CN115221396A

CN115221396A - Information recommendation method and device based on artificial intelligence and electronic equipment

Info

Publication number: CN115221396A
Application number: CN202110430589.9A
Authority: CN
Inventors: 肖小粤; 刘子璐
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-04-21
Filing date: 2021-04-21
Publication date: 2022-10-21

Abstract

The application provides an information recommendation method, an information recommendation device, electronic equipment and a computer-readable storage medium based on artificial intelligence; the method comprises the following steps: acquiring a plurality of sorting features of information to be recommended; performing feature cross processing based on a plurality of cross modes on the plurality of sequencing features 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 the behavior data of the target object in a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees; performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a prediction recommendation index of the target object corresponding to the information to be recommended; and performing recommendation operation aiming at the target object based on the different prediction recommendation indexes of the information to be recommended. Through the method and the device, the recommendation accuracy and the recommendation efficiency can be improved.

Description

Information recommendation method and device based on artificial intelligence and electronic equipment

Technical Field

The present disclosure relates to artificial intelligence technologies, and in particular, to an information recommendation method and apparatus based on artificial intelligence, an electronic device, and a computer-readable storage medium.

Background

Artificial Intelligence (AI) is a theory, method and technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results.

With the development of information technology and internet industry, information overload becomes a challenge for people to process information, and the personalized recommendation system effectively relieves the problems by more accurately understanding the user intention, but in the related art, the information which is ranked ahead but is not really liked by the user can still be recommended to the user in a way of uniformly estimating the recommendation index through a certain model aiming at the whole amount of users.

Disclosure of Invention

The embodiment of the application provides an information recommendation method and device based on artificial intelligence, an electronic device and a computer-readable storage medium, and recommendation accuracy and recommendation efficiency can be improved.

The technical scheme of the embodiment of the application is realized as follows:

the embodiment of the application provides an information recommendation method based on artificial intelligence, which comprises the following steps:

acquiring a plurality of sorting features of information to be recommended;

performing feature cross processing based on a plurality of cross modes on the plurality of sequencing features 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 the behavior data of the target object in a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees;

performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a prediction recommendation index of the target object corresponding to the information to be recommended;

and performing recommendation operation aiming at the target object based on the different prediction recommendation indexes of the information to be recommended.

The embodiment of the application provides an information recommendation device based on artificial intelligence, includes:

the acquisition module is used for acquiring a plurality of sorting features of the information to be recommended;

the cross module is used for carrying out feature cross processing based on a plurality of cross modes on the plurality of sequencing features to obtain a plurality of feature vectors which are in one-to-one correspondence with the plurality of cross modes;

the determining module is used for determining a target recommendation index prediction network matched with the behavior data of the target object in the candidate recommendation index prediction networks respectively corresponding to the different liveness degrees;

the prediction module is used for carrying out recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a prediction recommendation index of the target object corresponding to the information to be recommended;

and the recommending module is used for executing recommending operation aiming at the target object based on the different prediction recommending indexes of the information to be recommended.

In the foregoing solution, the obtaining module is further configured to: obtaining the sequencing data of the information to be recommended; when the type of the sequencing data is an information sequence, packaging the sequencing data into a numerical characteristic with a variable-length discrete characteristic; when the type of the sequencing data is numerical data, packaging the sequencing data into numerical characteristics with the type of continuous characteristics; when the type of the sequencing data does not belong to numerical data and does not belong to the information sequence, packaging the sequencing data into numerical characteristics of which the type is a fixed-length discrete characteristic; and embedding the numerical characteristics to obtain the sequencing characteristics.

In the foregoing solution, the obtaining module is further configured to: performing the following for any one of the numerical characteristics: weighting the numerical values by taking hidden vectors respectively corresponding to the numerical values in the numerical characteristics as weights to obtain an embedding dimension value of the numerical characteristics corresponding to one embedding dimension; and combining the embedded dimension value of the numerical characteristic corresponding to a plurality of embedded dimensions into the sorting characteristic of the sorting data.

In the above solution, the crossing module is further configured to: various processes among the following processes are performed: performing second-order feature cross processing on the plurality of sequencing features; performing explicit depth feature intersection processing on the plurality of ranking features; and performing implicit depth feature intersection processing on the plurality of ordering features.

In the foregoing solution, the crossing module is further configured to: combining the plurality of sorting features of the information to be recommended for a plurality of times to obtain a plurality of combined features of the information to be recommended; wherein the ranking features used for each of the combination processes are partially or completely different; weighting and summing each combined feature of the information to be recommended by taking a recommendation index influence factor of each combined feature of the information to be recommended as a weight parameter; wherein the recommendation index impact factor of the combined feature is a product of recommendation index association impact factors of ranking features included in the combined feature.

In the above solution, the crossing module is further configured to: performing explicit feature cross processing of a kth level to obtain an explicit feature vector of the kth level; pooling the explicit feature vectors of the k level to obtain a pooling result of the k level; splicing the K levels of pooling processing results; the method comprises the steps that K is an integer larger than or equal to 2, K is an integer variable with the value increasing from 1, the value range of K is that K is larger than or equal to 1 and is smaller than K, when the value of K is 1, input of the K-th level explicit feature cross processing is the multiple sorting features, and when the value of K is larger than or equal to 2 and is smaller than K, input of the K-th level explicit feature cross processing is an explicit feature vector of a K-1 level.

In the foregoing solution, the crossing module is further configured to: when k is 1, performing element-by-element multiplication processing on the plurality of sorting features and the plurality of sorting features; and when K is more than or equal to 2 and less than K, performing element-by-element multiplication on the plurality of sorting features and the explicit feature vector of the (K-1) th level.

In the foregoing solution, the crossing module is further configured to: carrying out N times of implicit feature cross processing on the plurality of sorting features; and when the value of N is not less than 2 and less than N, the input of the nth implicit feature cross processing is the implicit feature vector of the nth-1 time, and when the value of N is N-1, the output of the (N + 1) th implicit feature cross processing is the feature vector.

In the above solution, the crossing module is further configured to: the following processing is performed in the process of each implicit feature intersection processing: performing full-connection processing on the input of the nth implicit feature cross processing to obtain a full-connection processing result corresponding to the nth implicit feature cross processing; and activating the full-connection processing result to obtain the nth implicit characteristic vector.

In the above solution, the prediction module is further configured to: weighting the plurality of feature vectors through the target recommendation index prediction network to obtain semantic features of the target object corresponding to the information to be recommended; and mapping the semantic features into the prediction recommendation indexes of the target object corresponding to the information to be recommended through the target recommendation index prediction network.

In the foregoing aspect, the prediction module is further configured to: acquiring a plurality of reference target objects of which the portrait similarity of a user portrait with the target object is not less than a portrait similarity threshold; determining reference behavior data of each reference target object, and determining a reference recommendation index prediction network matched with the reference behavior data in a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees; performing recommendation index prediction processing on the plurality of feature vectors through the reference recommendation index prediction network to obtain a reference prediction recommendation index of the target object corresponding to the information to be recommended; performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain a prediction recommendation index of the target object corresponding to the information to be recommended; and correcting the prediction recommendation index of the target object corresponding to the information to be recommended according to the reference prediction recommendation index and the corresponding portrait similarity to obtain a new prediction recommendation index of the target object corresponding to the information to be recommended.

In the above scheme, the obtaining of the plurality of ranking features of the information to be recommended is realized through a feature extraction network, the performing of feature intersection processing on the plurality of ranking features based on a plurality of intersection modes is realized through an expert network, and the feature extraction network, the expert network and the plurality of candidate recommendation index prediction networks form a multi-target object prediction model; the apparatus further comprises a training module to: before obtaining a plurality of sorting features of information to be recommended, obtaining a plurality of target object samples corresponding to the information samples to be recommended, wherein behavior data of the plurality of target object samples are matched with a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees; associating each target object sample and the information to be recommended as a training sample corresponding to the target object sample; carrying out forward propagation on 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; carrying out forward propagation on the feature sample of each training sample in a candidate recommendation index prediction network corresponding to a target object sample of the training samples to obtain a training prediction recommendation index corresponding to each training sample; determining an error between the training prediction recommendation index of each training sample and the corresponding pre-labeled recommendation index; and reversely propagating the error in the multi-target object prediction model to determine a parameter change value of the multi-target object prediction model when the error is minimum, and updating the parameter of the multi-target object prediction model based on the parameter change value.

In the foregoing solution, the determining module is further configured to: determining the activity of the target object based on the behavior data of the target object; and determining the candidate recommendation index prediction network corresponding to the activity of the target object as a target recommendation index prediction network matched with the behavior data of the target object in a plurality of candidate recommendation index prediction networks respectively corresponding to different activities.

In the foregoing solution, the determining module is further configured to: acquiring the number of times of clicking operation of the target object in unit time and the online time of the target object in the unit time from the behavior data of the target object; and determining the liveness which is positively correlated with the number of the clicking operations and negatively correlated with the online time.

