponding to the kth recommendation task for learning to fit the recommendation index of each recommendation task, see formula (6):

[0149] (6);

[0150] wherein, is the output of the prediction fully connected layer of the indicator prediction network corresponding to the kth recommendation task, is the prediction fully connected layer (corresponding tower network) of the indicator prediction network corresponding to the kth recommendation task. is a concatenation operation, is the output of the attention network of the indicator prediction network corresponding to the kth recommendation task.

[0151] Each recommendation task plays a role of selectively focusing on the output information of the expert network through the attention distribution of the expert network, so as to focus attention on important information that needs to be focused on, in order to better control the shared granularity between different recommendation tasks, collaborative attention processing can be implemented between the indicator prediction networks of different recommendation tasks, so as to realize the sharing of the upper layer expression of the model.

[0152] Referring to Figure 6B , Figure 6B is a structural diagram of a multi-recommendation task model provided by an embodiment of the present application, and the focusing result of the attention processing of the recommendation task on the expert network is input to the prediction fully connected layer of the indicator prediction network corresponding to the recommendation task, and collaborative attention processing calculation is performed before the input, so as to realize the purpose of sharing useful information between the upper layer expressions, in the process of collaborative attention processing calculation, the dot product is used to measure the correlation between the attention processing result Attention A of the corresponding recommendation task A and the attention processing result Attention B of the corresponding recommendation task B, see formula (7):

[0153] (7);

[0154] wherein, is the correlation between the attention processing result Attention A of the corresponding recommendation task A and the attention processing result Attention B of the corresponding recommendation task B, is a dot product operator, is the attention processing result of the corresponding recommendation task A, is the attention processing result of the corresponding recommendation task B, is an activation function.

[0155] In some embodiments, the calculation of the prediction fully connected layer of the indicator prediction network corresponding to the recommendation task is as follows, based on formulas (8) and (9), the recommendation indicators of the recommendation task A and the recommendation task B can be obtained:

[0156] (8);

[0157] (9);

[0158] in, This corresponds to the recommendation metric for recommendation task A. This corresponds to the recommendation metric for recommendation task B. This is the prediction fully connected layer of the metric prediction network corresponding to recommendation task A. It is the prediction fully connected layer of the indicator prediction network corresponding to recommendation task B.

[0159] In some embodiments, there is a correlation between recommendation task A and recommendation task B. The upper-level representation information of recommendation task A can be used to guide the attention distribution of recommendation task B, and conversely, the upper-level representation information of recommendation task B can be used to guide the attention distribution of recommendation task A, thereby achieving mutual reinforcement between information utilization.

[0160] Compared to gating networks in multi-recommendation task models, attention processing offers a more direct, refined, and interpretable focus on feature subsets. AI-based information recommendation methods are implemented using multi-recommendation task models, which are attention-based models that learn the relationships between multiple recommendation tasks. Compared to gating networks in multi-recommendation task models, attention processing can more accurately distinguish between features in the feature subset that are relevant to the corresponding recommendation task and those that are not. It focuses attention on features that genuinely influence the corresponding recommendation task, resulting in more precise focus on feature subsets and better control over the granularity of sharing between multiple recommendation tasks. Both offline evaluation metrics (AUC) and online metrics are improved.

[0161] It is understood that in the embodiments of this application, data related to user profiles and profile similarity are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with relevant laws, regulations and standards.

[0162] The following description continues to illustrate the exemplary structure of the artificial intelligence-based information recommendation device 255 provided in the embodiments of this application as a software module. In some embodiments, such as... Figure 3As shown, the software modules stored in the artificial intelligence-based information recommendation apparatus 255 in the memory 250 can include: an acquisition module 2551 configured to acquire ranking features of to-be-recommended information; a mapping module 2552 configured to perform mapping processing on the ranking features based on multiple recommendation tasks to obtain first feature mapping results corresponding to each recommendation task; the mapping module 2552 is further configured to perform mapping processing on the ranking features based on multiple feature dimensions to obtain second feature mapping results corresponding to each feature dimension; an attention module 2553 configured to take any one recommendation task as a target recommendation task, and perform attention processing based on the first feature mapping results and the second feature mapping results corresponding to each feature dimension to obtain a recommendation index of the to-be-recommended information corresponding to the target recommendation task; and a recommendation module 2554 configured to perform a recommendation operation based on the recommendation index of the to-be-recommended information corresponding to each recommendation task.

