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. The method comprises: obtaining a plurality of ranking features of to-be-recommended information; performing feature cross processing on the plurality of ranking features based on a plurality of cross modes to obtain a plurality of feature vectors corresponding to the plurality of cross modes one by one; determining a target recommendation index prediction network matched with behavior data of a target object in a plurality of candidate recommendation index prediction networks corresponding to different activity levels; performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a predicted recommendation index of the target object corresponding to the to-be-recommended information; and performing a recommendation operation for the target object based on the predicted recommendation index of different to-be-recommended information. Through the application, the recommendation accuracy and efficiency can be improved.

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 related art still recommends to users who are ranked first but not really liked by the user through a unified model for recommendation index estimation. 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 recommendation accuracy and efficiency.

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

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

[0007] Obtain a plurality of ranking features of the information to be recommended;

[0008] Perform feature cross processing on the plurality of ranking features based on a plurality of cross methods to obtain a plurality of feature vectors corresponding to the plurality of cross methods one by one;

[0009] In a plurality of candidate recommendation index prediction networks corresponding to different activity levels, determine a target recommendation index prediction network matching the behavior data of the target object;

[0010] Perform recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a predicted recommendation index of the target object corresponding to the information to be recommended;

[0011] Perform a recommendation operation for the target object based on different predicted recommendation indexes of the information to be recommended.

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

[0013] obtain a plurality of ranking features of to-be-recommended information;

[0014] cross a plurality of ranking features based on a plurality of cross modes to obtain a plurality of feature vectors corresponding to the plurality of cross modes one by one;

[0015] determine a target recommendation index prediction network matching the behavior data of the target object in a plurality of candidate recommendation index prediction networks corresponding to different activity levels;

[0016] predict a recommendation index of the target object for the to-be-recommended information by performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network;

[0017] recommend the target object based on different predicted recommendation indexes of the to-be-recommended information.

[0018] In the above scheme, the obtaining module is further configured to: obtain ranking data of the to-be-recommended information; when the type of the ranking data is an information sequence, encapsulate the ranking data as a numerical feature of a variable-length discrete type; when the type of the ranking data is numerical data, encapsulate the ranking data as a numerical feature of a continuous type; when the type of the ranking data is neither numerical data nor the information sequence, encapsulate the ranking data as a numerical feature of a fixed-length discrete type; and perform embedding processing on a plurality of numerical features to obtain the plurality of ranking features.

[0019] In the above scheme, the obtaining module is further configured to: for any one of the numerical features, perform the following processing: weight a plurality of numerical values in the numerical feature based on a plurality of hidden vectors corresponding to the plurality of numerical values to obtain an embedding dimension value of the numerical feature corresponding to one embedding dimension; and combine embedding dimension values of the numerical feature corresponding to a plurality of embedding dimensions as a ranking feature of the ranking data.

[0020] In the above scheme, the cross module is further configured to: perform a plurality of the following processing: second-order feature cross processing on the plurality of ranking features; explicit deep feature cross processing on the plurality of ranking features; and implicit deep feature cross processing on the plurality of ranking features.

[0021] In the above scheme, the cross module is further configured to: perform multiple combination processing on the plurality of ranking features of the to-be-recommended information to obtain a plurality of combined features of the to-be-recommended information; wherein the ranking features used in each combination processing are partially different or completely different; and perform weighted summation processing on each combined feature of the to-be-recommended information by taking a recommendation index influence factor of each combined feature of the to-be-recommended information as a weight parameter; wherein the recommendation index influence factor of the combined feature is a product of recommendation index correlation influence factors of the ranking features included in the combined feature.

[0022] In the above scheme, the cross module is further configured to: perform k-level explicit feature cross processing to obtain a k-level explicit feature vector; perform pooling processing on the k-level explicit feature vector to obtain a k-level pooling processing result; and perform splicing processing on K-level pooling processing results; wherein K is an integer greater than or equal to 2, k is an integer variable starting from 1 and increasing, the value range of k is 1≤k

[0023] In the above scheme, the cross module is further configured to: when k is 1, perform element-by-element multiplication processing on the plurality of ranking features and the plurality of ranking features; and when k is 2≤k

[0024] In the above scheme, the cross module is further configured to: perform N times of implicit feature cross processing on the plurality of ranking features; wherein N is an integer greater than or equal to 2, n is an integer variable starting from 1 and increasing, the value range of n is 1≤n

[0025] In the above scheme, the cross module is further configured to: in each implicit feature cross processing, perform the following processing: perform full connection processing on the input of the n th implicit feature cross processing to obtain a full connection processing result corresponding to the n th implicit feature cross processing; and perform activation processing on the full connection processing result to obtain an n th implicit feature vector.

[0026] In the above scheme, the prediction module is further configured to: perform weighting processing on the plurality of feature vectors by using the target recommendation indicator prediction network to obtain semantic features of the target object corresponding to the to-be-recommended information; and map the semantic features to the predicted recommendation indicator of the target object corresponding to the to-be-recommended information by using the target recommendation indicator prediction network.

[0027] In the above scheme, the prediction module is further configured to: obtain a plurality of reference target objects whose user portrait similarity to the user portrait of the target object is not less than a portrait similarity threshold; determine reference behavior data of each reference target object, and determine a reference recommendation indicator prediction network that matches the reference behavior data from among a plurality of candidate recommendation indicator prediction networks corresponding to different activity levels; perform recommendation indicator prediction processing on the plurality of feature vectors by using the reference recommendation indicator prediction network to obtain a reference predicted recommendation indicator of the target object corresponding to the to-be-recommended information; perform recommendation indicator prediction processing on the plurality of feature vectors by using the target recommendation indicator prediction network to obtain a predicted recommendation indicator of the target object corresponding to the to-be-recommended information; and correct the predicted recommendation indicator of the target object corresponding to the to-be-recommended information by using the reference predicted recommendation indicator and the corresponding portrait similarity to obtain a new predicted recommendation indicator of the target object corresponding to the to-be-recommended information.

[0028] In the above scheme, the obtaining of the plurality of ranking features of the information to be recommended is implemented by a feature extraction network, the feature cross processing of the plurality of ranking features based on a plurality of cross modes is implemented by an expert network, and the feature extraction network, the expert network, and the plurality of candidate recommendation indicator prediction networks constitute a multi-target object prediction model; the apparatus further comprises a training module configured to: before obtaining the plurality of ranking features of the information to be recommended, obtain a plurality of target object samples corresponding to the information to be recommended samples, wherein the behavior data of the plurality of target object samples is matched with the candidate recommendation indicator prediction networks corresponding to different activity levels respectively; associate each target object sample with the information to be recommended as a training sample corresponding to the target object sample; perform forward propagation of the training sample corresponding to each target object sample in the feature extraction network and the plurality of expert networks to obtain a feature sample of each training sample; perform forward propagation of the feature sample of each training sample in the candidate recommendation indicator prediction network corresponding to the target object sample of the training sample to obtain a training prediction recommendation indicator corresponding to each training sample; determine the error between the training prediction recommendation indicator of each training sample and the corresponding pre-labeled recommendation indicator; and perform back propagation of the error in the multi-target object prediction model to determine the parameter change value of the multi-target object prediction model when the error reaches a minimum value, and update the parameters of the multi-target object prediction model based on the parameter change value.

[0029] In the above scheme, the determination module is further configured to: determine the activity level of the target object based on the behavior data of the target object; and in the plurality of candidate recommendation indicator prediction networks corresponding to different activity levels respectively, determine the candidate recommendation indicator prediction network corresponding to the activity level of the target object as the target recommendation indicator prediction network matched with the behavior data of the target object.

[0030] In the above scheme, the determination module is further configured to: obtain the number of click operations of the target object in a unit time and the online time of the target object in the unit time from the behavior data of the target object; and determine the activity level positively correlated with the number of click operations and negatively correlated with the online time.

[0031] Embodiments of the present application provide an electronic device, comprising:

[0032] a memory configured to store executable instructions;

[0033] a processor configured to execute the executable instructions stored in the memory to implement the information recommendation method based on artificial intelligence provided by the embodiments of the present application.

[0034] The embodiment of the application provides a computer readable storage medium, which stores executable instructions, and is used for implementing the information recommendation method based on artificial intelligence provided by the embodiment of the application when the processor executes the instructions.