An embodiment of the present application provides an electronic device, including:

a memory for storing executable instructions;

and the processor is used for realizing the artificial intelligence based information recommendation method provided by the embodiment of the application when the executable instructions stored in the memory are executed.

The embodiment of the application provides a computer-readable storage medium, which stores executable instructions and is used for realizing the artificial intelligence-based information recommendation method provided by the embodiment of the application when being executed by a processor.

The embodiment of the application has the following beneficial effects:

through carrying out cross processing on the sequencing characteristics in multiple modes, a plurality of characteristic vectors obtained through the cross processing learn rich semantic information, and then are subjected to recommendation index prediction processing aiming at different types of target objects, so that the activeness represented by behavior data of the target objects is directly associated with different candidate recommendation index prediction networks, and the prediction recommendation indexes of the target objects aiming at information to be recommended are predicted more accurately, and therefore accurate personalized recommendation is realized aiming at specific target objects, the recommendation efficiency is improved, and the utilization efficiency of machine resources is improved.

Drawings

FIG. 1 is a schematic structural diagram of an artificial intelligence-based information recommendation system provided by an embodiment of the present application;

fig. 2 is a schematic structural diagram of an electronic device provided in an embodiment of the present application;

3A-3D are schematic flow diagrams of artificial intelligence based information recommendation methods provided by embodiments of the present application;

4A-4C are product performance diagrams of artificial intelligence based information recommendation methods provided by embodiments of the present application;

FIG. 5 is a schematic structural diagram of a multi-target object prediction model provided in an embodiment of the present application;

FIG. 6 is a flowchart illustrating a training phase of an artificial intelligence based information recommendation method according to an embodiment of the present application;

FIG. 7 is a schematic processing diagram of a compressed cross expert network based on an artificial intelligence information recommendation method according to an embodiment of the present application;

FIG. 8 is a schematic processing diagram of a compressed cross expert network based on an artificial intelligence based information recommendation method according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a multi-target object prediction model according to an embodiment of the present application.

Detailed Description

In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be described in further detail with reference to the accompanying drawings, the described embodiments should not be considered as limiting the present application, and all other embodiments obtained by a person of ordinary skill in the art without making creative efforts fall within the protection scope of the present application.

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

In the following description, references to the terms "first \ second \ third" are only to distinguish similar objects and do not denote a particular order or importance, but rather "first \ second \ third" may, where permissible, be interchanged in a particular order or sequence so that embodiments of the present application described herein can be practiced in other than the order shown or described herein.

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

Before further detailed description of the embodiments of the present application, terms and expressions referred to in the embodiments of the present application will be described, and the terms and expressions referred to in the embodiments of the present application will be used for the following explanation.

1) The recommendation system comprises: recommendation systems are a tool for automatically contacting users and information, which can help users find information of interest to them in an information overload environment, and can push information to users of interest to them.

2) Since the target object, i.e., the target for information recommendation, is a terminal and the target for information recommendation is a user operating the corresponding terminal, the "object" and the "user" are equivalently described hereinafter. It is understood that the user may be a natural person capable of operating the terminal, or a robot program running in the terminal and capable of simulating human behavior.

3) And information to be recommended: i.e. information that can be sent to the terminal for presentation for recommendation to the user (target object) of the respective terminal.

In the related art, recommendation accuracy is improved by modeling a recommendation index prediction task, for example, all users are served by the same model to predict a prediction recommendation index of all users for a plurality of information to be recommended, such as a predicted click rate, and the applicant finds that user data distribution is different for different types of users when implementing the embodiment of the present application, so that the applicant proposes a technical concept of independent modeling. The high-activity users (users with high activity) and the low-activity users (users with low activity) are distinguished according to the behavior data of the users, the users with low activity are users with the number of interactive operations within unit time not exceeding a number threshold, the users with high activity are users with the number of interactive operations within unit time exceeding the number threshold, and the interactive operations comprise click operations, forwarding operations and the like.

In the process of implementing the embodiment of the present application, an applicant finds that, for a unified modeling manner, a same model is used to serve all users, which cannot explicitly give consideration to data distribution differences between high-activity users and low-activity users, and when implementing the embodiment of the present application, the applicant finds that data distribution differences between the high-activity users and the low-activity users are large, for example, data of the high-activity users are fuller, a positive sample ratio is higher, a data loss rate and a noise rate of the low-activity users are larger, and a positive sample ratio is lower, but the low-activity users are a core user group of a recommendation system, and in some service scenarios, the recommendation system often needs to pay more attention to a recommendation effect for the low-activity users, for example: in the recommendation scenario, the high-activity users already have stickiness, so the low-activity users are the user group that the recommendation system needs to pay attention to, and if the recommendation system cannot serve the low-activity users well, the user stickiness and the user retention rate are negatively affected. Aiming at an independent modeling mode, two models are used for respectively fitting the data stream of the high-activity user and the data stream of the low-activity user, and although the mode can explicitly give consideration to the data difference distribution of the high-activity user and the low-activity user, more machine resource services are required to be consumed for online reasoning, the online engineering complexity is increased by times, and the online engineering is difficult to implement truly. Therefore, the applicant finds that, when implementing the embodiment of the present application, the data difference distribution of high-activity users and low-activity users cannot be taken into account explicitly in the related art without occupying more machine resources.

An embodiment of the present application provides an information recommendation method and apparatus based on artificial intelligence, an electronic device, and a computer-readable storage medium, which can explicitly consider data difference distribution of high-activity users and low-activity users without occupying more machine resources, so as to improve recommendation accuracy and recommendation efficiency of a recommendation system without increasing resource occupancy. In the following, an exemplary application when the electronic device is implemented as a server will be explained.

Referring to fig. 1, fig. 1 is a schematic structural diagram of an artificial intelligence-based information recommendation system provided in an embodiment of the present application, and is described by taking recommendation indexes as click through rates and predicted recommendation indexes as predicted click through rates as examples, in order to support a news application, a terminal 400 is connected to a server 200 through a network 300 (for example, the server 200 includes a training server 200-1 and an application server 200-2), and the network 300 may 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 pieces of information to be recommended, and determines the pieces of information to be recommended with the predicted click rate ranked in the front to return to the terminal 400 for presentation.

In some embodiments, when the information recommendation system is applied to a video recommendation scene, the terminal 400 receives a video to be uploaded, the terminal 400 sends the video to the application server 200-2, and the application server 200-2 determines information to be recommended, which is ranked in the front of the predicted click rate through a multi-target object prediction model, and sends the information to be recommended, which is ranked in the front of the predicted click rate, to the terminal 400, so that the terminal 400 directly presents the information to be recommended, which is ranked in the front of the predicted click rate, in a video recommendation home page.

In other embodiments, when the information recommendation method provided by the embodiment of the present application is implemented by a terminal alone, in the above-described various application scenarios, the terminal may operate a multi-target object prediction model to determine information to be recommended, which is ranked in the top by predicted click rate, and directly present the information to be recommended, which is ranked in the top by predicted click rate, at the terminal.

In some embodiments, the training server 200-1 and the application server 200-2 may be independent physical servers, may also be a server cluster or a distributed system formed by a plurality of physical servers, and may also be cloud servers that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, web services, cloud communication, middleware services, domain name services, security services, CDNs, and big data and artificial intelligence platforms. The terminal 400 may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, and the like. The terminal and the server may be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiment of the present invention.

Referring to fig. 2, fig. 2 is a schematic structural diagram of an electronic device provided in an embodiment of the present application, taking the electronic device as an application server 200-2 as an example, the application server 200-2 shown in fig. 2 includes: at least one processor 210, memory 250, at least one network interface 220. The various components in application server 200-2 are coupled together by bus system 240. It is understood that the bus system 240 is used to enable connected communication between these components. The bus system 240 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 240 in fig. 2.

The Processor 210 may be an integrated circuit chip having Signal processing capabilities, such as a general purpose Processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, etc., wherein the general purpose Processor may be a microprocessor or any conventional Processor, etc.

The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Memory 250 optionally includes one or more storage devices physically located remotely from processor 210.

The memory 250 includes volatile memory or nonvolatile memory, and may include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 250 described in embodiments of the invention is intended to comprise any suitable type of memory.

In some embodiments, memory 250 may be capable of storing data to support various operations, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplified below.

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

a network communication module 252 for communicating to other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), among others.

In some embodiments, the artificial intelligence based information recommendation device provided by the embodiments of the present invention may be implemented in software, and fig. 2 shows an artificial intelligence based information recommendation device 255 stored in a memory 250, which may be software in the form of programs and plug-ins, and includes the following software modules: an obtaining module 2551, a crossing module 2552, a determining module 2553, a predicting module 2554, a recommending module 2555, and a training module 2556, which are logical and therefore can be arbitrarily combined or further split according to the implemented functions, which will be described below.