[0163] In some embodiments, the acquisition module 2551 is further configured to: acquire discrete features corresponding to multiple feature types of the recommendation information, wherein the multiple feature types include: user features, information features, and environment features; and perform the following processing on the discrete features of any one of the feature types: perform weighting processing on multiple discrete values in the discrete features by taking hidden vectors corresponding to the multiple discrete values as weights to obtain embedding dimension values of the discrete features corresponding to one embedding dimension, and combine the embedding dimension values of the discrete features corresponding to multiple embedding dimensions into dense features of the feature type; and perform splicing processing on the dense features of the multiple feature types to obtain the ranking features corresponding to the to-be-recommended information.

[0164] In some embodiments, the acquisition module 2551 is further configured to: before acquiring the ranking features of the to-be-recommended information, acquire the recommendation information by at least one of the following ways: acquiring information similar in content to historical browsing information of a to-be-recommended object and having a content similarity not less than a content similarity threshold to serve as the to-be-recommended information; and acquiring multiple reference to-be-recommended objects similar in user portrait to the to-be-recommended object and having a portrait similarity not less than a portrait similarity threshold, and acquiring historical browsing information corresponding to each reference to-be-recommended object to serve as the to-be-recommended information.

[0165] In some embodiments, the mapping module 2552 is further configured to: perform mapping processing on the ranking features by a feature full connection layer in an index prediction network corresponding to each recommendation task to obtain the first feature mapping results corresponding to each recommendation task; and map the ranking features to feature spaces corresponding to each expert network by multiple expert networks to obtain second feature mapping results of feature dimensions associated with each expert network.

[0166] In some embodiments, the attention module 2553 is further configured to: perform attention processing on the first feature mapping result and the second feature mapping result corresponding to each feature dimension to obtain an attention processing result corresponding to the target recommendation task; and perform recommendation task prediction processing on the information to be recommended by combining the attention processing result corresponding to the target recommendation task and a prediction full connection layer of an index prediction network corresponding to the target recommendation task in the multi-recommendation task model to obtain a recommendation index of the information to be recommended corresponding to the target recommendation task.

[0167] In some embodiments, the attention module 2553 is further configured to: multiply the first feature mapping result and the transpose of the second feature mapping result corresponding to each feature dimension, perform linear processing on the obtained multiplication result corresponding to each feature dimension, perform maximum likelihood processing on the obtained linear processing result corresponding to each feature dimension, multiply the obtained maximum likelihood processing result and the second feature mapping result corresponding to each feature dimension, and obtain an attention processing result corresponding to each feature dimension; and perform splicing processing on the attention processing result corresponding to each feature dimension to obtain an attention processing result corresponding to the target recommendation task.

[0168] In some embodiments, the attention module 2553 is further configured to: perform splicing processing on the attention processing result corresponding to each feature dimension to obtain an attention processing result corresponding to the target recommendation task, perform linear transformation processing on the attention processing results of the plurality of recommendation tasks to obtain query vectors, key vectors and value vectors corresponding to the plurality of recommendation tasks, determine a co-attention processing result for the target recommendation task based on the query vectors, the key vectors and the value vectors corresponding to the plurality of recommendation tasks, and update the attention processing result corresponding to the target recommendation task based on the co-attention processing result for the target recommendation task.

[0169] In some embodiments, the attention module 2553 is further configured to: perform linear transformation processing on the attention processing result of the target recommendation task to obtain a query vector, a key vector and a value vector of the attention processing result corresponding to the target recommendation task, and perform linear transformation processing on the attention processing result of at least one other recommendation task to obtain a query vector, a key vector and a value vector of the attention processing result corresponding to each other recommendation task, wherein the other recommendation task is a recommendation task other than the target recommendation task in the plurality of recommendation tasks.

[0170] In some embodiments, the attention module 2553 is further configured to: for each other recommendation task, perform the following processing: perform dot product processing on the query vector of the attention processing result of the target recommendation task and the key vector of the attention processing result of the other recommendation task to obtain the relevance between the target recommendation task and the other recommendation task; and perform multiplication processing on the relevance and the value vector of the attention processing result of the other recommendation task to obtain the co-attention processing result of the other recommendation task for the target recommendation task.