[0035] The embodiment of the application has the following beneficial effects:

[0036] Through the cross processing of the sorting features in multiple ways, the multiple feature vectors obtained through the cross processing learn rich semantic information, and then the recommendation index prediction processing is performed on different types of target objects, so that the activity represented by the behavior data of the target object is directly associated with different candidate recommendation index prediction networks, the prediction recommendation index of the target object for the to-be-recommended information is more accurately predicted, precise personalized recommendation is realized for specific target objects, and the recommendation efficiency and the machine resource utilization efficiency are improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a structural schematic diagram of an information recommendation system based on artificial intelligence provided by the embodiment of the application;

[0038] Figure 2 is a structural schematic diagram of an electronic device provided by the embodiment of the application;

[0039] Figures 3A-3D is a flow schematic diagram of an information recommendation method based on artificial intelligence provided by the embodiment of the application;

[0040] Figures 4A-4C is a product performance schematic diagram of the information recommendation method based on artificial intelligence provided by the embodiment of the application;

[0041] Figure 5 is a structural schematic diagram of a multi-target object prediction model provided by the embodiment of the application;

[0042] Figure 6 is a flow schematic diagram of a training stage of the information recommendation method based on artificial intelligence provided by the embodiment of the application;

[0043] Figure 7 is a processing schematic diagram of a compression cross expert network of the information recommendation method based on artificial intelligence provided by the embodiment of the application;

[0044] Figure 8 is a processing schematic diagram of a compression cross expert network of the information recommendation method based on artificial intelligence provided by the embodiment of the application;

[0045] Figure 9 is a structural schematic diagram of a multi-target object prediction model provided by the embodiment of the application. DETAILED DESCRIPTION

[0046] In order to make the purposes, technical solutions and advantages of the present application clearer, the following further describes the present application in conjunction with the accompanying drawings, the described embodiments should not be regarded as limiting the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0047] In the following description, "some embodiments" are referred to, 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.

[0048] In the following description, the terms "first\second\third" are only to distinguish similar objects, and do not represent a specific order of the objects, and it can be understood that "first\second\third" 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.

[0049] 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.

[0050] The relevant 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 within the scope of authorization of laws and regulations and the personal information subject, carry out subsequent data use and processing.

[0051] Before further detailing the embodiments of the present application, the terms and terms involved in the embodiments of the present application are explained, the terms and terms involved in the embodiments of the present application are applicable to the following explanations.

[0052] 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.

[0053] 2) Target object, i.e. the target of information recommendation, since the medium of information presentation is a terminal, the target of information recommendation is the user operating the corresponding terminal, therefore, "object" and "user" are described equivalently in the following. It can be understood that the user here can be a natural person capable of operating the terminal, or a robot program capable of simulating human behavior running in the terminal.

[0054] 3) Information to be recommended: i.e. information that can be sent to the terminal for presentation to the user (target object) of the corresponding terminal for recommendation.

[0055] In the related art, the recommendation accuracy is improved by modeling the recommendation index prediction task, for example, by using one model to serve all users to predict the predicted recommendation index of all users for multiple to-be-recommended information, such as predicted click rate. The applicant found during implementation of the embodiments of the present application that there are differences in user data distribution for different types of users, and therefore proposed the technical concept of independent modeling. In independent modeling, a data stream of high-active users and a data stream of low-active users need to be divided, and two independent models with the same structure are used to fit the two data streams respectively, and an independent model corresponding to high-active users and an independent model corresponding to low-active users are constructed respectively. High-active users (high-active users) and low-active users (low-active users) are distinguished according to user behavior data. The low-active users are users whose interactive operation times in a unit time do not exceed a time threshold, and the high-active users are users whose interactive operation times in a unit time exceed the time threshold. The interactive operations include click operations, forwarding operations, and the like.

[0056] During implementation of the embodiments of the present application, the applicant found that for the unified modeling manner, one model is used to serve all users, and this manner cannot explicitly take into account the data distribution differences between high-active users and low-active users. The applicant found during implementation of the embodiments of the present application that the data distribution differences between high-active users and low-active users are large. For example, the data of high-active users is more complete, and the proportion of positive samples is higher, while the data missing rate and noise rate of low-active users are larger, and the proportion of positive samples is lower. However, low-active users are the core user group of the recommendation system, and in some business scenarios, the recommendation system often needs to pay more attention to the recommendation effect for low-active users. For example, in the recommendation scenario, high-active users already have stickiness, and therefore low-active users are the user group that the recommendation system needs to focus on. If the recommendation system cannot serve low-active users well, the user stickiness and user retention rate will be negatively affected. For the independent modeling manner, two models are used to fit the data stream of high-active users and the data stream of low-active users respectively. Although this manner can explicitly take into account the data distribution differences between high-active users and low-active users, more machine resources are consumed for online reasoning, and the online engineering complexity is often doubled, which is difficult to truly implement.

[0057] The embodiment of the present application provides an information recommendation method and device based on artificial intelligence, electronic equipment and computer readable storage medium, which can explicitly take into account the data difference distribution of high-active users and low-active users without occupying more machine resources, thereby improving the recommendation accuracy and efficiency of the recommendation system without improving the resource occupancy rate. The following describes an exemplary application of the electronic equipment provided by the embodiment of the present application. The electronic equipment provided by the embodiment of the present application can be a server. The following describes an exemplary application when the electronic equipment is implemented as a server.

[0058] Referring to Figure 1 , Figure 1 is a structural schematic diagram of an information recommendation system based on artificial intelligence provided by the embodiment of the present application. Taking the recommendation index of click rate and the predicted recommendation index of predicted click rate as examples, the following describes an exemplary application of the information recommendation system based on artificial intelligence provided by the embodiment of the present application. In order to support a news application, the terminal 400 is connected to the server 200 (exemplarily, the server 200 includes a training server 200-1 and an application server 200-2) through the 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-target object prediction model to the application server 200-2. The terminal 400 used by the user sends a user request to the application server 200-2. The application server 200-2 determines the predicted click rate of the user for a plurality of to-be-recommended information, and determines the to-be-recommended information with a high predicted click rate to return to the terminal 400 for presentation.

[0059] In some embodiments, when the information recommendation system is applied to a video recommendation scenario, the terminal 400 receives a video to be uploaded. The terminal 400 sends the video to the application server 200-2. The application server 200-2 determines the to-be-recommended information with a high predicted click rate through the multi-target object prediction model and sends it to the terminal 400, so that the terminal 400 directly presents the to-be-recommended information with a high predicted click rate in the video recommendation home page.

[0060] In some embodiments, when the information recommendation method provided by the embodiment of the present application is implemented by a terminal alone, in the various application scenarios described above, the terminal can run the multi-target object prediction model to determine the to-be-recommended information with a high predicted click rate, and directly present the to-be-recommended information with a high predicted click rate on the terminal.

[0061] In some embodiments, the training server 200-1 and the application server 200-2 can be a standalone physical server, or a server cluster or a 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, and the like, but is not limited thereto. The terminal and the server can be connected directly or indirectly through wired or wireless communication, and the present embodiment is not limited thereto.

[0062] Referring to Figure 2 , Figure 2 is a structural schematic diagram of an electronic device provided by the present embodiment, taken the application server 200-2 as an example, Figure 2 The application server 200-2 shown in the figure includes at least one processor 210, a memory 250, and at least one network interface 220. The various components in the application server 200-2 are coupled together through a bus system 240. It can be understood that the bus system 240 is used to realize the connection and communication between the components. In addition to the data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for the purpose of clear illustration, all kinds of buses are marked as the bus system 240 in the figure. Figure 2 The processor 210 can be an integrated circuit chip with signal processing capability, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.

[0063] The processor 210 can be an integrated circuit chip with signal processing capability, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.

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

[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 present embodiment 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 a subset or superset 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 Fidelity (WiFi), and Universal Serial Bus (USB), etc.

[0069] In some embodiments, the artificial intelligence-based information recommendation device provided by the embodiments of the present application can be implemented in a software manner, Figure 2 An artificial intelligence-based information recommendation device 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 cross module 2552, a determination module 2553, a prediction module 2554, a recommendation module 2555, and a training module 2556. These modules are logical, so they can be combined or further split according to the functions implemented, and the functions of each module will be described below.

[0070] The artificial intelligence-based information recommendation method provided by the embodiments of the present application will be described in conjunction with the exemplary applications and implementations of the application server 200-2 provided by the embodiments of the present application.

[0071] Referring to Figure 9 , Figure 9is a structural schematic diagram of a multi-target object prediction model provided by an embodiment of the present application. If applied to a news recommendation system, there are two candidate recommendation indicator prediction networks matched with different activity levels, for example, a candidate recommendation indicator prediction network A corresponding to a low-active user (a user with low activity) and a candidate recommendation indicator prediction network B corresponding to a high-active user (a user with high activity). The high-active user and the low-active user are distinguished according to the behavior data of the user. The high-active user (a user with high activity) and the low-active user (a user with low activity) are distinguished according to the behavior data of the user. The user with low activity is a user whose interactive operation times in a unit time do not exceed a times threshold. The user with high activity is a user whose interactive operation times in a unit time exceed a times threshold. The interactive operation includes a click operation, a forwarding operation, etc. A candidate recommendation indicator prediction network includes a prediction fully connected layer A and a threshold network A. Hereinafter, the recommendation indicator is taken as an example of a click rate, and the prediction of the recommendation indicator is taken as the prediction of the click rate. The recommendation indicator includes a click rate, a reading completion degree, and other interactive evaluation parameters. The prediction fully connected layer A outputs a prediction click rate A of a target object belonging to a low-active user for the to-be-recommended information. Another candidate recommendation indicator prediction network includes a prediction fully connected layer B and a threshold network B. The prediction fully connected layer B outputs a prediction click rate B of a target object belonging to a high-active user for the to-be-recommended information. The two candidate recommendation indicator prediction networks and the candidate recommendation indicator prediction network share a plurality of expert networks (for example, an expert network 0, an expert network 1, and an expert network 2). Each expert network corresponds to a different feature cross mode. The two candidate recommendation indicator prediction networks and the candidate recommendation indicator prediction network also share a feature input layer.