The artificial intelligence based information recommendation method provided by the embodiment of the present application will be described in conjunction with an exemplary application and implementation of the application server 200-2 provided by the embodiment of the present application.

Referring to fig. 9, fig. 9 is a schematic structural diagram of a multi-target object prediction model provided in this embodiment, if the multi-target object prediction model is applied to a news recommendation system, it is assumed that there are two candidate recommendation index prediction networks matching different liveness degrees, for example, a candidate recommendation index prediction network a corresponding to a low-liveness user (a user with low liveness) and a candidate recommendation index prediction network B corresponding to a high-liveness user (a user with high liveness), the high-liveness user and the low-liveness user are distinguished according to behavior data of the users, the high-liveness user (a user with high liveness) and the low-liveness user (a user with low liveness) are distinguished according to behavior data of the users, the user with low liveness is a user whose number of interactive operations in unit time does not exceed a number threshold, the user with high liveness is a user whose number of interactive operations in unit time exceeds a number threshold, the interactive operation includes a click operation, a forwarding operation, and the like, one candidate recommendation index prediction network includes a prediction full-link layer a and a threshold network a, which are described below by taking a recommendation index as a click rate as an example and taking a prediction recommendation index as a prediction click rate, the recommendation index includes interactive evaluation parameters such as click rate, reading completion degree, and the like, the prediction full-link layer a outputs a predicted click rate a of a target object belonging to a low-activity user with respect to information to be recommended, another candidate recommendation index prediction network includes a prediction full-link layer B and a threshold network B, the prediction full-link layer B outputs a predicted click rate B of a target object belonging to a high-activity user with respect to information to be recommended, the two candidate recommendation index prediction networks and the candidate recommendation index prediction network share a plurality of expert networks (for example, expert network 0, expert network 1 and expert network 2), each expert network corresponds to different feature crossing modes, and the two candidate recommendation index prediction networks and the candidate recommendation index prediction network also share a feature input layer.

Referring to fig. 3A, fig. 3A is a schematic flowchart of an artificial intelligence-based information recommendation method provided in an embodiment of the present application, which will be described in conjunction with steps 101 to 105 shown in fig. 3A.

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

In some embodiments, referring to fig. 3B, fig. 3B is a schematic flowchart of an artificial intelligence based information recommendation method provided in an embodiment of the present application, and the obtaining of the multiple ranking features of the information to be recommended in step 101 may be implemented through steps 1011 to 1015.

In step 1011, ranking data of the information to be recommended is obtained.

By way of example, the type of sequencing data includes at least one of: the information sequence to which the behavior data of the target object aims, the type of the information to be recommended, the user portrait data of the target object, the behavior data of the target object, and the behavior data of a reference target object related to the target object.

In step 1012, when the type of the sorted data is the information sequence, the sorted data is packaged into the numerical characteristic of which the type is the variable-length discrete type characteristic.

As an example, when the type of the sorted data is an information sequence to be recommended, for example, an information sequence to be recommended clicked by a target object, an information sequence to be recommended that is exposed but not clicked, an information sequence to be recommended that is exposed and clicked, an information sequence to be recommended that is approved, an information sequence to be recommended that is reviewed, an information sequence to be recommended that is forwarded, an information sequence to be recommended that is collected, an information sequence to be recommended that is attended to, an information sequence to be recommended that is reported, a main sequence of public numbers of information to be recommended that is attended by the target object, and the like, the information sequences to be recommended are packaged as numerical features of variable-length discrete type features, for example, for the information sequences to be recommended (information 1 to be recommended, information 2 to be recommended, and information 3 to be recommended), the numerical features of vector form that are correspondingly packaged as corresponding to the information 1 to be recommended, the numerical features of vector form that are corresponding to the information 2 to be recommended, the information sequence to be recommended that is corresponding to the information 3, and the numerical features of vector form are composed of a plurality of numerical values.

In step 1013, when the type of the sorted data is numerical data, the sorted data is packaged as a numerical signature of which the type is a continuous signature.

As an example, when the type of the sort data is numerical data, for example, the age of the target object, the statistical value of the target object based on a plurality of categories of information to be recommended (exposure/number of clicks/click rate), the statistical value of the target object's current gender's crowd's information to be recommended (exposure/number of clicks/click rate), the statistical value of the target object's current age's crowd's information to be recommended (exposure/number of clicks/click rate), the statistical value of the information to be recommended (exposure/number of clicks/click rate/number of shares/number of relays/number of concerns/number of raises/average reading time/number of pictures, etc.), and the sort data is packaged as numerical features of which type is a continuous type feature, for example, "15" is correspondingly packaged 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 a plurality of numerical values.

In step 1014, when the type of the sorted data does not belong to the numerical data and does not belong to the information sequence, the sorted data is packaged into the numerical characteristic of which the type is the fixed-length discrete characteristic.

As an example, when the type of the sorted data does not belong to numerical data and does not belong to the information sequence to be recommended, for example, gender of the target object, or the like, information category label to be recommended (entertainment/science/sports, or the like), phone brand category of the target object, phone electricity amount of the target object (one/two/full, or the like), phone brightness of the target object (dark/general/bright, or the like), network category of the target object (4G/5G/Wi-Fi, or the like), current time period of the target object (early morning/midday/late, or the like), gender of the target object (male/female), age of the target object (old/middle year/young, or the like), activity level of the target object (low activity/middle activity/high activity), work category of the target object (programmer/driver/cleaner, or the like), current location of the target object (guangdong/north lake/north, or the like), interest category of the target object (basketball/reading/listening to the like), or the like, these data are packaged as numerical features in the form of a numerical vector, and the target object is a plurality of numerical vector corresponding gender vector.

In step 1015, a plurality of numerical features are embedded to obtain a plurality of ranking features.

As an example, a plurality of numerical features obtained through steps 1012 to 1014 belong to sparse features, and the numerical value corresponding to a position exceeding a set threshold in the sparse features (vectors) is zero, so in order to characterize all the numerical features, the dimension of each numerical feature is high, for example, the dimension of the sparse feature may reach 10000 dimensions, for better performing subsequent feature processing and semantic learning, the numerical features need to be packaged as ranking features participating in feature intersection processing, the ranking features belong to dense features relative to the sparse features, and the numerical value corresponding to a position exceeding the set threshold in the dense features (vectors) is not zero.

In some embodiments, the embedding processing on the plurality of numerical features to obtain the plurality of ranking features may be implemented by the following technical solutions: the following processing is performed for any one of the numerical features: taking hidden vectors respectively corresponding to a plurality of numerical values in the numerical characteristics as weights, and carrying out weighting processing on the plurality of numerical values to obtain an embedding dimension value of the numerical characteristics corresponding to one embedding dimension; and combining the embedding dimension value sets of the numerical characteristic corresponding to the multiple embedding dimensions into a sorting characteristic of the sorting data.

As an example, when the embedding dimension of the ranking feature is 5, the numerical feature may have a value of 1 at only one position, and for the position of the numerical feature, the multiple embedding dimension values mapped to the ranking feature are all 1, then in the process of obtaining the ranking feature from the input numerical feature, only one neuron in the sparse feature layer is active, and the values corresponding to five lines connected to the active neuron are v _m1 、v _m2 、v _m3 、v _m4 、v _m5 The combination of these five values is the ordering attribute e _m For any one embedding dimension of the sorted features, the value of the embedding dimension is obtained based on numerical values of a plurality of positions of the numerical feature, each numerical value contributes to the embedding dimension, a hidden vector corresponding to the numerical values of the numerical feature is taken as a weight, and if the field length of the numerical feature is 3, three discrete values exist, and the corresponding hidden vector is, for example, v _1n 、v _2n 、v _3n Then the embedding dimension value corresponding to the nth embedding dimension is the first bitWeighting processing results of the set numerical value, the numerical value at the second position and the numerical value at the third position aiming at 5 embedding dimensions to obtain the sorting characteristic e _m 。

In step 102, a plurality of sorting features are subjected to feature intersection processing based on a plurality of intersection ways, so as to obtain a plurality of feature vectors corresponding to the plurality of intersection ways one by one.

In some embodiments, referring to fig. 3C, fig. 3C is a schematic flowchart of an artificial intelligence-based information recommendation method provided in this embodiment, and the step 102 of performing feature interleaving processing based on multiple interleaving manners on multiple sorting features may be implemented by executing multiple (i.e., at least two) steps of steps 1021 to 1023, where it needs to be pointed out that steps 1021 to 1023 are parallel, and there is no restriction on the execution order.

In step 1021, a second order feature intersection process is performed on the plurality of ranked features.

In some embodiments, the above-mentioned performing second-order feature intersection processing on a plurality of ranking features may be implemented by the following technical solutions: combining a plurality of sorting features of the information to be recommended for a plurality of times to obtain a plurality of combined features of the information to be recommended; wherein the ranking features used for each combined treatment are partially or completely different; weighting and summing each combined feature of the information to be recommended by taking the recommendation index influence factor of each combined feature of the information to be recommended as a weight parameter; and 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.