[0171] In some embodiments, the attention module 2553 is further configured to: when the number of other recommendation tasks is one, perform addition processing on the co-attention processing result of the other recommendation task for the target recommendation task and the attention processing result of the target recommendation task, and update the attention processing result of the corresponding target recommendation task based on the obtained addition processing result; and when the number of other recommendation tasks is multiple, perform addition processing on the co-attention processing result of each other recommendation task for the target recommendation task and the attention processing result of the target recommendation task, and update the attention processing result of the corresponding target recommendation task based on the obtained addition processing result.

[0172] In some embodiments, the recommendation module 2554 is further configured to: obtain an aggregation rule corresponding to the multiple recommendation tasks; perform calculation processing on the recommendation indexes respectively corresponding to the multiple recommendation tasks based on the operators included in the aggregation rule to obtain comprehensive recommendation indexes of the multiple to-be-recommended information respectively corresponding to the multiple recommendation tasks; determine the recommendation order of the multiple to-be-recommended information according to the comprehensive recommendation indexes of the multiple to-be-recommended information respectively corresponding to the multiple recommendation tasks, and perform the recommendation operation on the corresponding to-be-recommended object based on the recommendation order of the multiple to-be-recommended information.

[0173] In some embodiments, the information recommendation apparatus 255 based on artificial intelligence further comprises: a training module 2555 configured to: before obtaining the ranking features of the to-be-recommended information, perform the following processing in each iteration training process of the multi-recommendation task model: perform forward propagation of the to-be-recommended information sample in the feature extraction network of the multi-recommendation task model, the multiple expert networks, and the index prediction network corresponding to the multiple recommendation tasks one by one to perform recommendation task prediction on the to-be-recommended information sample to obtain the predicted recommendation indexes of the to-be-recommended information sample corresponding to each recommendation task; wherein the index prediction network comprises a feature full connection layer corresponding to the multiple recommendation tasks one by one, an attention network, and a prediction full connection layer; determine the error between the predicted recommendation indexes of the to-be-recommended information sample corresponding to each recommendation task and the pre-labeled recommendation indexes corresponding to each recommendation task; aggregate the errors corresponding to each recommendation task according to a loss function to obtain an aggregated error, and perform backward propagation of the aggregated error in the multi-recommendation task model to determine the parameter change value of the multi-recommendation task model when the loss function reaches a minimum value, and update the parameters of the multi-recommendation task model based on the parameter change value.

[0174] The embodiment of the present application provides a computer program product or computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the information recommendation method based on artificial intelligence provided by the embodiment of the present application.

[0175] The embodiment of the present application provides a computer readable storage medium storing executable instructions, wherein the executable instructions are stored, and when the executable instructions are executed by a processor, the processor executes the information recommendation method based on artificial intelligence provided by the embodiment of the present application, for example, as shown in the information recommendation method based on artificial intelligence. Figures 4A-4D The embodiment of the present application provides a computer readable storage medium storing executable instructions, wherein the executable instructions are stored, and when the executable instructions are executed by a processor, the processor executes the information recommendation method based on artificial intelligence provided by the embodiment of the present application, for example, as shown in the information recommendation method based on artificial intelligence.

[0176] In some embodiments, the computer readable storage medium can be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; or various devices including one or any combination of the above memories.

[0177] In some embodiments, the executable instructions can be in the form of programs, software, software modules, scripts or codes, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and can be deployed in any form, including being deployed as independent programs or being deployed as modules, components, subroutines or other units suitable for use in a computing environment.

[0178] As an example, the executable instructions can but not necessarily correspond to files in a file system, can be stored in a part of a file storing other programs or data, for example, stored in one or more scripts in a hyper text markup language (HTML, Hyper Text Markup Language) document, stored in a single file dedicated to the program in question, or stored in multiple cooperative files (for example, files storing one or more modules, subroutines or code parts).

[0179] As an example, the executable instructions can be deployed to execute on one computing device, or on multiple computing devices located in one place, or on multiple computing devices distributed in multiple places and interconnected through a communication network.