[0072] Referring to Figure 3A , Figure 3A is a flowchart of an information recommendation method based on artificial intelligence provided by an embodiment of the present application. Steps 101-105 shown in Figure 3A will be described.

[0073] In step 101, a plurality of ranking features of to-be-recommended information are obtained.

[0074] In some embodiments, referring to Figure 3B , Figure 3B is a flowchart of an information recommendation method based on artificial intelligence provided by an embodiment of the present application. In step 101, a plurality of ranking features of to-be-recommended information are obtained, which can be implemented by steps 1011-1015.

[0075] In step 1011, ranking data of to-be-recommended information is obtained.

[0076] As an example, the type of the sorting data includes at least one of the following: an information sequence to which the behavior data of the target object is directed, a category of the to-be-recommended information, user portrait data of the target object, behavior data of the target object, and behavior data of a reference target object related to the target object.

[0077] In step 1012, when the type of the sorting data is an information sequence, the sorting data is encapsulated into a numerical feature of a variable-length discrete type.

[0078] As an example, when the type of the sorting data is a to-be-recommended information sequence, such as a to-be-recommended information sequence clicked by the target object, a to-be-recommended information sequence exposed but not clicked, a to-be-recommended information sequence exposed and clicked, a to-be-recommended information sequence liked, a to-be-recommended information sequence commented, a to-be-recommended information sequence forwarded, a to-be-recommended information sequence collected, a to-be-recommended information sequence followed, a to-be-recommended information sequence reported, a main sequence of a to-be-recommended information publishing public number followed by the target object, etc., these to-be-recommended information sequences are encapsulated into numerical features of a variable-length discrete type, for example, for a to-be-recommended information sequence (to-be-recommended information 1, to-be-recommended information 2, and to-be-recommended information 3), a numerical feature in the form of a vector corresponding to the to-be-recommended information 1, a numerical feature in the form of a vector corresponding to the to-be-recommended information 2, and a numerical feature in the form of a vector corresponding to the to-be-recommended information 3 are encapsulated, and the numerical feature in the form of a vector is a vector composed of multiple numerical values.

[0079] In step 1013, when the type of the sorting data is numerical data, the sorting data is encapsulated into a numerical feature of a continuous type.

[0080] As an example, when the type of the sorting data is numerical data, such as the age of the target object, a statistical value (exposure number / click number / click rate) of the target object based on multiple to-be-recommended information categories, a statistical value (exposure number / click number / click rate) of to-be-recommended information by a crowd of the current gender of the target object, a statistical value (exposure number / click number / click rate) of to-be-recommended information by a crowd of the current age group of the target object, a statistical number (exposure number / click number / click rate / share number / forward number / follow number / report number / average reading duration / picture number, etc.) of to-be-recommended information, etc., these sorting data are encapsulated into numerical features of a continuous type, for example, when the age of the target object is “15”, “15” is encapsulated into a numerical feature in the form of a vector, and the numerical feature in the form of a vector is a vector composed of multiple numerical values.

[0081] In step 1014, when the type of the sorting data does not belong to numerical data and does not belong to an information sequence, the sorting data is encapsulated into a numerical feature of a fixed-length discrete type.

[0082] As an example, when the type of the sorting data does not belong to numerical data and does not belong to the information sequence to be recommended, for example, the gender of the target object, the information category label to be recommended (entertainment / technology / sports, etc.), the mobile phone brand category of the target object, the mobile phone power of the target object (one grid / two grids / full grid, etc.), the mobile phone brightness of the target object (dark / normal / bright, etc.), the network category of the target object (4G / 5G / Wi-Fi, etc.), the current time period of the target object (early morning / noon / late afternoon, etc.), the gender of the target object (male / female), the age group of the target object (old / middle-aged / young, etc.), the activity level of the target object (low activity / medium activity / high activity), the work category of the target object (programmer / driver / cleaner, etc.), the current location of the target object (Guangdong / Hubei / Beijing, etc.), the interest category of the target object (basketball / reading / singing, etc.), etc., these data are encapsulated as numerical features of the type of fixed-length discrete type, for example, for the gender of the target object “female”, “female” is encapsulated as a numerical feature in the form of a vector, and the numerical feature in the form of a vector is a vector composed of multiple numerical values.

[0083] In step 1015, the multiple numerical features are embedded to obtain multiple sorting features.

[0084] As an example, the multiple numerical features obtained through steps 1012-1014 belong to sparse features, and the numerical values corresponding to positions exceeding a set threshold in the sparse features (vectors) are zero. Therefore, in order to represent all numerical features, the dimension of each numerical feature is high, for example, the dimension of the sparse feature can reach 10,000 dimensions. In order to better perform subsequent feature processing and semantic learning, it is necessary to encapsulate the numerical features into sorting features participating in feature cross processing. The sorting features belong to dense features relative to the sparse features, and the numerical values corresponding to positions exceeding a set threshold in the dense features (vectors) are not zero.

[0085] In some embodiments, the embedding processing of the multiple numerical features to obtain the multiple sorting features can be realized through the following technical solutions: for any numerical feature, the following processing is performed: the multiple numerical values in the numerical feature are weighted by using the hidden vectors corresponding to the multiple numerical values as weights, to obtain an embedding dimension value of the numerical feature corresponding to one embedding dimension; and the embedding dimension values of the numerical feature corresponding to multiple embedding dimensions are combined into the sorting feature of the sorting data.

[0086] As an example, when the embedding dimension of the sorting feature is 5, and the numerical feature has only one position with a numerical value of 1, for the position of the numerical feature, the multiple embedding dimension values of the sorting feature are all 1, and then in the process of obtaining the sorting feature from the input numerical feature, only one neuron in the sparse feature layer works, and the values corresponding to the five lines connected to the neuron are v m1、v m2 、v m3 、v m4 、v m5 The five values are combined to be the sorting feature e m For any embedding dimension of the sorting feature, the value of the embedding dimension is obtained based on the values of multiple positions of the numerical feature, each value contributes to the embedding dimension, and the embedding dimension value corresponding to the n th embedding dimension is the weighted processing result of the value of the first position, the value of the second position, and the value of the third position, and the weighted processing is performed on the five embedding dimensions to obtain the sorting feature e 1n 、v 2n 、v 3n . m .

[0087] In step 102, the plurality of sorting features are subjected to feature intersection processing based on a plurality of intersection modes, and a plurality of feature vectors corresponding one by one to the plurality of intersection modes are obtained.

[0088] In some embodiments, referring to Figure 3C , Figure 3C is a flowchart of the information recommendation method based on artificial intelligence provided by the embodiments of the present application, and the feature intersection processing of the plurality of sorting features based on a plurality of intersection modes in step 102 can be realized by executing a plurality of (i.e. at least two) steps in steps 1021-1023. It should be noted that steps 1021-1023 are parallel, and there is no restriction on the execution sequence.

[0089] In step 1021, the plurality of sorting features are subjected to second-order feature intersection processing.

[0090] In some embodiments, the above-mentioned second-order feature intersection processing of the plurality of sorting features can be realized by the following technical solutions: a plurality of sorting features of the information to be recommended are subjected to multiple combination processing to obtain a plurality of combined features of the information to be recommended; wherein the sorting features used in each combination processing are partially different or completely different; and each combined feature of the information to be recommended is weighted and summed by using the recommendation index influence factor of each combined feature of the information to be recommended as a weight parameter; wherein the recommendation index influence factor of the combined feature is the product of the recommendation index correlation influence factors of the sorting features included in the combined feature.

[0091] As an example, the sorting features used in each combination processing are partially the same, which means that the sorting feature x1 can be combined with the sorting feature x2 to form the combined feature x 12 , and the sorting feature x1 can be combined with the sorting feature x3 to form the combined feature x13 The sorting features used in the two combinations are different, but both are combined with the sorting feature x1 to form a combination feature. The sorting feature x1 can be combined with the sorting feature x2 to form a combination feature x 12 The sorting feature x3 can be combined with the sorting feature x4 to form a combination feature x 34 The sorting features used in the two combinations are completely different.

[0092] As an example, the second-order feature cross processing uses a factorization machine model. Any two features are combined in pairs, and these combined features can be regarded as a new feature. The weight of the combined feature is learned in the training stage. See formula (1):

[0093]

[0094] Where the value of v is determined in the training process, and the component v i The dot product of the two features is the fusion parameter v i The recommendation index correlation influence factor is x i The sorting feature is x, and n is the number of sorting features.

[0095] In step 1022, the plurality of sorting features are subjected to explicit deep feature cross processing.