As an example, the ranking features used are partially the same for each combination process, meaning that the ranking feature x is ₁ Can sort the features x ₂ Combined to form a combined feature x ₁₂ Ordering feature x ₁ Can sort the features x ₃ Combined to form combined features x ₁₃ The ranking features used in the two combinations involved are partly different, but are each identical to the ranking feature x ₁ The combination characteristics are formed, and the completely different ranking characteristics used in each combination processing means that the ranking characteristicsx ₁ Can sort the features x ₂ Combined to form combined features x ₁₂ Ordering feature x ₃ Can sort the features x ₄ Combined to form a combined feature x ₃₄ The ordering characteristics used in the two combinations involved are quite different.

As an example, the second-order feature intersection processing uses a factorization model, any two features are combined in pairs, the combined features can be regarded as a new feature, and the weight of the combined features is obtained through learning in a training phase, which is shown in formula (1):

wherein the value of v is determined during the training process, the component v _i The dot product of (a) is the fusion parameter of the two features, v _i I.e. the above mentioned recommendation index association influence factor, x _i I.e., the ordering attribute, n is the number of ordering attributes.

In step 1022, explicit depth feature intersection processing is performed on the plurality of ranking features.

In some embodiments, performing explicit depth feature interleaving on the plurality of ranking features in step 1022 may be implemented by the following technical solutions: performing explicit feature cross processing on the kth level to obtain an explicit feature vector of the kth level; pooling the explicit feature vectors of the k level to obtain a pooling result of the k level; splicing the K levels of pooling treatment results; the method comprises the steps that K is an integer larger than or equal to 2, K is an integer variable with the value increasing from 1, the value range of K is that K is larger than or equal to 1 and is smaller than K, when the value of K is 1, the input of the K-th level explicit feature cross processing is a plurality of sorting features, and when the value of K is larger than or equal to 2 and is smaller than K, the input of the K-th level explicit feature cross processing is an explicit feature vector of a K-1 level.

In some embodiments, the performing of the explicit feature interleaving processing of the kth level may be implemented by the following technical solutions: when the value of k is 1, multiplying the plurality of sorting features by the plurality of sorting features element by element; and when K is more than or equal to 2 and less than K, performing element-by-element multiplication on the plurality of ordering characteristics and the explicit characteristic vector of the K-1 level.

For example, referring to fig. 7 and 8, fig. 7 and 8 are schematic processing diagrams of a compressed cross expert network based on an artificial intelligence information recommendation method provided in an embodiment of the present application, and in order to implement automatic learning of explicit high-order feature interaction and enable interaction to occur at a vector level, display feature interaction is performed according to the calculation principle shown in fig. 7 and 8, referring to fig. 8, an input ranking feature and a feature output by a hidden layer of the compressed cross expert network respectively form a matrix, which is denoted as X ₀ (dimension D, m ordering characteristics) and X _k Referring to fig. 7, neurons in each hidden layer in the compressed cross expert network are derived from the hidden layer of the previous layer and the original input ranking features, and the hidden layer of the k-th layer contains H _k The calculation of the hidden layer of each neuron vector can be divided into two steps: (1) According to the state X of the previous hidden layer _k And a matrix X of ordering characteristics of the original input ₀ Calculating an intermediate result, wherein the intermediate result is a three-dimensional tensor; (2) On the basis of this intermediate result, H is used _k+1 Size of m x H _k The convolution kernel of the compressed cross expert network generates the state of the next hidden layer, the operation is generally consistent with the convolution neural network in computer vision, the only difference is the design of the convolution kernel, the acceptance domain related to one neuron in the compressed cross expert network is the whole plane perpendicular to the characteristic dimension D, the acceptance domain is the local small range area around the current neuron, therefore, the characteristic diagram obtained by the convolution operation in the compressed cross expert network is a vector instead of a matrix, see fig. 7, the finally learned order of the characteristic interaction is determined by the number of the network layers, each hidden layer is connected to the output unit of the expert network through a pooling operation, thereby ensuring that the output unit can obtain the characteristic interaction modes with different orders, the structure of the compressed cross expert network is similar to the convolution neural network, namely, the state of each layer is calculated by the value of the previous hidden layer and an additional input data, but the compressed cross expert network is obtained by the calculation of the value of the previous hidden layer and the additional input dataThe parameters of different layers in the expert network are different, and the additional input data in the compressed cross expert network is fixed and always is the matrix X of the sequencing characteristics of the original input ₀ High-order semantic features can be learned through high-order explicit crossing, accuracy of subsequent click rate prediction is improved, and the high-order explicit crossing is interpretable.

In step 1023, an implicit depth feature interleaving process is performed on the plurality of ranked features.

In some embodiments, the implicit depth feature interleaving processing on the plurality of sorted features in step 1023 can be implemented by the following technical solutions: carrying out N times of implicit feature cross processing on the plurality of sequencing features; the method comprises the following steps that N is an integer larger than or equal to 2, N is an integer variable with the value increasing from 1, the value range of N is that N is larger than or equal to 1 and is smaller than N, when the value of N is 1, the input of the nth implicit feature cross processing is a plurality of sorting features, when the value of N is larger than or equal to 2 and is smaller than N, the input of the nth implicit feature cross processing is an N-1-th implicit feature vector, and when the value of N is N-1, the output of the (N + 1) -th implicit feature cross processing is a feature vector.

In some embodiments, the above-mentioned performing implicit feature intersection processing on a plurality of ranking features for N times may be implemented by the following technical solutions: the following processing is performed in the process of each implicit feature intersection processing: performing full connection processing on the input of the nth implicit characteristic cross processing to obtain a full connection processing result corresponding to the nth implicit characteristic cross processing; and activating the full-connection processing result to obtain the nth implicit characteristic vector.

As an example, each implicit feature crossing process is implemented by a plurality of neural networks, each neural network includes a full connection layer and an activation layer, the plurality of ordered features are propagated forward in N cascaded neural networks to obtain feature vectors obtained through the feature crossing process, in the process of processing each neural network, two operations need to be performed, the full connection process is performed through the full connection layer, then the activation process is performed through the activation layer, a random deletion process is performed in the training stage of the neural network, that is, part of neurons are discarded at random.

In step 103, a target recommendation index prediction network matching the behavior data of the target object is determined among the candidate recommendation index prediction networks respectively corresponding to the plurality of different liveness degrees.

In some embodiments, in step 103, determining a target recommendation index prediction network matching with the behavior data of the target object in the candidate recommendation index prediction networks respectively corresponding to a plurality of different liveness degrees may be implemented by the following technical solutions: determining the activity of the target object based on the behavior data of the target object; and determining the candidate recommendation index prediction network corresponding to the activity of the target object as the target recommendation index prediction network matched with the behavior data of the target object in the candidate recommendation index prediction networks respectively corresponding to the different activities.

In some embodiments, the determining the liveness of the target object in step 103 based on the behavior data of the target object may be implemented by the following technical solutions: acquiring the number of times of clicking operation of the target object in unit time and the online time of the target object in unit time from the behavior data of the target object; and determining the activeness positively correlated to the number of the clicking operations and negatively correlated to the online time.

As an example, the behavior data of the target object records a timestamp of the click operation of the target object, the statistical timestamp belongs to the number of times of the click operation within a certain unit time, the liveness of the target object is in positive correlation with the number of times, the larger the number of times represents that the click operation is initiated by the target object within the unit time more frequently, and the online time of the target object within the unit time is counted, and the online time has a reduction effect on the liveness, for example, the online time of the target object within one month is as long as 600 hours, although the number of times of the click operation within one month is 600, only one click operation exists per hour on average, the liveness representing the target object is lower, but the online time of the target object within one month is only 10 hours, although the number of times of the click operation within one month is 300, only thirty click operations exist per hour on average, the liveness representing the target object is higher, the high or low of the liveness of the target object can be distinguished by an interval, for example, the interval to which the ratio of the number of the click operation to the online time within the unit time of the target object within the unit time is determined, and the second interval, for example, the liveness exists, and the second interval is determined, and the third interval. When the ratio falls in a first interval, representing that the target object belongs to a user with low liveness; when the ratio falls in a second interval, representing that the target object belongs to a normal activity user; when the ratio falls in the third interval, the target object is represented to belong to a user with high activity, each activity corresponds to a respective candidate recommendation index prediction network, namely three candidate recommendation index prediction networks are respectively in one-to-one correspondence with low activity, normal activity and high activity, and the activity represented by the behavior data of the target object is directly associated with different candidate recommendation index prediction networks, so that the prediction click rate of the target object for the information to be recommended can be predicted more accurately.

As an example, the click operation may be interpreted narrowly as a click operation for a certain information to be recommended, or may be interpreted broadly as an active operation including a like operation, a forward operation, a comment operation, and the like.