[0180] In summary, by the embodiment of the application, the mapping processing based on the feature dimension and based on the task of the sorting feature is performed, and the attention processing of the respective mapping results is performed, so that the recommendation task is directly associated with different feature dimensions, the indicators corresponding to multiple recommendation tasks are more accurately predicted, and accurate personalized recommendation is realized.

[0181] The above merely describes the embodiments of the present application, but does not serve to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and scope of the present application shall be included in the protection scope of the present application.

Claims

1. An information recommendation method based on artificial intelligence, characterized in that, The method comprises: obtaining a sorting feature of to-be-recommended information; wherein the sorting feature is obtained by fusing discrete features of multiple feature types of the to-be-recommended information; performing mapping processing on the sorting feature based on multiple recommendation tasks to obtain a task-specific first feature mapping result corresponding to each of the recommendation tasks; inputting the sorting feature into multiple expert networks shared by the multiple recommendation tasks to perform mapping processing on the sorting feature based on multiple feature dimensions to obtain a second feature mapping result corresponding to each of the feature dimensions; taking any one of the recommendation tasks as a target recommendation task, using the first feature mapping result corresponding to the target recommendation task as a query, performing attention processing based on the first feature mapping result and the second feature mapping result corresponding to each of the feature dimensions to obtain a recommendation index of the to-be-recommended information corresponding to the target recommendation task; and performing a recommendation operation based on the recommendation index of the to-be-recommended information corresponding to each of the recommendation tasks.

2. The method of claim 1, wherein, The method comprises: obtaining discrete features of multiple feature types of the to-be-recommended information; wherein the multiple feature types include user features, information features, and environment features; for the discrete features of any one of the feature types, performing the following processing: weighting multiple discrete values in the discrete features using hidden vectors corresponding to the multiple discrete values as weights to obtain embedding dimension values of one embedding dimension corresponding to the discrete features, and combining the embedding dimension values of multiple embedding dimensions corresponding to the discrete features into a dense feature of the feature type; and performing splicing processing on the dense features of the multiple feature types to obtain the sorting feature corresponding to the to-be-recommended information.

3. The method of claim 1, wherein, Before the method of obtaining the sorting feature of the to-be-recommended information, the method further comprises: obtaining the to-be-recommended information by at least one of the following methods: obtaining information with a content similarity to historical browsing information of a to-be-recommended object not less than a content similarity threshold as the to-be-recommended information; obtaining multiple reference to-be-recommended objects with a user portrait similarity to a user portrait of the to-be-recommended object not less than a portrait similarity threshold, and obtaining historical browsing information corresponding to each of the reference to-be-recommended objects as the to-be-recommended information.

4. The method of claim 1, wherein the mapping processing on the sorting feature based on multiple recommendation tasks to obtain a task-specific first feature mapping result corresponding to each of the recommendation tasks comprises: performing mapping processing on the sorting feature through a feature full connection layer in an index prediction network corresponding to each of the recommendation tasks to obtain a first feature mapping result corresponding to each of the recommendation tasks; and the inputting of the sorting feature into multiple expert networks shared by the multiple recommendation tasks to perform mapping processing on the sorting feature based on multiple feature dimensions to obtain a second feature mapping result corresponding to each of the feature dimensions comprises: ​ The ranking features are respectively mapped to feature spaces corresponding to each of the expert networks through the multiple expert networks, to obtain second feature mapping results of feature dimensions associated with each of the expert networks.

5. The method of claim 1, wherein, The attention processing based on the first feature mapping results and the second feature mapping results corresponding to each of the feature dimensions obtains a recommendation index of the information to be recommended corresponding to the target recommendation task, including: The attention processing based on the first feature mapping results and the second feature mapping results corresponding to each of the feature dimensions obtains an attention processing result corresponding to the target recommendation task. The recommendation index of the information to be recommended corresponding to the target recommendation task is obtained through a prediction fully connected layer in an index prediction network corresponding to the target recommendation task and in combination with the attention processing result corresponding to the target recommendation task.