[0096] In some embodiments, the explicit deep feature cross processing of the plurality of sorting features in step 1022 can be implemented by the following technical solutions: performing k-level explicit feature cross processing to obtain a k-level explicit feature vector; performing pooling processing on the k-level explicit feature vector to obtain a k-level pooling processing result; and splicing the K-level pooling processing results; wherein K is an integer greater than or equal to 2, k is an integer variable with a value increasing from 1, and the value of k ranges from 1 to K. When k is 1, the input of the k-level explicit feature cross processing is the plurality of sorting features. When k is 2≤k

[0097] In some embodiments, the k-level explicit feature cross processing can be implemented by the following technical solutions: when k is 1, the plurality of sorting features are multiplied element by element with the plurality of sorting features; and when k is 2≤k

[0098] As an example, see Figure 7 and Figure 8 ,Figure 7 and Figure 8 is a processing schematic diagram of the compressed cross-expert network of the information recommendation method based on artificial intelligence provided by the embodiment of the application, in order to realize automatic learning of explicit high-order feature interaction, and make the interaction occur at the vector level, the display feature interaction is performed through the calculation principle shown in Figure 7 and Figure 8 , see Figure 8 , the input ranking features and the features of the hidden layer output of the compressed cross-expert network are respectively constituted into a matrix, denoted as X0 (dimension D, m ranking features) and X k , see Figure 7 , the neurons of each hidden layer in the compressed cross-expert network are calculated according to the hidden layer of the previous layer and the ranking features of the original input, the kth hidden layer contains H k neuron vectors, the calculation of the hidden layer can be divided into two steps: (1) according to the state X k of the hidden layer of the previous layer and the matrix X0 of the ranking features of the original input, an intermediate result is calculated, the intermediate result is a three-dimensional tensor; (2) on the basis of the intermediate result, the state of the next layer hidden layer is generated by using H k+1 convolution kernels with a size of m*H k , this operation is basically consistent with the convolutional neural network in computer vision, the only difference is the design of the convolution kernel, the receptive field related to a neuron in the compressed cross-expert network is the entire plane perpendicular to the feature dimension D, and the receptive field is usually a small local area around the current neuron, therefore, the feature map obtained through the convolution operation in the compressed cross-expert network is a vector, not a matrix, see Figure 7 , the order of the learned feature interaction is determined by the number of layers of the network, each layer of the hidden layer is connected to the output unit of the expert network through a pooling operation, so as to ensure that the output unit can obtain different order feature interaction modes, the structure of the compressed cross-expert network is similar to that of the recurrent neural network, that is, the state of each layer is calculated from the value of the hidden layer of the previous layer and an additional input data, but the parameters of different layers in the compressed cross-expert network are different, and the additional input data in the compressed cross-expert network is fixed, that is, the matrix X0 of the ranking features of the original input, through high-order explicit cross, high-order semantic features can be learned, the accuracy of subsequent prediction of the click rate is improved, and the high-order explicit cross has interpretability.

[0099] In step 1023, the plurality of ranking features are processed by implicit deep feature cross.

[0100] In some embodiments, the implicit deep feature cross processing on the plurality of ranking features in step 1023 can be implemented by the following technical solution: performing N times of implicit feature cross processing on the plurality of ranking features; wherein N is an integer greater than or equal to 2, n is an integer variable with a value increasing from 1, the value of n ranges from 1 to N, when n is 1, the input of the n th implicit feature cross processing is the plurality of ranking features, when n is 2≤nN, the input of the n th implicit feature cross processing is the (n-1) th implicit feature vector, and when n is N-1, the output of the (n+1) th implicit feature cross processing is the feature vector.

[0101] In some embodiments, the N times of implicit feature cross processing on the plurality of ranking features can be implemented by the following technical solution: performing the following processing in each implicit feature cross processing: performing full connection processing on the input of the n th implicit feature cross processing to obtain the full connection processing result corresponding to the n th implicit feature cross processing; and performing activation processing on the full connection processing result to obtain the n th implicit feature vector.

[0102] As an example, each implicit feature cross processing is implemented by a plurality of neural networks, each neural network includes a full connection layer and an activation layer, the plurality of ranking features are forward propagated in the N cascaded neural networks to obtain the feature vector obtained by the feature cross processing, and in the processing of each neural network, two operations need to be processed, that is, full connection processing through the full connection layer and activation processing through the activation layer. In the training stage of the neural network, random deletion processing is performed, that is, part of the neurons are randomly discarded. Since the feature full connection processing and the activation processing are performed through the plurality of cascaded neural networks, high-order cross of the features can be implemented. Due to the characteristics of the neural network, the high-order cross of the features is implicit cross. Through the high-order implicit cross, high-order semantic features can be learned, and the accuracy of subsequent prediction of the click rate can be improved.

[0103] In step 103, among the plurality of candidate recommendation indicator prediction networks corresponding to different activity levels, a target recommendation indicator prediction network matching the behavior data of the target object is determined.

[0104] In some embodiments, the determination of the target recommendation indicator prediction network matching the behavior data of the target object among the plurality of candidate recommendation indicator prediction networks corresponding to different activity levels in step 103 can be implemented by the following technical solution: determining the activity level of the target object based on the behavior data of the target object; and among the plurality of candidate recommendation indicator prediction networks corresponding to different activity levels, determining the candidate recommendation indicator prediction network corresponding to the activity level of the target object as the target recommendation indicator prediction network matching the behavior data of the target object.

[0105] In some embodiments, the determining, in step 103, the activity level of the target object based on the behavior data of the target object can be implemented by the following technical solution: obtaining, from the behavior data of the target object, the number of click operations of the target object in a unit time and the online time of the target object in the unit time; and determining the activity level positively correlated with the number of click operations and negatively correlated with the online time.

[0106] For example, the behavior data of the target object records the time stamp of the click operation of the target object, the number of click operations in a unit time is counted, the activity level of the target object is positively correlated with the number of click operations, the more the number of click operations, the more frequent the click operation of the target object in the unit time, and the online time of the target object in the unit time is counted, the online time will reduce the activity level, for example, the online time of a certain target object in a month is as long as 600 hours, although the number of click operations in a month is 600, but there is only one click operation per hour on average, which represents a low activity level of the target object, but the online time of a certain target object in a month is only 10 hours, although the number of click operations in a month is 300, but there are thirty click operations per hour on average, which represents a high activity level of the target object, the high and low of the activity level can be distinguished by interval, for example, the ratio of the number of click operations of the target object in a unit time to the online time of the target object in a unit time is determined to belong to an interval, and the activity level is determined, for example, there are a first interval, a second interval and a third interval. When the ratio falls within the first interval, it represents that the target object belongs to a low activity user; when the ratio falls within the second interval, it represents that the target object belongs to a normal activity user; when the ratio falls within the third interval, it represents that the target object belongs to a high activity user. Each activity level corresponds to a candidate recommendation index prediction network, that is, there are three candidate recommendation index prediction networks corresponding to low activity, normal activity and high activity respectively. The activity level represented by the behavior data of the target object is directly associated with different candidate recommendation index prediction networks, which can more accurately predict the predicted click rate of the target object for the to-be-recommended information.

[0107] For example, the click operation can be interpreted as a click operation for a certain to-be-recommended information in a narrow sense, or can be interpreted as an active operation including a like operation, a forwarding operation, a comment operation, etc.

[0108] In step 104, the target recommendation index prediction network is used to perform click rate prediction processing on the plurality of feature vectors to obtain the predicted click rate of the target object for the to-be-recommended information.

[0109] In some embodiments, referring to Figure 3D , Figure 3Dis a flowchart of an information recommendation method based on artificial intelligence provided by the embodiment of the application. In step 104, the target recommendation index prediction network is used to perform click rate prediction processing on the plurality of feature vectors to obtain the predicted click rate of the target object for the to-be-recommended information. This can be achieved by performing steps 1041-1042.

[0110] In step 1041, the target recommendation index prediction network is used to perform weighting processing on the plurality of feature vectors to obtain the semantic feature of the target object for the to-be-recommended information.

[0111] In step 1042, the target recommendation index prediction network is used to map the semantic feature to the predicted click rate of the target object for the to-be-recommended information.

[0112] As an example, the target recommendation index prediction network includes a threshold network and a corresponding prediction full connection layer. The plurality of feature vectors output based on the plurality of feature cross processing modes are weighted and summed by using the weight parameters configured by the threshold network to obtain the semantic feature, and then the semantic feature is mapped to the predicted click rate of the target object for the to-be-recommended information by using the prediction full connection layer. Since the target recommendation index prediction network corresponds to the activity level represented by the behavior data of the target object, the target recommendation index prediction network can accurately fit the predicted click rate of the target object for the to-be-recommended information with the corresponding activity level.

[0113] In some embodiments, in step 104, the target recommendation index prediction network is used to perform click rate prediction processing on the plurality of feature vectors to obtain the predicted click rate of the target object for the to-be-recommended information. This can be achieved by the following technical solution: obtaining a plurality of reference target objects with a portrait similarity of a user portrait of the target object not less than a portrait similarity threshold; determining reference behavior data of each reference target object, and determining a reference recommendation index prediction network matching the reference behavior data in a plurality of candidate recommendation index prediction networks corresponding to different activity levels; performing click rate prediction processing on the plurality of feature vectors by using the reference recommendation index prediction network to obtain the reference predicted click rate of the target object for the to-be-recommended information; performing click rate prediction processing on the plurality of feature vectors by using the target recommendation index prediction network to obtain the predicted click rate of the target object for the to-be-recommended information; and correcting the predicted click rate of the target object for the to-be-recommended information by using the reference predicted click rate and the corresponding portrait similarity to obtain a new predicted click rate of the target object for the to-be-recommended information.