In step 104, click rate prediction processing is performed on the plurality of feature vectors through a target recommendation index prediction network to obtain a predicted click rate of the target object corresponding to the information to be recommended.

In some embodiments, referring to fig. 3D, fig. 3D is a schematic flowchart of an information recommendation method based on artificial intelligence provided in the embodiment of the present application, and in step 104, click rate prediction processing is performed on a plurality of feature vectors through a target recommendation index prediction network to obtain a predicted click rate of a target object corresponding to information to be recommended, which may be implemented by executing steps 1041 to 1042.

In step 1041, performing weighting processing on the multiple feature vectors through a target recommendation index prediction network to obtain semantic features of the target object corresponding to the information to be recommended;

in step 1042, the semantic features are mapped to the predicted click rate of the target object corresponding to the information to be recommended by the target recommendation index prediction network.

As an example, the target recommendation index prediction network comprises a threshold network and a corresponding prediction full-link layer, weighting and summing processing is performed on a plurality of feature vectors output based on a plurality of feature cross processing modes through weight parameters configured by the threshold network to obtain semantic features, and then the semantic features are mapped to the predicted click rate of the target object corresponding to the information to be recommended through the prediction full-link layer.

In some embodiments, in step 104, click rate prediction processing is performed on the plurality of feature vectors through a target recommendation index prediction network to obtain a predicted click rate of the target object corresponding to the information to be recommended, which may be implemented by the following technical solutions: acquiring a plurality of reference target objects of which the portrait similarity of a user portrait with the target object is not less than a portrait similarity threshold; determining reference behavior data of each reference target object, and determining a reference recommendation index prediction network matched with the reference behavior data in a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees; carrying out click rate prediction processing on the plurality of characteristic vectors through a reference recommendation index prediction network to obtain a reference prediction click rate of the target object corresponding to the information to be recommended; carrying out click rate prediction processing on the plurality of characteristic vectors through a target recommendation index prediction network to obtain the predicted click rate of the target object corresponding to the information to be recommended; and correcting the predicted click rate of the target object corresponding to the information to be recommended by referring to the predicted click rate and the corresponding portrait similarity to obtain a new predicted click rate of the target object corresponding to the information to be recommended.

As an example, on the basis of steps 1041-1042, a reference predicted click rate of a reference target object for information to be recommended may be predicted, the predicted click rate in step 1042 is modified by the reference predicted click rate, a plurality of reference target objects with a portrait similarity of a user of the target object not less than a portrait similarity threshold are first obtained, so that the obtained reference target objects are users with similar interests to the target object, reference behavior data of each reference target object is determined, and a reference recommendation index prediction network matching the reference behavior data is determined in candidate recommendation index prediction networks respectively corresponding to a plurality of different liveness degrees, although the reference target object and the target object have similar interests, the liveness represented by the behavior data of the reference target object may be different from the liveness represented by the behavior data of the target object, the click rate prediction processing is carried out on a plurality of characteristic vectors through a reference recommendation index prediction network to obtain a reference prediction click rate of the information to be recommended corresponding to the target object, the prediction click rate of the information to be recommended corresponding to the target object is corrected through the reference prediction click rate and the corresponding portrait similarity to obtain a new prediction click rate of the information to be recommended corresponding to the target object, for example, the reference prediction click rate obtained for a certain reference target object is 0.8, the portrait similarity of a user portrait of the reference target object and a user portrait of the target object is 0.9, the prediction click rate of the information to be recommended corresponding to the target object 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 are averaged by 0.5 to obtain a new prediction click rate of the information to be recommended corresponding to the target object, the prediction accuracy can be effectively improved and the interpretability is provided by correcting the prediction result of the target object by referring to the prediction result of the target object.

In step 105, a recommendation operation for the target object is executed based on the predicted click rate of different information to be recommended.

As an example, the recommendation order of the multiple pieces of information to be recommended is determined according to the predicted click rate of the information to be recommended, the recommendation operation corresponding to the object to be recommended is executed based on the recommendation order of the multiple pieces of information to be recommended, that is, the information to be recommended is sorted in a descending order according to the predicted click rate of the information to be recommended, at least one piece of information to be recommended sorted in the top is selected from the result of the sorting in the descending order, and the recommendation operation corresponding to the object to be recommended is executed based on the selected information to be recommended.

As an example, a plurality of pieces of information to be recommended which are sorted in the descending order are selected as information to be recommended to an object to be recommended, and a recommendation operation corresponding to the object to be recommended is executed based on the selected information to be recommended, where the recommendation operation may be directly pushed to the object to be recommended, or may be reordered based on the selected information to be recommended, where reordering refers to reordering of the selected information to be recommended in a manner of aggregation processing of different multi-recommendation tasks, so as to obtain recommendation information liked by the object to be recommended from a more comprehensive perspective.

In some embodiments, the obtaining of the plurality of ranking features of the information to be recommended in step 101 is realized by a feature extraction network, the performing of feature intersection processing on the plurality of ranking features based on a plurality of intersection modes in step 102 is realized by an expert network, and the feature extraction network, the expert network and a plurality of candidate recommendation index prediction networks form a multi-target object prediction model; before obtaining a plurality of sorting features of information to be recommended in step 101, obtaining a plurality of target object samples corresponding to the information samples to be recommended, wherein behavior data of the plurality of target object samples are matched with a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees; associating each target object sample with information to be recommended to form a training sample corresponding to the target object sample; carrying out forward propagation on the training sample corresponding to each target object sample in a feature extraction network and a plurality of expert networks to obtain a feature sample of each training sample; carrying out forward propagation on the characteristic sample of each training sample in a candidate recommendation index prediction network corresponding to a target object sample of the training samples to obtain a training prediction click rate corresponding to each training sample; determining the error between the training predicted click rate of each training sample and the corresponding pre-marked click rate; and reversely propagating the error in the multi-target object prediction model to determine a parameter change value of the multi-target object prediction model when the error obtains a minimum value, and updating the parameter of the multi-target object prediction model based on the parameter change value.

As an example, before online service of a multi-target object prediction model, a large number of training samples are required to be used for fully training the multi-target object prediction model, so that the data distribution of the fitting samples of the multi-target object prediction model can be realized, trainable parameters of the multi-target object prediction model can be fully learned, and historical click logs of all target object samples are extracted through the embedded point service of an online recommendation system; the method comprises the steps of splicing an information sample to be recommended in a historical click label and a target object sample corresponding to the historical click label into an available training sample, obtaining a target object sample with an exposure label but no historical click label aiming at the information sample to be recommended, splicing the target object sample with the information sample to be recommended into an available training sample, wherein the historical click label is an effective click label aiming at click operation of entering an information detail page for more than 5 seconds, the target object sample without the historical click label represents that the target object sample browses the information sample to be recommended but does not click the information sample to be recommended, after obtaining a large number of training samples, performing a series of preprocessing such as null value filling, abnormal sample removing and outlier characteristic value correction on the training samples, dividing a training set and a verification set after finishing processing, performing forward propagation and reverse parameter updating on a multi-target object prediction model based on the divided training set, stopping training when training indexes on the verification set are not lifted, and outputting the target prediction model with fixed parameters for use of an online recommendation service.

As an example, the following training samples are targeted: if the target object sample of the training sample has a click operation for the information sample to be recommended, calculating an error by using the following formula (2):

L＝-y _i logp _i (2)；

wherein L is an error, p _i Is a prediction of trainingClick rate, y _i Is the pre-mark click rate (which typically takes a value of 1 in the presence of a click operation).

As an example, the following training samples are targeted: if the target object sample of the training sample has no click operation for the information sample to be recommended, but the target object sample browses the information to be recommended, calculating an error by using the following formula (3):

L＝-(1-y _i )log(1-p _i ) (3)；

wherein L is an error, p _i Is to train the predicted click rate, y _i Is the pre-mark click rate (which typically takes a value of 0 in the presence of a click operation).

Next, an exemplary application of the embodiment of the present application in a practical application scenario will be described.

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 predicted click rates of the user for a plurality of pieces of information to be recommended, and determines pieces of information to be recommended, which are ranked in the front of the predicted click rates, to return to the terminal for presentation.

Referring to fig. 4A to 4C, fig. 4A to 4C are schematic product representation diagrams of an artificial intelligence based information recommendation method according to an embodiment of the present disclosure, where a message page 501A of a social client is presented in fig. 4A, and a referrer page 501C of a news product in fig. 4C is switched to in response to a click operation on a news product dialog box 502A in the message page 501A, and a public number page 501B of a social client is presented in fig. 4B, and a referrer page 501C of a news product in fig. 4C is switched to in response to a click operation on an entry control 502B of a public number in the public number page 501B, where the referrer page 501C includes various information referred to a user by a personalized recommendation algorithm.