6. The method of claim 5, wherein, The attention processing based on the first feature mapping results and the second feature mapping results corresponding to each of the feature dimensions obtains an attention processing result corresponding to the target recommendation task, including: The first feature mapping results are multiplied by transpositions of the second feature mapping results corresponding to each of the feature dimensions, and linear processing is performed on the obtained multiplication results corresponding to each of the feature dimensions. The linear processing results corresponding to each of the feature dimensions are subjected to maximum likelihood processing, and the obtained maximum likelihood processing results are multiplied by the second feature mapping results corresponding to each of the feature dimensions to obtain attention processing results corresponding to each of the feature dimensions. The attention processing results corresponding to each of the feature dimensions are spliced to obtain an attention processing result corresponding to the target recommendation task.

7. The method of claim 6, wherein, After the attention processing results corresponding to each of the feature dimensions are spliced to obtain an attention processing result corresponding to the target recommendation task, the method further includes: The attention processing results of the multiple recommendation tasks are subjected to linear transformation processing to obtain query vectors, key vectors, and value vectors corresponding to the multiple recommendation tasks. Based on the query vectors, key vectors, and value vectors corresponding to the multiple recommendation tasks, a co-attention processing result corresponding to the target recommendation task is determined. Based on the co-attention processing result corresponding to the target recommendation task, the attention processing result corresponding to the target recommendation task is updated.

8. The method of claim 7, wherein, The attention processing results of the multiple recommendation tasks are subjected to linear transformation processing to obtain query vectors, key vectors, and value vectors corresponding to the multiple recommendation tasks, including: The attention processing result of the target recommendation task is subjected to linear transformation processing to obtain query vectors, key vectors, and value vectors of the attention processing result corresponding to the target recommendation task. The attention processing results of at least one other recommendation task are subjected to linear transformation processing to obtain query vectors, key vectors, and value vectors of the attention processing result corresponding to each of the other recommendation tasks. The other recommendation tasks are recommendation tasks in the multiple recommendation tasks that are different from the target recommendation task.

9. The method of claim 8, wherein, The query vector, the key vector, and the value vector corresponding to the plurality of recommendation tasks are used to determine a co-attention processing result corresponding to the target recommendation task, including: The following processing is performed for each of the other recommendation tasks: The query vector of the attention processing result of the target recommendation task is dot multiplied with the key vector of the attention processing result of the other recommendation task to obtain a relevance between the target recommendation task and the other recommendation task; The relevance is multiplied with the value vector of the attention processing result of the other recommendation task to obtain a co-attention processing result of the other recommendation task for the target recommendation task.

10. The method of claim 8, wherein, The co-attention processing result of the other recommendation task for the target recommendation task is added to the attention processing result of the target recommendation task, and the attention processing result of the target recommendation task is updated based on the obtained addition processing result. When the number of other recommendation tasks is one, the co-attention processing result of the other recommendation task for the target recommendation task is added to the attention processing result of the target recommendation task, and the attention processing result of the target recommendation task is updated based on the obtained addition processing result. When the number of other recommendation tasks is one, the co-attention processing result of the other recommendation task for the target recommendation task is added to the attention processing result of the target recommendation task, and the attention processing result of the target recommendation task is updated based on the obtained addition processing result.

11. The method of claim 1, wherein, The recommendation indicators corresponding to each of the recommendation tasks are calculated based on the aggregation rules included in the aggregation rules, and the comprehensive recommendation indicators of the plurality of to-be-recommended information corresponding to the plurality of recommendation tasks are obtained. According to the comprehensive recommendation indicators of the plurality of to-be-recommended information corresponding to the plurality of recommendation tasks, the recommendation order of the plurality of to-be-recommended information is determined, and the recommendation operation corresponding to the to-be-recommended object is performed based on the recommendation order of the plurality of to-be-recommended information.

12. The method of any one of claims 1-11, wherein, Before the ordering feature of the to-be-recommended information is obtained, the method further comprises: The following processing is performed in each iteration training process of the multi-recommendation task model: The to-be-recommended information sample is forward propagated in the feature extraction network, the plurality of expert networks, and the index prediction network corresponding to the plurality of recommendation tasks of the multi-recommendation task model to perform recommendation index prediction on the to-be-recommended information sample, and obtain the predicted recommendation index of the to-be-recommended information sample corresponding to each of the recommendation tasks; The index prediction network includes a feature full connection layer, an attention network, and a prediction full connection layer corresponding to each of the recommendation tasks; An error between the predicted recommendation index of the to-be-recommended information sample corresponding to each of the recommendation tasks and the pre-labeled recommendation index corresponding to each of the recommendation tasks is determined. ​ ​ aggregate errors corresponding to the respective recommendation tasks according to a loss function to obtain an aggregated error, and back-propagate the aggregated error in the multi-recommendation task model to determine a parameter change value of the multi-recommendation task model when the loss function reaches a minimum value, and update parameters of the multi-recommendation task model based on the parameter change value.