[0114] As an example, on the basis of steps 1041-1042, the reference prediction click rate of the target object for the to-be-recommended information can be predicted, and the prediction click rate in step 1042 is corrected by the reference prediction click rate. First, a plurality of reference target objects with a portrait similarity of a user portrait of the target object not less than a portrait similarity threshold are obtained, so that the obtained reference target objects are users with similar interests to the target object. The reference behavior data of each reference target object is determined, and the reference recommendation index prediction network matched with the reference behavior data is determined in the candidate recommendation index prediction network corresponding to a plurality of different active degrees. Although the reference target object has similar interests to the target object, the active degree represented by the behavior data of the reference target object may be different from the active degree represented by the behavior data of the target object. The reference recommendation index prediction network is used to perform click rate prediction processing on a plurality of feature vectors to obtain the reference prediction click rate of the target object for the to-be-recommended information. The prediction click rate of the target object for the to-be-recommended information is corrected by the reference prediction click rate and the corresponding portrait similarity to obtain the new prediction click rate of the target object for the to-be-recommended information. For example, the reference prediction click rate obtained for a certain reference target object is 0.8, the portrait similarity of the user portrait of the reference target object to the user portrait of the target object is 0.9, and the prediction click rate of the target object for the to-be-recommended information is 0.5. Therefore, the portrait similarity and the reference prediction click rate can be multiplied, and the multiplication result and the prediction click rate 0.5 can be averaged to obtain the new prediction click rate of the target object for the to-be-recommended information. The prediction result of the target object is corrected by the prediction result of the reference target object, which can effectively improve the prediction accuracy and has interpretability.

[0115] In step 105, a recommendation operation for the target object is performed based on the prediction click rates of different to-be-recommended information.

[0116] As an example, the recommendation order of a plurality of to-be-recommended information is determined according to the prediction click rate of the to-be-recommended information, and the recommendation operation for the corresponding to-be-recommended object is performed based on the recommendation order of the plurality of to-be-recommended information, that is, the to-be-recommended information is sorted in descending order according to the prediction click rate, at least one to-be-recommended information with a high ranking is selected from the descending order sorting result, and the recommendation operation for the corresponding to-be-recommended object is performed based on the selected to-be-recommended information.

[0117] As an example, in the result of descending order sorting, a plurality of recommended information ranked in the front are 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 aggregation processing modes of the plurality of recommendation tasks, so as to obtain the recommended information liked by the to-be-recommended object from a more comprehensive perspective.

[0118] In some embodiments, the plurality of sorting features of the to-be-recommended information obtained in step 101 are obtained through a feature extraction network, and the feature cross processing of the plurality of sorting features based on a plurality of cross modes in step 102 is implemented through an expert network. The feature extraction network, the expert network, and the plurality of candidate recommendation indicator prediction networks constitute a multi-target object prediction model. Before obtaining the plurality of sorting features of the to-be-recommended information in step 101, a plurality of target object samples corresponding to the to-be-recommended information samples are obtained, wherein the behavior data of the plurality of target object samples are matched with the candidate recommendation indicator prediction networks corresponding to the plurality of different activity levels. Each target object sample is associated with the to-be-recommended information as a training sample corresponding to the target object sample. The training sample corresponding to each target object sample is forward propagated in the feature extraction network and the plurality of expert networks to obtain a feature sample of each training sample. The feature sample of each training sample is forward propagated in the candidate recommendation indicator prediction network corresponding to the target object sample of the training sample to obtain a training prediction click rate corresponding to each training sample. The error between the training prediction click rate of each training sample and the corresponding pre-labeled click rate is determined. The error is back propagated in the multi-target object prediction model to determine the parameter change value of the multi-target object prediction model when the error reaches the minimum value, and the parameters of the multi-target object prediction model are updated based on the parameter change value.

[0119] As an example, before the multi-target object prediction model is put into service, a large number of training samples need to be used to fully train the multi-target object prediction model, so that the multi-target object prediction model fits the data distribution of the samples, the trainable parameters of the multi-target object prediction model can be fully learned, and through the online recommendation system, the historical click logs of each target object sample are extracted; the to-be-recommended information sample in the historical click label and the target object sample corresponding to the historical click label are spliced into available training samples, and target object samples that exist exposure label but do not exist historical click label for to-be-recommended information samples can also be obtained, and spliced into available training samples with to-be-recommended information samples. The historical click label is an effective click label, and the effective click label is for a click operation that enters the information detail page for more than 5 seconds. The target object sample without a historical click label represents that the target object sample has browsed the to-be-recommended information sample, but has not clicked the to-be-recommended information sample. After a large number of training samples are obtained, a series of preprocessing such as null value filling, abnormal sample elimination, and outlier feature value correction need to be performed on the training samples. After the processing is completed, the training set and the validation set are divided, the multi-target object prediction model is trained based on the divided training set and the validation set, and the forward propagation and the reverse parameter update are performed. When the training index on the validation set no longer improves, the training is stopped, and the multi-target object prediction model with fixed parameters is output for online recommendation service.

[0120] As an example, for the following training sample: the target object sample of the training sample exists a click operation for the to-be-recommended information sample, the error is calculated by using the following formula (2):

[0121] L = -y i logp i (2);

[0122] Wherein, L is the error, p i is the training prediction click rate, y i is the pre-labeled click rate (usually takes 1 when there is a click operation).

[0123] As an example, for the following training sample: the target object sample of the training sample does not exist a click operation for the to-be-recommended information sample, but the target object sample has browsed the to-be-recommended information, the error is calculated by using the following formula (3):

[0124] L = -(1-y i ) log (1-p i ) (3);

[0125] Wherein, L is the error, p i is the training prediction click rate, y i is the pre-labeled click rate (usually takes 0 when there is no click operation).

[0126] Below, an exemplary application of the embodiments of the present application in an actual application scenario will be described.

[0127] In some embodiments, the information recommendation method based on artificial intelligence provided by the embodiments of the present application is applied to a news recommendation application scenario. A training server pushes a trained multi-target object prediction model to an application server. A terminal used by a user sends a user request to the application server. The application server determines a predicted click rate of the user for a plurality of to-be-recommended information, and determines to-be-recommended information with a high predicted click rate to return to the terminal for presentation.

[0128] Referring to Figures 4A-4C , Figures 4A-4C is a product performance schematic diagram of the information recommendation method based on artificial intelligence provided by the embodiments of the present application, Figure 4A a message page 501A of a social client is presented in Figure 4C , in response to a click operation on a news product conversation bar 502A in the message page 501A, switching to Figure 4B a recommendation main page 501C of the news product in Figure 4C , the recommendation main page 501C includes various information recommended to the user by a personalized recommendation algorithm. The main purpose of the information recommendation method based on artificial intelligence provided by the embodiments of the present application is to fit the data distribution of high and low active users by a multi-target model, so as to fully mine the data distribution difference of high and low active users, recommend the most suitable information to users with different activity levels, and thus improve the recommendation accuracy and efficiency of the recommendation system.

[0129] In some embodiments, referring to Figure 5 , Figure 5 is a structural schematic diagram of the multi-target object prediction model provided by the embodiments of the present application. The multi-target object prediction model includes a feature input layer, a semantic layer, and an output layer. The feature input layer can be implemented as a feature extraction network. The semantic layer can be implemented as an expert network. The threshold network and the output layer can be implemented as a candidate recommendation index prediction network.

[0130] In some embodiments, the input layer mainly includes three parts, which are variable-length discrete features, fixed-length discrete features and continuous features, and each feature is described as follows: 1. Variable-length discrete features, which refer to discrete features with indefinite length, often sequence features, such as information sequence clicked by the user, etc. The variable-length discrete features mainly include: information sequence features exposed but not clicked, information sequence features exposed and clicked, information sequence features liked, information sequence features commented, information sequence features forwarded, information sequence features collected, information sequence features followed, information sequence features reported, main sequence features of information publishing public accounts followed by the user, etc. 2. Fixed-length discrete features, which refer to discrete features with fixed length. The fixed-length discrete features have discrete values, and the values have no actual mathematical meaning and are not used to represent sequences, such as the gender of the user, etc. The fixed-length discrete features mainly include: information category labels (entertainment / technology / sports, etc.), user mobile phone brand categories, user mobile phone power (one grid / two grids / full grid, etc.), user mobile phone brightness (dark / normal / bright, etc.), user network categories (4G / 5G / Wi-Fi, etc.), user current time period (early morning / noon / late afternoon, etc.), user gender (male / female), user age group (old / middle-aged / young, etc.), user activity level (low activity / medium activity / high activity), user work category (programmer / driver / cleaner, etc.), user current location (Guangdong / Hubei / Beijing, etc.), user interest category (basketball / reading / singing, etc.), etc. 3. Continuous features, which refer to features with continuous values and actual mathematical meaning, such as the age of the user, etc. The continuous features mainly include: user statistical values based on multiple information categories (exposure number / click number / click rate), statistical values of the current gender of the user (exposure number / click number / click rate), statistical values of the current age group of the user (exposure number / click number / click rate), statistical numbers of information (exposure number / click number / click rate / share number / forward number / follow number / report number / average reading time / picture number, etc.), etc.