In some embodiments, referring to fig. 5, fig. 5 is a schematic structural diagram of a multi-target object prediction model provided in an embodiment of the present application, where the multi-target object prediction model includes a feature input layer, a semantic layer, and an output layer, the feature input layer may be practiced as a feature extraction network, the semantic layer may be practiced as an expert network, and the threshold network and the output layer may be practiced as a candidate recommendation indicator prediction network.

In some embodiments, the input layer comprises three components, namely, a variable-length discrete feature, a fixed-length discrete feature, and a continuous feature, each of which is described in detail below: 1. variable-length discrete features refer to discrete features of variable length, often sequence features, such as: the information sequence clicked by the user, and the like, the variable-length discrete type characteristics mainly comprise: the information sequence characteristics of the information release public number, such as the exposed and un-clicked information sequence characteristics, the exposed and clicked information sequence characteristics, the approved information sequence characteristics, the commented information sequence characteristics, the forwarded information sequence characteristics, the collected information sequence characteristics, the concerned information sequence characteristics, the reported information sequence characteristics, the main sequence characteristics of the information release public number concerned by the user and the like; 2. the fixed-length discrete feature refers to a discrete feature with a fixed length, and the value of the fixed-length discrete feature is a discrete value, does not have actual mathematical meaning and is not used for representing a sequence, for example: the sex of the user, etc., the fixed-length discrete type characteristics mainly include: information category labels (entertainment/science and technology/sports, etc.), user mobile phone brand categories, user mobile phone electricity (one/two/full, etc.), user mobile phone brightness (dark/general/bright, etc.), user network categories (4G/5G/Wi-Fi, etc.), user current time periods (early morning/noon/evening, etc.), user gender (male/female), user age period (old/middle year/young, etc.), user activity level (low activity/middle activity/high activity), user work categories (programmer/driver/cleaner, etc.), user current location (Guangdong/Hubei/Beijing, etc.), user interest categories (basketball/reading/listening to songs, etc.), etc.; 3. the continuous type feature refers to a feature that a value is a continuous value and has an actual mathematical meaning, for example: age of the user, etc., the continuous type features mainly include: the user is based on the statistics of a plurality of information categories (exposure/click rate), the statistics of the information of the crowd of the current gender of the user (exposure/click rate), the statistics of the information of the crowd of the current age of the user (exposure/click rate), the statistics of the information (exposure/click rate/sharing/forwarding/attention/reporting/average reading time/number of pictures, etc.), etc.

In some embodiments, to enable features of the input layer to intersect with a richer semantic implication representation, and provides the following output layers corresponding to different target objects with the implicit semantic features capable of covering different crowds with different activities, the structural design optimization is carried out on the multi-target object prediction model in the semantic layer, so that the multi-target object prediction model can more fully mine and collect the data distribution difference of high and low living users, the vector representation of the variable-length discrete type characteristic, the fixed-length discrete type characteristic and the continuous type characteristic of the input layer can be input to a plurality of expert networks (Experts) together, in order to improve the feature crossing capability of the expert Network, a double-crossing expert Network (Bi-Interaction), a Compressed crossing expert Network (CIN) and a Deep crossing expert Network (Deep Module) are adopted, the double-crossing expert Network carries out factorization type second-order crossing on vector representation, the Compressed crossing expert Network carries out explicit high-order crossing on the vector representation, the Deep crossing expert Network carries out implicit high-order crossing on the vector representation, or modes such as a deformable convolution crossing expert Network and an automatic crossing expert Network can be adopted, wherein, the implicit high-order crossing comprises a plurality of hidden layer processing, each hidden layer processing comprises full connection processing and activation function processing, the training process also includes the process of deleting the neurons randomly, after each expert network executes the cross operation, the feature vector corresponding to each expert network is obtained, weighting the feature vectors corresponding to each expert network through the threshold networks corresponding to different target users, furthermore, different semantic features for different target users can be obtained, and the weighting processing can be carried out by using a forward attention processing mode, a self-attention processing mode and the like.

In some embodiments, in order to implement automatic learning of explicit high-order feature interaction and enable interaction to occur at a vector level, display feature interaction is performed according to the calculation principle shown in fig. 7 and 8, referring to fig. 8, a hidden vector is a unit object, so that an input ranking feature and a feature output by a hidden layer of a compressed cross expert network form a matrix, which is denoted as X ₀ And X _k The neuron of each hidden layer in the compressed cross expert network is obtained by calculation according to the hidden layer of the previous layer and the sequencing feature of the original input, and the hidden layer of the k layer contains H _k The calculation of the individual neuron vectors and hidden layers can be divided into two steps: (1) According to the state X of the previous hidden layer _k And a matrix X of ordering characteristics of the original input ₀ Calculating an intermediate result, wherein the intermediate result is a three-dimensional tensor; (2) On the basis of this intermediate result, H is used _k+1 Size of m x H _k The convolution kernel generates the state of the next hidden layer, the operation is generally consistent with the convolution neural network in computer vision, the only difference is the design of the convolution kernel, the acceptance domain related to one neuron in the compression cross expert network is the whole plane perpendicular to the characteristic dimension D, the acceptance domain is the local small-range area around the current neuron, therefore the characteristic diagram obtained by the convolution operation in the compression cross expert network is a vector instead of a matrix, see FIG. 7, the order of the finally learned characteristic interaction is determined by the number of the network layers, each hidden layer is connected to the output unit of the expert network through a pooling operation, thereby ensuring that the output unit can obtain the characteristic interaction modes with different orders, the structure of the compression cross expert network is similar to that of the convolution neural network, namely, the state of each layer is calculated by the value of the previous hidden layer and an additional input data, but 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, and the original input characteristic matrix is always the X of the ordering characteristic matrix ₀ 。

In some embodiments, the multi-target object prediction model provided in the embodiment of the present application finally needs to predict a high live click rate or a low live click rate of a user, the output layer is composed of two task modules, which are a high live task module and a low live task module respectively, if the target object is a high live target user, the semantic layer outputs semantic features corresponding to the high live target user to input the semantic features to the high live task module, the high live task module performs mapping conversion on the semantic features through a full connection layer corresponding to the high live target object, and finally determines the predicted click rate of the target object for the information to be recommended through an activation function, if the target object is a high live target user, the semantic layer outputs semantic features corresponding to the low live target user to input the semantic features to the low live task module, the low live task module performs mapping conversion through the full connection layer corresponding to the low live target object, and finally determines the predicted click rate of the target object for the information to be recommended through the activation function.

In some embodiments, referring to fig. 6, fig. 6 is a schematic flowchart of a training phase of an artificial intelligence based information recommendation method provided in an embodiment of the present application, before the multi-target object prediction model is served online, a large number of training samples are required to be used to fully train the multi-target object prediction model, so that the multi-target object prediction model fits a data distribution of the samples, 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 5 phases: in the first stage (historical clicking of a user), extracting a historical clicking log of a user sample through a buried point service of an online recommendation system; in the second stage (constructing samples), extracting sequencing characteristics through a characteristic reporting system, wherein the sequencing characteristics comprise variable-length discrete characteristics, fixed-length discrete characteristics and continuous characteristics, splicing user samples corresponding to characteristic data and historical click labels into usable training samples, the historical click labels are effective click labels, the effective click labels aim at click operations of entering information detail pages for more than 5 seconds, in the third stage (dividing a training set verification set), after a large number of training samples are obtained, a series of preprocessing steps such as null value filling, abnormal sample removing and outlier characteristic value correcting are required to be carried out on the training samples, after the preprocessing steps are finished, the training set and the verification set are divided, in the fourth stage (training model), full training is carried out on the multi-target object prediction model based on the divided training set and verification set, an early-stop strategy is adopted to train the multi-target object prediction model, namely, training is stopped when training indexes on the verification set are not improved any more, and in the fifth stage (output model), the multi-target object prediction model with fixed parameters is output to be used for on-line multi-target service.

In some embodiments, the training phase of the artificial intelligence-based information recommendation method provided in the embodiment of the present application uses a cross-entropy piecewise loss function as a whole loss function for training, where the whole loss function is shown in formula (4):

wherein L is _total Is a loss of the whole body, and the loss of the whole body,

is the high loss of activity for the ith training sample,

is the low activity loss of the ith sample, and N is the total number of training samples.

In some embodiments, the high loss of activity for the ith training sample is seen in equation (5):

wherein, the first and the second end of the pipe are connected with each other,

is the high loss of activity, p, of the ith training sample _i Is the predicted click rate probability, y _i Is the actual index value (usually 1), user _i Users of the ith training sample.

In some embodiments, the low loss of activity for the ith training sample is seen in equation (6):

wherein the content of the first and second substances,

is the low activity loss, p, of the ith training sample _i Is the predicted click rate probability, y _i Is the actual index value (usually 1), user _i Users of the ith training sample.