13. An information recommendation apparatus based on artificial intelligence, characterized by comprising: comprise: an acquisition module, configured to acquire a ranking feature of to-be-recommended information, wherein the ranking feature is obtained by fusing discrete features of multiple feature types of the to-be-recommended information; a mapping module, configured to perform mapping processing on the ranking feature based on multiple recommendation tasks to obtain a first feature mapping result specific to each of the recommendation tasks; the mapping module is further configured to input the ranking feature into multiple expert networks shared by the multiple recommendation tasks to perform mapping processing on the ranking feature based on multiple feature dimensions to obtain a second feature mapping result corresponding to each of the feature dimensions; an attention module, configured to take any one of the recommendation tasks as a target recommendation task, use the first feature mapping result corresponding to the target recommendation task as a query, and perform attention processing based on the first feature mapping result and the second feature mapping result corresponding to each of the feature dimensions to obtain a recommendation indicator of the to-be-recommended information corresponding to the target recommendation task; a recommendation module, configured to perform a recommendation operation based on the recommendation indicator of the to-be-recommended information corresponding to each of the recommendation tasks.

14. The apparatus of claim 13, wherein, the acquisition module is further configured to: acquire discrete features of multiple feature types of the recommendation information; wherein the multiple feature types comprise user features, information features, and environmental features; for the discrete features of any one of the feature types, perform the following processing: weight multiple discrete values in the discrete features using hidden vectors corresponding to the multiple discrete values as weights to obtain embedding dimension values of one embedding dimension corresponding to the discrete features, and combine the embedding dimension values of multiple embedding dimensions corresponding to the discrete features into a dense feature of the feature type; perform splicing processing on the dense features of the multiple feature types to obtain the ranking feature corresponding to the to-be-recommended information.

15. The apparatus of claim 13, wherein, the acquisition module is further configured to: acquire the recommendation information by at least one of the following methods: acquire information with a content similarity to historical browsing information of a to-be-recommended object that is not less than a content similarity threshold, as the to-be-recommended information; acquire multiple reference to-be-recommended objects with a user portrait similarity to a user portrait of the to-be-recommended object that is not less than a portrait similarity threshold, and acquire historical browsing information corresponding to each of the reference to-be-recommended objects as the to-be-recommended information.

16. The apparatus of claim 13, wherein, the mapping module is further configured to: perform mapping processing on the ranking feature through a feature fully connected layer in an indicator prediction network corresponding to each of the recommendation tasks to obtain the first feature mapping result corresponding to each of the recommendation tasks; map the ranking feature to a feature space corresponding to each of the expert networks respectively to obtain the second feature mapping result of the feature dimension associated with each of the expert networks.

17. The apparatus of claim 13, wherein, the attention module is further configured to: perform attention processing on the first feature mapping result and the second feature mapping result corresponding to each feature dimension to obtain an attention processing result corresponding to the target recommendation task; perform recommendation index prediction processing on the to-be-recommended information through a prediction fully connected layer in an index prediction network corresponding to the target recommendation task and in combination with the attention processing result corresponding to the target recommendation task to obtain a recommendation index of the to-be-recommended information corresponding to the target recommendation task.

18. The apparatus of claim 13, wherein, The attention module is further configured to: multiply the first feature mapping result and a transpose of the second feature mapping result corresponding to each feature dimension, perform linear processing on a multiplication result corresponding to each feature dimension, and perform maximum likelihood processing on a linear processing result corresponding to each feature dimension, multiply a maximum likelihood processing result and the second feature mapping result corresponding to each feature dimension, and obtain an attention processing result corresponding to each feature dimension. perform splicing processing on the attention processing result corresponding to each feature dimension to obtain the attention processing result corresponding to the target recommendation task.