[0131] In some embodiments, in order to enable the features of the input layer to cross more rich semantic implication representation, and provide the output layer corresponding to different target objects with the implicit semantic features that can cover high and low active different groups, the multi-target object prediction model is structurally designed and optimized at the semantic layer, so that the multi-target object prediction model can more fully mine the data distribution differences of high and low active users. The vector representation of the variable-length discrete features, fixed-length discrete features and continuous features of the input layer is input to multiple expert networks (Experts). In order to improve the feature crossing ability of the expert network, a Bi-Interaction, a Compressed Interaction Network (CIN) and a Deep Module are used. The Bi-Interaction performs a second-order cross of the vector representation by a factorization machine. The CIN performs an explicit high-order cross of the vector representation. The Deep Module performs an implicit high-order cross of the vector representation. A deformable convolutional cross expert network and an automatic cross expert network can also be used. The implicit high-order cross includes multiple hidden layer processing. Each hidden layer processing includes full connection processing and activation function processing. Random neuron deletion processing is also included in the training. After the expert network performs the cross operation, the feature vector corresponding to each expert network is obtained. The threshold network corresponding to different target users is used to weight the feature vector corresponding to each expert network, so as to obtain different semantic features for different target users. Forward attention processing and self-attention processing can also be used when weighting.

[0132] In some embodiments, in order to realize automatic learning of explicit high-order feature interaction, and enable the interaction to occur at the vector level, the explicit feature interaction is performed by the calculation principle shown in Figure 7 and Figure 8 . Referring to Figure 8 , the implicit vector is a single unit object. Therefore, the input ordered features and the features output by the hidden layer of the compressed interaction expert network are respectively formed into a matrix, denoted as X0and X k . The neurons of each hidden layer in the compressed interaction expert network are calculated based on the hidden layer of the previous layer and the ordered features of the original input. The kth hidden layer contains H k neuron vectors. The calculation of the hidden layer can be divided into two steps: (1) based on the state X k of the hidden layer of the previous layer and the matrix X0of the ordered features of the original input, an intermediate result is calculated. The intermediate result is a three-dimensional tensor; (2) based on the intermediate result, H k+1 neurons with a size of m*H kThe convolution kernel generates the state of the next layer of the hidden layer, which is generally consistent with the convolutional neural network in computer vision, the only difference being the design of the convolution kernel. The receptive field of a neuron in the compression cross expert network is the entire plane perpendicular to the feature dimension D. The receptive field is usually a small local area around the current neuron. Therefore, the feature map obtained by the convolution operation in the compression cross expert network is a vector, not a matrix. See Figure 7 The order of the learned feature interaction is determined by the number of network layers. Each layer of the hidden layer is connected to the output unit of the expert network through a pooling operation, thereby ensuring that the output unit can obtain different orders of feature interaction patterns. The structure of the compression cross expert network is similar to that of the recurrent neural network, that is, the state of each layer is calculated from the value of the previous layer of the hidden layer and an additional input data. However, the parameters of different layers in the compression cross expert network are different, and the additional input data in the compression cross expert network is fixed, which is always the matrix X0 of the sorted features of the original input.

[0133] In some embodiments, the multi-target object prediction model provided by the embodiments of the present application finally needs to predict the high-active click rate or the low-active click rate of the user. The output layer is composed of two task modules, namely the high-active task module and the low-active task module. If the target object is a high-active target user, the semantic layer outputs the semantic features corresponding to the high-active target user to input into the high-active task module. The high-active task module maps and converts the semantic features through the full connection layer corresponding to the high-active target object, and finally determines the predicted click rate of the target object for the to-be-recommended information through the activation function. If the target object is a high-active target user, the semantic layer outputs the semantic features corresponding to the low-active target user to input into the low-active task module. The low-active task module maps and converts the semantic features through the full connection layer corresponding to the low-active target object, and finally determines the predicted click rate of the target object for the to-be-recommended information through the activation function.

[0134] In some embodiments, see Figure 6 , Figure 6is a flowchart of a training phase of an information recommendation method based on artificial intelligence provided by the embodiment of the application. Before the multi-target object prediction model is put into service, a large number of training samples are used to fully train the multi-target object prediction model, so that the multi-target object prediction model fits the data distribution of the samples, the trainable parameters of the multi-target object prediction model can be fully learned, and the training process of the multi-target object prediction model includes five stages: in the first stage (user's historical clicks), the historical click logs of the user samples are extracted through the online recommendation system's burying point service; in the second stage (constructing samples), the sorting features are extracted through the feature reporting system, the sorting features include variable-length discrete features, fixed-length discrete features and continuous features, the feature data and the user samples corresponding to the historical click labels are spliced into usable training samples, the historical click label is an effective click label, and the effective click label is directed to a click operation of entering an information detail page for more than 5 seconds; in the third stage (dividing training set and validation set), after a large number of training samples are obtained, a series of preprocessing such as null value filling, abnormal sample elimination and outlier feature value correction are required, and the training set and the validation set are divided after the processing; in the fourth stage (training model), based on the divided training set and the validation set, the multi-target object prediction model is fully trained, and the early stopping strategy is used to train the multi-target object prediction model, that is, the training is stopped when the training index on the validation set no longer improves; in the fifth stage (output model), the multi-target object prediction model with fixed parameters is output for online recommendation service.

[0135] In some embodiments, the training phase of the information recommendation method based on artificial intelligence provided by the embodiment of the application uses a cross-entropy segmented loss function as the overall loss function for training, and the overall loss function is shown in formula (4):

[0136]

[0137] wherein L total is the overall loss, is the high activity loss of the i th training sample, is the low activity loss of the i th sample, and N is the total number of training samples.

[0138] In some embodiments, the high activity loss of the i th training sample is shown in formula (5):

[0139]

[0140] wherein, is the high activity loss of the i th training sample, p i is the predicted click rate probability, y i is the actual index value (usually taking the value 1), user iThe user of the i-th training sample.

[0141] In some embodiments, the low-active loss of the i-th training sample is seen from formula (6):

[0142]

[0143] wherein, is the low-active loss of the i-th training sample, p i is the predicted click rate probability, y i is the actual index value (usually taking the value of 1), user i The user of the i-th training sample.

[0144] In some embodiments, the above-mentioned loss is explained as error. For each training sample, there is a corresponding error. Whether it is a low-active user training sample or a high-active user training sample, the error is calculated in the form of the cross-entropy loss function shown in formulas (5) and (6). The AdamW optimization algorithm is used to solve the optimization problem. The AdamW optimization algorithm integrates the adaptive learning rate and the momentum term, and adopts the weight decay strategy, so that the multi-target object prediction model requires fewer resources and converges faster, and can efficiently find a local optimal solution to the optimization problem.

[0145] In some embodiments, the multi-target object prediction model provided by the embodiments of the present application can better fit the data distribution difference of high and low active user groups, and the training parameters can be better learned. After the training parameters are fixed, the multi-target object prediction model can be exported and subsequently applied to online recommendation services.

[0146] In some embodiments, the application stage includes the following steps: 1. The application server determines whether the current time is a recommendation opportunity. If so, long discrete features, fixed discrete features, and continuous features are constructed based on user data and data of information to be recommended, thereby generating test feature samples; 2. The test feature samples are processed by feature cross processing at the semantic layer, and the index prediction is performed using the task module matched with the user's activity level in the multi-target object prediction model according to the user's activity level, thereby obtaining the predicted click rate probability of all information to be recommended; 3. All information to be recommended is sorted according to the predicted click rate probability, and the information to be recommended with the highest recommendation score is recommended.

[0147] The embodiment of the application provides a kind of information recommendation method based on artificial intelligence, applied to the ranking model of recommendation system, give consideration to the data distribution problem of learning high active and low active user, to improve the reading conversion effect of different active user groups in recommendation system, through the candidate recommendation index prediction network of high active user and low active user, respectively fitting the data distribution of high active user and low active user, fully model the conversion rate of high active user and low active user, improve the performance of ranking model in high active user and low active user, to improve the number of plays, the amount of browsing, the number of daily active users, conversion rate and other key indicators of recommendation system.

[0148] In some embodiments, compared with the modeling manner in the related art, whether unified modeling or independent modeling, the multi-target object prediction model provided in the embodiment of the application explicitly takes into account the data difference distribution of high active and low active user groups without occupying more machine resources, through novel neural network structure design, the multi-target object prediction model can ingeniously mine the data distribution difference between high active data and low active data, and through double-cross expert network, compressed cross expert network and deep cross expert network, the input ranking features are fully cross-mined, which greatly improves the precision and efficiency of online recommendation, thereby improving user experience.

[0149] The following continues to illustrate the exemplary structure of the information recommendation device 255 based on artificial intelligence provided in the embodiment of the application as a software module, in some embodiments, as shown in Figure 2 The software module in the information recommendation device 255 based on artificial intelligence stored in the memory 250 can include an acquisition module 2551 for acquiring a plurality of ranking features of information to be recommended, a cross module 2552 for performing feature cross processing based on a plurality of cross modes on the plurality of ranking features to obtain a plurality of feature vectors corresponding one-to-one to the plurality of cross modes, a determination module 2553 for determining a target recommendation index prediction network matching the behavior data of the target object in a plurality of candidate recommendation index prediction networks corresponding to different active degrees, a prediction module 2554 for performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a predicted recommendation index of the target object corresponding to the information to be recommended, and a recommendation module 2555 for performing a recommendation operation for the target object based on the predicted recommendation index of different information to be recommended.