In some embodiments, the above-mentioned loss is interpreted as an error, for each training sample, there is a corresponding error, and the calculation manner of the error is a cross entropy loss function shown in formulas (5) and (6) no matter the training sample of the low-activity user or the training sample of the high-activity user, and an AdamW optimization algorithm is used to solve the optimization problem, and integrates an adaptive learning rate and a momentum term, and a weight attenuation strategy is adopted, so that the multi-target object prediction model requires fewer resources, converges faster, and can find a local optimal solution to the optimization problem with higher efficiency.

In some embodiments, the multi-target object prediction model provided by the embodiment of the application can better fit the data distribution difference of high and low living user groups, the training parameters can be better learned, and after the training parameters are fixed, the multi-target object prediction model can be derived and then applied to online recommendation services.

In some embodiments, the application phase includes the following steps: 1. when the application server judges that the current moment is the recommendation opportunity, constructing a long discrete type feature, a fixed length discrete type feature and a continuous type feature based on the user data and the data of the information to be recommended so as to generate a tested feature sample; 2. performing feature cross processing on the tested feature samples through a semantic layer, and performing index prediction by using a task module matched with the activity of the user in a multi-target object prediction model according to the activity of the user to obtain the predicted click rate probability of all information to be recommended; 3. and sequencing all the information to be recommended according to the predicted click rate probability, and recommending the information to be recommended with the highest score.

The embodiment of the application provides an information recommendation method based on artificial intelligence, which is applied to a sequencing model of a recommendation system, and is used for learning the data distribution problem of high-activity users and low-activity users, so that the reading conversion effect of user groups with different activity degrees in the recommendation system is improved, the data distribution of the high-activity users and the data distribution of the low-activity users are fitted through candidate recommendation index prediction networks of the high-activity users and the low-activity users respectively, the conversion rates of the high-activity users and the low-activity users are modeled fully, the performance of the sequencing model in the high-activity users and the low-activity users is improved, and key indexes such as the playing number, the browsing amount, the daily active user number and the conversion rate of the recommendation system are improved.

In some embodiments, compared with modeling manners in the related art, no matter the modeling manner is unified modeling or independent modeling, the multi-target object prediction model provided by the embodiment of the application explicitly considers the data difference distribution of high-activity and low-activity user groups under the condition of not occupying more machine resources, through the novel neural network structure design, the multi-target object prediction model can skillfully mine the data distribution difference of the high-activity data and the low-activity data, and through a double-cross expert network, a compression cross expert network and a deep cross expert network, the input ranking characteristics are fully and cross mined, the accuracy and the efficiency of online recommendation are greatly improved, and the user experience is improved.

Continuing with the exemplary structure of the artificial intelligence based information recommendation device 255 provided by the embodiment of the present application as a software module, in some embodiments, as shown in fig. 2, the software module stored in the artificial intelligence based information recommendation device 255 of the memory 250 may include an obtaining module 2551 for obtaining a plurality of ranking features of the information to be recommended; a cross module 2552, configured to perform feature cross processing based on multiple cross manners on the multiple sorting features, to obtain multiple feature vectors corresponding to the multiple cross manners one to one; a determining module 2553, configured to determine, in the candidate recommendation indicator prediction networks respectively corresponding to the multiple different liveness degrees, a target recommendation indicator prediction network that matches the behavior data of the target object; the prediction module 2554 is configured to perform recommendation index prediction processing on the plurality of feature vectors through a target recommendation index prediction network to obtain a prediction recommendation index of the target object corresponding to the information to be recommended; and the recommending module 2555 is used for executing recommending operation aiming at the target object based on the prediction recommending indexes of different information to be recommended.

In some embodiments, the obtaining module 2551 is further configured to: acquiring sequencing data of information to be recommended; when the type of the sequencing data is an information sequence, packaging the sequencing data into a numerical characteristic with a variable-length discrete characteristic; when the type of the sequencing data is numerical data, packaging the sequencing data into numerical characteristics with the type of continuous characteristics; when the type of the sequencing data does not belong to numerical data and does not belong to an information sequence, packaging the sequencing data into numerical characteristics of which the type is a fixed-length discrete characteristic; and embedding the numerical characteristics to obtain a plurality of sequencing characteristics.

In some embodiments, the obtaining module 2551 is further configured to: the following processing is performed for any one of the numerical features: weighting a plurality of numerical values by taking hidden vectors respectively corresponding to the plurality of numerical values in the numerical characteristics as weights to obtain an embedding dimension value of the numerical characteristics corresponding to one embedding dimension; and combining the embedding dimension value values of the numerical characteristic corresponding to a plurality of embedding dimensions into a sorting characteristic of the sorting data.

In some embodiments, crossover module 2552, is further configured to: various processes among the following processes are performed: performing second-order feature cross processing on the plurality of sequencing features; performing explicit depth feature intersection processing on the plurality of ordering features; and carrying out implicit depth feature intersection processing on the plurality of ordering features.

In some embodiments, crossover module 2552, is further configured to: combining the plurality of sorting features of the information to be recommended for a plurality of times to obtain a plurality of combined features of the information to be recommended; wherein the ranking features used for each combined treatment are partially or completely different; taking the recommendation index influence factor of each combined feature of the information to be recommended as a weight parameter, and carrying out weighted summation processing on each combined feature of the information to be recommended; and 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.

In some embodiments, crossover module 2552 is further configured to: performing explicit feature cross processing of a kth level to obtain an explicit feature vector of the kth level; pooling the explicit feature vectors of the kth level to obtain a pooling result of the kth level; splicing the K levels of pooling processing results; the method comprises the steps that K is an integer larger than or equal to 2, K is an integer variable with values increasing from 1, the value range of K is that K is larger than or equal to 1 and is smaller than K, when the value of K is 1, the input of the K-th level explicit feature cross processing is a plurality of ordering features, and when the value of K is larger than or equal to 2 and is smaller than K, the input of the K-th level explicit feature cross processing is an explicit feature vector of a K-1 level.

In some embodiments, crossover module 2552, is further configured to: when k is 1, performing element-by-element multiplication processing on the plurality of sorting features and the plurality of sorting features; and when K is more than or equal to 2 and less than K, performing element-by-element multiplication on the plurality of sorting features and the explicit feature vector of the K-1 level.

In some embodiments, crossover module 2552 is further configured to: carrying out N times of implicit feature cross processing on the plurality of sequencing features; the method comprises the following steps of performing implicit feature cross processing on a plurality of input features, wherein N is an integer greater than or equal to 2, N is an integer variable with the value increasing from 1, the value range of N is that N is greater than or equal to 1 and is less than or equal to N, when the value of N is 1, the input of the nth implicit feature cross processing is a plurality of sorting features, when the value of N is greater than or equal to 2 and is less than or equal to N, the input of the nth implicit feature cross processing is an N-1-th implicit feature vector, and when the value of N is N-1, the output of the N + 1-th implicit feature cross processing is a feature vector.

In some embodiments, crossover module 2552, is further configured to: the following processing is performed in the process of each implicit feature intersection processing: performing full connection processing on the input of the nth implicit characteristic cross processing to obtain a full connection processing result corresponding to the nth implicit characteristic cross processing; and activating the full-connection processing result to obtain the nth implicit characteristic vector.

In some embodiments, the prediction module 2554 is further configured to: weighting the plurality of feature vectors through a target recommendation index prediction network to obtain semantic features of the target object corresponding to the information to be recommended; and mapping the semantic features into a prediction recommendation index of the target object corresponding to the information to be recommended through a target recommendation index prediction network.

In some embodiments, the prediction module 2554 is further configured to: acquiring a plurality of reference target objects of which the portrait similarity of a user portrait with the target object is not less than a portrait similarity threshold; determining reference behavior data of each reference target object, and determining a reference recommendation index prediction network matched with the reference behavior data in a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees; performing recommendation index prediction processing on the plurality of feature vectors through a reference recommendation index prediction network to obtain a reference prediction recommendation index of the target object corresponding to the information to be recommended; performing recommendation index prediction processing on the plurality of characteristic vectors through a target recommendation index prediction network to obtain a prediction recommendation index of the target object corresponding to the information to be recommended; and correcting the prediction recommendation index of the information to be recommended corresponding to the target object by referring to the prediction recommendation index and the corresponding portrait similarity to obtain a new prediction recommendation index of the information to be recommended corresponding to the target object.

In some embodiments, the obtaining of the plurality of ranking features of the information to be recommended is realized through a feature extraction network, the feature cross processing of the plurality of ranking features based on a plurality of cross modes is realized through an expert network, and the feature extraction network, the expert network and a plurality of candidate recommendation index prediction networks form a multi-target object prediction model; the apparatus further includes a training module 2556 for: before obtaining a plurality of sorting features of information to be recommended, obtaining a plurality of target object samples corresponding to the information samples to be recommended, wherein behavior data of the plurality of target object samples are matched with candidate recommendation index prediction networks respectively corresponding to a plurality of different activeness; associating each target object sample with information to be recommended as a training sample corresponding to the target object sample; carrying out forward propagation on the training sample corresponding to each target object sample in a feature extraction network and a plurality of expert networks to obtain a feature sample of each training sample; carrying out forward propagation on the characteristic sample of each training sample in a candidate recommendation index prediction network corresponding to a target object sample of the training sample to obtain a training prediction recommendation index corresponding to each training sample; determining an error between the training prediction recommendation index of each training sample and the corresponding pre-labeled recommendation index; and reversely propagating the error in the multi-target object prediction model to determine a parameter change value of the multi-target object prediction model when the error obtains a minimum value, and updating the parameter of the multi-target object prediction model based on the parameter change value.