19. The apparatus of claim 13, wherein, The attention module is further configured to: perform linear transformation processing on the attention processing results of the plurality of recommendation tasks to obtain query vectors, key vectors, and value vectors corresponding to the plurality of recommendation tasks, determine a co-attention processing result corresponding to the target recommendation task based on the query vectors, the key vectors, and the value vectors corresponding to the plurality of recommendation tasks, and update the attention processing result corresponding to the target recommendation task based on the co-attention processing result corresponding to the target recommendation task.

20. The apparatus of claim 13, wherein, The attention module is further configured to: perform linear transformation processing on the attention processing result of the target recommendation task to obtain query vectors, key vectors, and value vectors of the attention processing result corresponding to the target recommendation task, perform linear transformation processing on the attention processing result of at least one other recommendation task to obtain query vectors, key vectors, and value vectors of the attention processing result corresponding to each other recommendation task, wherein the other recommendation task is a recommendation task that is different from the target recommendation task in the plurality of recommendation tasks.

21. The apparatus of claim 13, wherein, The attention module is further configured to: perform the following processing for each other recommendation task: perform dot product processing on the query vectors of the attention processing result of the target recommendation task and the key vectors of the attention processing result of the other recommendation task to obtain a relevance between the target recommendation task and the other recommendation task, multiply the relevance and the value vectors of the attention processing result of the other recommendation task to obtain a co-attention processing result of the other recommendation task for the target recommendation task.

22. The apparatus of claim 13, wherein, The attention module is further configured to: when the number of the other recommendation tasks is one, perform addition processing on the co-attention processing result of the other recommendation task corresponding to the target recommendation task and the attention processing result of the target recommendation task, and update the attention processing result corresponding to the target recommendation task based on a result of the addition processing. When the number of other recommendation tasks is multiple, the attention processing result of the target recommendation task corresponding to each of the other recommendation tasks is added to the attention processing result of the target recommendation task, and the attention processing result of the target recommendation task is updated based on the obtained addition processing result.

23. The apparatus of claim 13, wherein, The recommendation module is further configured to: obtain an aggregation rule corresponding to the multiple recommendation tasks; perform calculation processing on the recommendation indexes corresponding to the multiple recommendation tasks respectively based on the operators included in the aggregation rule, to obtain comprehensive recommendation indexes of the multiple pieces of to-be-recommended information corresponding to the multiple recommendation tasks respectively; determine a recommendation order of the multiple pieces of to-be-recommended information according to the comprehensive recommendation indexes of the multiple pieces of to-be-recommended information corresponding to the multiple recommendation tasks respectively, and perform a recommendation operation corresponding to a to-be-recommended object based on the recommendation order of the multiple pieces of to-be-recommended information.

24. The apparatus of claim 13, wherein, The device further includes a training module configured to: perform the following processing in each iteration training process of the multi-recommendation task model: perform forward propagation of a to-be-recommended information sample in a feature extraction network, multiple expert networks, and index prediction networks corresponding to the multiple recommendation tasks one by one, to perform recommendation index prediction on the to-be-recommended information sample, to obtain predicted recommendation indexes of the to-be-recommended information sample corresponding to each of the recommendation tasks; wherein the index prediction network includes a feature full connection layer, an attention network, and a prediction full connection layer corresponding to the multiple recommendation tasks one by one; determine errors between the predicted recommendation indexes of the to-be-recommended information sample corresponding to each of the recommendation tasks and pre-labeled recommendation indexes corresponding to each of the recommendation tasks; aggregate the errors corresponding to each of the recommendation tasks according to a loss function to obtain an aggregated error, and perform backward propagation of the aggregated error in the multi-recommendation task model, to determine a parameter change value of the multi-recommendation task model when the loss function reaches a minimum value, and update parameters of the multi-recommendation task model based on the parameter change value.

25. An electronic device, comprising: comprise: a memory configured to store executable instructions; a processor configured to execute the executable instructions stored in the memory, to implement the artificial intelligence-based information recommendation method in any one of claims 1 to 12.

26. A computer-readable storage medium, characterized in that, executable instructions stored in the memory, to implement the artificial intelligence-based information recommendation method in any one of claims 1 to 12.

27. A computer program product comprising computer instructions, characterized in that, The computer instructions are executed by the processor to implement the artificial intelligence-based information recommendation method in any one of claims 1 to 12.

Images