[0150] In some embodiments, the obtaining module 2551 is further configured to: obtain ranking data of the information to be recommended; when the type of the ranking data is information sequence, encapsulate the ranking data as a numerical feature of the type of variable-length discrete feature; when the type of the ranking data is numerical data, encapsulate the ranking data as a numerical feature of the type of continuous feature; when the type of the ranking data is neither numerical data nor information sequence, encapsulate the ranking data as a numerical feature of the type of fixed-length discrete feature; and perform embedding processing on the plurality of numerical features to obtain a plurality of ranking features.

[0151] In some embodiments, the obtaining module 2551 is further configured to: for any one numerical feature, perform the following processing: weight the plurality of values in the numerical feature with the plurality of vectors corresponding to the plurality of values respectively as weights to obtain an embedding dimension value of the numerical feature corresponding to one embedding dimension; and combine the embedding dimension values of the numerical feature corresponding to a plurality of embedding dimensions as a ranking feature of the ranking data.

[0152] In some embodiments, the cross module 2552 is further configured to: perform one or more of the following processing: second-order feature cross processing on the plurality of ranking features; explicit deep feature cross processing on the plurality of ranking features; and implicit deep feature cross processing on the plurality of ranking features.

[0153] In some embodiments, the cross module 2552 is further configured to: perform a plurality of combination processing on the plurality of ranking features of the information to be recommended to obtain a plurality of combination features of the information to be recommended; wherein the ranking features used in each combination processing are partially different or completely different; and perform weighted summation processing on each combination feature of the information to be recommended with a recommendation index influence factor of each combination feature of the information to be recommended as a weight parameter; wherein the recommendation index influence factor of the combination feature is a product of the recommendation index correlation influence factors of the ranking features included in the combination feature.

[0154] In some embodiments, the cross module 2552 is further configured to: perform explicit feature cross processing of the kth level to obtain an explicit feature vector of the kth level; perform pooling processing on the explicit feature vector of the kth level to obtain a pooling processing result of the kth level; and perform splicing processing on the K-level pooling processing results; wherein K is an integer greater than or equal to 2, k is an integer variable starting from 1 and increasing, the value range of k is 1≤k

[0155] In some embodiments, the cross module 2552 is further configured to: when k is 1, perform element-wise multiplication processing on the plurality of ranking features and the plurality of ranking features; and when k is 2≤k

[0156] In some embodiments, the cross module 2552 is further configured to: perform N times of implicit feature cross processing on the plurality of ranking features; wherein N is an integer greater than or equal to 2, n is an integer variable starting from 1 and increasing, and n is in the range of 1≤n

[0157] In some embodiments, the cross module 2552 is further configured to: in each implicit feature cross processing, perform the following processing: perform full connection processing on the input of the n th implicit feature cross processing to obtain a full connection processing result corresponding to the n th implicit feature cross processing; and perform activation processing on the full connection processing result to obtain the n th implicit feature vector.

[0158] In some embodiments, the prediction module 2554 is further configured to: perform weighting processing on the plurality of feature vectors by using the target recommendation indicator prediction network to obtain semantic features of the to-be-recommended information corresponding to the target object; and map the semantic features to the predicted recommendation indicator of the to-be-recommended information corresponding to the target object by using the target recommendation indicator prediction network.

[0159] In some embodiments, the prediction module 2554 is further configured to: obtain a plurality of reference target objects whose user portrait similarity with the user portrait of the target object is not less than a portrait similarity threshold; determine reference behavior data of each reference target object, and determine a reference recommendation indicator prediction network matching the reference behavior data in a plurality of candidate recommendation indicator prediction networks corresponding to different activity levels; perform recommendation indicator prediction processing on the plurality of feature vectors by using the reference recommendation indicator prediction network to obtain a reference predicted recommendation indicator of the to-be-recommended information corresponding to the target object; perform recommendation indicator prediction processing on the plurality of feature vectors by using the target recommendation indicator prediction network to obtain a predicted recommendation indicator of the to-be-recommended information corresponding to the target object; and correct the predicted recommendation indicator of the to-be-recommended information corresponding to the target object by using the reference predicted recommendation indicator and the corresponding portrait similarity to obtain a new predicted recommendation indicator of the to-be-recommended information corresponding to the target object.

[0160] In some embodiments, the plurality of ranking features of the information to be recommended are obtained through a feature extraction network, the feature cross processing of the plurality of ranking features based on a plurality of cross modes is implemented through an expert network, and the feature extraction network, the expert network, and the plurality of candidate recommendation indicator prediction networks constitute a multi-target object prediction model; the device further comprises a training module 2556 configured to: before obtaining the plurality of ranking features of the information to be recommended, obtain a plurality of target object samples corresponding to a sample of the information to be recommended, wherein the behavior data of the plurality of target object samples are matched with the candidate recommendation indicator prediction networks corresponding to the plurality of different activity levels; associate each target object sample with the information to be recommended as a training sample of the corresponding target object sample; perform forward propagation of the training sample corresponding to each target object sample in the feature extraction network and the plurality of expert networks to obtain a feature sample of each training sample; perform forward propagation of the feature sample of each training sample in the candidate recommendation indicator prediction network corresponding to the target object sample of the training sample to obtain a training prediction recommendation indicator corresponding to each training sample; determine the error between the training prediction recommendation indicator of each training sample and the corresponding pre-labeled recommendation indicator; and perform back propagation of the error in the multi-target object prediction model to determine the parameter change value of the multi-target object prediction model when the error reaches a minimum value, and update the parameters of the multi-target object prediction model based on the parameter change value.

[0161] In some embodiments, the determining module 2553 is further configured to: determine the activity level of the target object based on the behavior data of the target object; and in the plurality of candidate recommendation indicator prediction networks corresponding to the plurality of different activity levels, determine the candidate recommendation indicator prediction network corresponding to the activity level of the target object as the target recommendation indicator prediction network matched with the behavior data of the target object.

[0162] In some embodiments, the determining module 2553 is further configured to: obtain the number of click operations of the target object in a unit time and the online time of the target object in a unit time from the behavior data of the target object; and determine the activity level positively correlated with the number of click operations and negatively correlated with the online time.

[0163] The embodiments of the present application provide a computer program product or a 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 to enable the computer device to perform the information recommendation method based on artificial intelligence provided in the embodiments of the present application.

[0164] The embodiment of the present application provides a computer readable storage medium storing executable instructions, wherein the executable instructions are executed by a processor to execute the information recommendation method based on artificial intelligence provided by the embodiment of the present application, for example, as shown in the following. Figures 3A-3D The information recommendation method based on artificial intelligence is shown.

[0165] In some embodiments, the computer readable storage medium can be FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disc, or CD-ROM memory, etc.; and can also be various devices including one or any combination of the above memories.

[0166] 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 as modules, components, subroutines or other units suitable for use in a computing environment.

[0167] As an example, the executable instructions can but not necessarily correspond to files in a file system, can be stored in part of a file storing other programs or data, for example, stored in one or more scripts in a Hyper Text Markup Language (HTML) 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 portions).

[0168] As an example, the executable instructions can be deployed to execute on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed at multiple sites and interconnected through a communication network.

[0169] In summary, through the cross processing of the sorting features in multiple ways by the embodiment of the present application, the multiple feature vectors obtained through the cross processing learn rich semantic information, and then the recommendation index prediction processing is performed on different types of target objects, so that the activity represented by the behavior data of the target object is directly associated with different candidate recommendation index prediction networks, the prediction recommendation index of the target object for the to-be-recommended information is more accurately predicted, and the precise personalized recommendation for the specific target object is realized.

[0170] The above is only an embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement and improvement made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. An information recommendation method based on artificial intelligence, characterized in that, include: Multiple ranking features of the information to be recommended are obtained through a feature extraction network; The following feature crossover processes are performed by an expert network to obtain multiple feature vectors corresponding to each feature crossover process: second-order feature crossover process on the multiple ranking features; explicit deep feature crossover process on the multiple ranking features; implicit deep feature crossover process on the multiple ranking features; wherein different feature crossover methods correspond to different expert networks; Based on the behavioral data of the target object, the activity level of the target object is determined, wherein the activity level is positively correlated with the number of clicks per unit time and negatively correlated with the online time per unit time. Among multiple candidate recommendation index prediction networks corresponding to different activity levels, a target recommendation index prediction network that matches the behavioral data of the target object is determined; wherein, the candidate recommendation index prediction networks share the feature extraction network and the expert network; The target recommendation index prediction network performs weighted processing on the multiple feature vectors to obtain the semantic features of the target object corresponding to the information to be recommended; the target recommendation index prediction network then maps the semantic features to a predicted recommendation index for the target object corresponding to the information to be recommended. The predicted recommendation index for the target object corresponding to the information to be recommended is corrected using the following method to obtain a new predicted recommendation index for the target object corresponding to the information to be recommended: Multiple reference target objects with a user profile similarity of not less than a profile similarity threshold to the target object are obtained; reference behavior data for each reference target object is determined, and a reference recommendation index prediction network matching the reference behavior data is selected from multiple candidate recommendation index prediction networks corresponding to different activity levels; recommendation index prediction processing is performed on the multiple feature vectors through the reference recommendation index prediction network to obtain a reference predicted recommendation index for the target object corresponding to the information to be recommended; recommendation index prediction processing is performed on the multiple feature vectors through the target recommendation index prediction network to obtain a predicted recommendation index for the target object corresponding to the information to be recommended; the predicted recommendation index for the target object corresponding to the information to be recommended is corrected using the reference predicted recommendation index and the corresponding profile similarity to obtain a new predicted recommendation index for the target object corresponding to the information to be recommended. Based on the different predictive recommendation metrics of the information to be recommended, a recommendation operation is performed for the target object; Before obtaining multiple ranking features of the information to be recommended through the feature extraction network, multiple target object samples of the information to be recommended are obtained. The behavioral data of these multiple target object samples are matched with candidate recommendation metric prediction networks corresponding to different activity levels. Each target object sample is associated with the information to be recommended as a training sample corresponding to that target object sample. The training samples corresponding to each target object sample are forward propagated in the feature extraction network and the multiple expert networks to obtain feature samples for each training sample. The feature samples of each training sample are forward propagated in the candidate recommendation metric prediction network corresponding to the target object samples of the training samples to obtain the training predicted click-through rate (CTR) for each training sample. The error between the training predicted CTR and the corresponding pre-labeled CTR for each training sample is determined. The error is backpropagated in the multi-target object prediction model to determine the parameter change value of the multi-target object prediction model when the error reaches its minimum value, and the parameters of the multi-target object prediction model are updated based on the parameter change value.

2. The method according to claim 1, characterized in that, The step of determining the target recommendation indicator prediction network that matches the behavioral data of the target object from among the multiple candidate recommendation indicator prediction networks corresponding to different activity levels includes: Among multiple candidate recommendation indicator prediction networks corresponding to different activity levels, the candidate recommendation indicator prediction network corresponding to the activity level of the target object is determined as the target recommendation indicator prediction network that matches the behavioral data of the target object.

3. The method according to claim 1, characterized in that, The multiple ranking features for obtaining the information to be recommended include: Obtain the sorting data of the information to be recommended; When the type of the sorted data is an information sequence, the sorted data is encapsulated as a numerical feature of type variable-length discrete feature; When the sorted data is of numerical type, the sorted data is encapsulated as a numerical feature of continuous type. When the type of the sorted data is not numerical data and does not belong to the information sequence, the sorted data is encapsulated as a numerical feature of type fixed-length discrete feature; The numerical features are embedded to obtain the sorting features.

4. The method according to claim 3, characterized in that, The embedding process of the multiple numerical features to obtain the multiple ranking features includes: For any one of the numerical features, perform the following processing: Using the latent vectors corresponding to multiple values ​​in the numerical features as weights, the multiple values ​​are weighted to obtain the embedding dimension value of an embedding dimension corresponding to the numerical features. The numerical features are combined with the embedding dimension values ​​of multiple embedding dimensions to form the sorting features of the sorted data.

5. The method according to claim 1, characterized in that, The second-order feature cross processing of the multiple sorting features includes: The multiple ranking features of the information to be recommended are combined multiple times to obtain multiple combined features of the information to be recommended. The sorting features used in each of the combined processes are different or completely different. Using the recommendation index influence factor of each of the combined features of the information to be recommended as the weight parameter, a weighted summation is performed on each of the combined features of the information to be recommended. The recommendation index influence factor of the combined feature is the product of the recommendation index association influence factors of the ranking features included in the combined feature.

6. The method according to claim 1, characterized in that, The explicit deep feature cross processing of the multiple ranking features includes: Perform explicit feature cross processing at the k-th level to obtain the explicit feature vector at the k-th level. The explicit feature vectors of the k-th level are pooled to obtain the pooling result of the k-th level. The pooling results from K levels are concatenated. Where K is an integer greater than or equal to 2, and k is an integer variable whose value starts from 1 and increments, with a range of values ​​for k. k When k is 1, the input to the explicit feature cross-processing at the k-th level is the multiple ranking features; when k is 2... k When K, the input to the explicit feature cross-processing of the k-th level is the explicit feature vector of the (k-1)-th level.

7. The method according to claim 6, characterized in that, The explicit feature cross-processing at the k-th level includes: When k is 1, the multiple sorting features are multiplied element by element; When k is 2 k When K, the multiple sorting features are multiplied element-wise with the explicit feature vector of the (k-1)th level.

8. The method according to claim 1, characterized in that, The acquisition of multiple ranking features for the information to be recommended is achieved through the feature extraction network. The feature cross-processing based on multiple cross-methods for the multiple ranking features is implemented through the expert network. The feature extraction network, the expert network, and the multiple candidate recommendation index prediction networks constitute the multi-target object prediction model.

9. An information recommendation device based on artificial intelligence, characterized in that, include: The acquisition module is used to obtain multiple ranking features of the information to be recommended through a feature extraction network. The cross-processing module is used to perform multiple of the following feature cross-processing operations through an expert network to obtain multiple feature vectors corresponding one-to-one with the feature cross-processing operations: performing second-order feature cross-processing on the multiple ranked features; performing explicit deep feature cross-processing on the multiple ranked features; and performing implicit deep feature cross-processing on the multiple ranked features; wherein different feature cross-processing methods correspond to different expert networks; The determination module is used to determine the activity level of the target object based on the target object's behavioral data, wherein the activity level is positively correlated with the number of clicks per unit time and negatively correlated with the online time per unit time; among multiple candidate recommendation indicator prediction networks corresponding to different activity levels, a target recommendation indicator prediction network that matches the behavioral data of the target object is determined; wherein the candidate recommendation indicator prediction networks share the feature extraction network and the expert network; The prediction module is used to weight the multiple feature vectors through the target recommendation metric prediction network to obtain the semantic features of the target object corresponding to the information to be recommended; to map the semantic features to the predicted recommendation metric of the target object corresponding to the information to be recommended through the target recommendation metric prediction network; and to correct the predicted recommendation metric of the target object corresponding to the information to be recommended by the following method to obtain a new predicted recommendation metric of the target object corresponding to the information to be recommended: obtaining multiple reference target objects whose profile similarity to the user profile of the target object is not less than a profile similarity threshold; determining the reference behavior data of each reference target object, and applying it to the multiple different activity levels. A reference recommendation index prediction network matching the reference behavior data is determined from the candidate recommendation index prediction networks respectively; the reference recommendation index prediction network is used to perform recommendation index prediction processing on the multiple feature vectors to obtain a reference predicted recommendation index for the target object corresponding to the information to be recommended; the target recommendation index prediction network is used to perform recommendation index prediction processing on the multiple feature vectors to obtain a predicted recommendation index for the target object corresponding to the information to be recommended; the predicted recommendation index for the target object corresponding to the information to be recommended is corrected using the reference predicted recommendation index and the corresponding profile similarity to obtain a new predicted recommendation index for the target object corresponding to the information to be recommended. The recommendation module is used to perform recommendation operations for the target object based on the predicted recommendation index of different information to be recommended; The training module is used to acquire multiple target object samples of the information to be recommended before acquiring multiple ranking features of the information to be recommended through the feature extraction network. The behavioral data of the multiple target object samples are matched with candidate recommendation metric prediction networks corresponding to different activity levels. Each target object sample is associated with the information to be recommended as a training sample corresponding to the target object sample. The training samples corresponding to each target object sample are forward propagated in the feature extraction network and the multiple expert networks to obtain feature samples for each training sample. The feature samples of each training sample are forward propagated in the candidate recommendation metric prediction network corresponding to the target object samples of the training samples to obtain the training predicted click-through rate (CTR) for each training sample. The error between the training predicted CTR and the corresponding pre-labeled CTR of each training sample is determined. The error is backpropagated in the multi-target object prediction model to determine the parameter change value of the multi-target object prediction model when the error reaches its minimum value, and the parameters of the multi-target object prediction model are updated based on the parameter change value.

10. The apparatus according to claim 9, characterized in that, The acquisition module is further configured to acquire the sorted data of the information to be recommended; when the type of the sorted data is an information sequence, the sorted data is encapsulated into a numerical feature of type variable-length discrete feature; when the type of the sorted data is numerical data, the sorted data is encapsulated into a numerical feature of type continuous feature; when the type of the sorted data is neither numerical data nor an information sequence, the sorted data is encapsulated into a numerical feature of type fixed-length discrete feature; and the multiple numerical features are embedded to obtain the multiple sorted features.

11. The apparatus according to claim 9, characterized in that, The acquisition module is further configured to perform the following processing on any numerical feature: using the latent vectors corresponding to the multiple values ​​in the numerical feature as weights, perform weighted processing on the multiple values ​​to obtain the embedding dimension value of an embedding dimension corresponding to the numerical feature; and combine the embedding dimension values ​​of the multiple embedding dimensions corresponding to the numerical feature to form the sorting feature of the sorted data.

12. An electronic device, characterized in that, include: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the information recommendation method based on artificial intelligence as described in any one of claims 1 to 8.

13. A computer-readable storage medium, characterized in that, It stores executable instructions for implementing the information recommendation method based on artificial intelligence as described in any one of claims 1 to 8 when executed by a processor.

14. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the information recommendation method based on artificial intelligence as described in any one of claims 1 to 8.

Images