In some embodiments, determining module 2553 is further configured to: determining the activity of the target object based on the behavior data of the target object; and determining the candidate recommendation index prediction network corresponding to the activity of the target object as the target recommendation index prediction network matched with the behavior data of the target object in the candidate recommendation index prediction networks respectively corresponding to the different activities.

In some embodiments, determining module 2553 is further configured to: acquiring the number of times of clicking operation of the target object in unit time and the online time of the target object in unit time from the behavior data of the target object; and determining the liveness which is positively correlated with the number of the clicking operations and negatively correlated with the online time.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, so that the computer device executes the artificial intelligence based information recommendation method described in the embodiment of the present application.

The embodiment of the present application provides a computer-readable storage medium storing executable instructions, wherein the executable instructions are stored, and when being executed by a processor, the executable instructions are to be executed by the processor to perform the artificial intelligence based information recommendation method provided by the embodiment of the present application, for example, the artificial intelligence based information recommendation method shown in fig. 3A-3D.

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

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

By way of example, executable instructions may correspond, but do not necessarily have to correspond, to files in a file system, and may be stored in a portion of a file that holds other programs or data, such as in one or more scripts in a hypertext Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).

As an example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices located at one site or distributed across multiple sites and interconnected by a communication network.

In summary, the ranking features are subjected to cross processing in multiple ways through the embodiment of the application, so that a plurality of feature vectors obtained through the cross processing learn rich semantic information, and then recommendation indexes of different types of target objects are subjected to prediction processing, so that the liveness represented by behavior data of the target objects is directly associated with different candidate recommendation index prediction networks, and the prediction recommendation indexes of the target objects for the information to be recommended are predicted more accurately, and thus accurate personalized recommendation is realized for specific target objects.

The above description is only an example of the present application, and is not intended to limit the 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 is characterized by comprising the following steps:

acquiring a plurality of sorting features of information to be recommended;

2. The method according to claim 1, wherein determining a target recommendation index prediction network matching the behavior data of the target object among the candidate recommendation index prediction networks corresponding to the plurality of different liveness degrees respectively comprises:

determining the activity of the target object based on the behavior data of the target object;

and determining the candidate recommendation index prediction network corresponding to the activity of the target object as a target recommendation index prediction network matched with the behavior data of the target object in the candidate recommendation index prediction networks respectively corresponding to the different activities.

3. The method of claim 2, wherein determining the liveness of the target object based on the behavior data of the target object comprises:

acquiring the number of times of clicking operation of the target object in unit time and the online time of the target object in unit time from the behavior data of the target object;

and determining the liveness which is positively correlated with the number of the clicking operations and negatively correlated with the online time.

4. The method according to claim 1, wherein the performing, by the target recommendation index prediction network, recommendation index prediction processing on the plurality of feature vectors to obtain a prediction recommendation index of the target object corresponding to the information to be recommended includes:

weighting the plurality of feature vectors through the target recommendation index prediction network to obtain semantic features of the target object corresponding to the information to be recommended;

and mapping the semantic features into a prediction recommendation index of the target object corresponding to the information to be recommended through the target recommendation index prediction network.

5. The method according to claim 1, wherein the performing recommendation index prediction processing on the plurality of feature vectors through the target recommendation index prediction network to obtain the prediction recommendation index of the target object corresponding to the information to be recommended includes:

acquiring a plurality of reference target objects of which the portrait similarity of a user portrait with the target object is not less than a portrait similarity threshold;

determining reference behavior data of each reference target object, and determining a reference recommendation index prediction network matched with the reference behavior data in a plurality of candidate recommendation index prediction networks respectively corresponding to different liveness degrees;

performing recommendation index prediction processing on the plurality of feature vectors through the reference recommendation index prediction network to obtain a reference prediction recommendation index of the target object corresponding to the information to be recommended;

and correcting the prediction recommendation index of the target object corresponding to the information to be recommended according to the reference prediction recommendation index and the corresponding portrait similarity to obtain a new prediction recommendation index of the target object corresponding to the information to be recommended.

6. The method according to claim 1, wherein the obtaining of the plurality of ranking features of the information to be recommended includes:

acquiring the sequencing data of the information to be recommended;

when the type of the sequencing data is an information sequence, packaging the sequencing data into a numerical characteristic with a variable-length discrete characteristic;

when the type of the sequencing data is numerical data, packaging the sequencing data into numerical characteristics with the type of continuous characteristics;

when the type of the sequencing data does not belong to numerical data and does not belong to the information sequence, packaging the sequencing data into numerical characteristics of which the type is a fixed-length discrete characteristic;

and embedding the numerical characteristics to obtain the sequencing characteristics.

7. The method of claim 6, wherein said embedding the plurality of numerical features to obtain the plurality of ranking features comprises:

performing the following for any one of the numerical characteristics:

weighting the numerical values by taking hidden vectors respectively corresponding to the numerical values in the numerical characteristics as weights to obtain an embedding dimension value of the numerical characteristics corresponding to one embedding dimension;

and combining the embedded dimension values of the numerical characteristic corresponding to a plurality of the embedded dimensions into a sorting characteristic of the sorting data.

8. The method according to claim 1, wherein the performing feature interleaving processing on the plurality of ranking features based on a plurality of interleaving manners includes:

various processes among the following processes are performed:

performing second-order feature cross processing on the plurality of sequencing features;

performing explicit depth feature intersection processing on the plurality of ranking features;

and performing implicit depth feature intersection processing on the plurality of sequencing features.

9. The method of claim 8, wherein the second order feature interleaving processing the plurality of ordered features comprises:

combining the plurality of sorting features of the information to be recommended for a plurality of times to obtain a plurality of combined features of the information to be recommended;

wherein the ranking features used for each of the combination processes are partially or completely different;

weighting and summing each combined feature of the information to be recommended by taking the recommendation index influence factor of each combined feature of the information to be recommended as a weight parameter;

and 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.

10. The method of claim 8, wherein the explicit depth feature interleaving the plurality of ordered features comprises:

performing explicit feature cross processing of a kth level to obtain an explicit feature vector of the kth level;

pooling the explicit feature vectors of the kth level to obtain a pooling result of the kth level;

splicing the K levels of pooling processing results;

the method comprises the steps that K is an integer larger than or equal to 2, K is an integer variable with values increasing from 1, the value range of K is that K is larger than or equal to 1 and is smaller than or equal to K, when the value of K is 1, input of the K-th level explicit feature cross processing is the multiple sorting features, and when the value of K is larger than or equal to 2 and is smaller than K, input of the K-th level explicit feature cross processing is an explicit feature vector of a K-1 level.

11. The method of claim 10, wherein performing the k-th level explicit feature interleaving comprises:

when the value of k is 1, performing element-by-element multiplication processing on the plurality of sorting features and the plurality of sorting features;

and when K is more than or equal to 2 and less than K, performing element-by-element multiplication on the plurality of ordering characteristics and the explicit characteristic vector of the K-1 level.

12. The method of claim 1,

the obtaining of the plurality of ranking features of the information to be recommended is realized through a feature extraction network,

the feature interleaving processing based on a plurality of interleaving modes is carried out on the plurality of sequencing features through an expert network,

the feature extraction network, the expert network and the candidate recommendation index prediction networks form a multi-target object prediction model;

before obtaining a plurality of ranking features of information to be recommended, the method further comprises:

obtaining a plurality of target object samples corresponding to the information sample to be recommended, wherein behavior data of the target object samples are matched with candidate recommendation index prediction networks respectively corresponding to different activeness;

associating each target object sample and the information to be recommended as a training sample corresponding to the target object sample;

carrying out forward propagation on 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;

carrying out forward propagation on the feature sample of each training sample in a candidate recommendation index prediction network corresponding to a target object sample of the training samples to obtain a training prediction recommendation index corresponding to each training sample;

determining an error between the training prediction recommendation index of each training sample and the corresponding pre-labeled recommendation index;

and reversely propagating the error in the multi-target object prediction model to determine a parameter change value of the multi-target object prediction model when the error takes a minimum value, and updating the parameter of the multi-target object prediction model based on the parameter change value.

13. An artificial intelligence-based information recommendation device, comprising:

14. An electronic device, comprising:

a memory for storing executable instructions;

a processor for implementing the artificial intelligence based information recommendation method of any one of claims 1 to 12 when executing the executable instructions stored in the memory.

15. A computer-readable storage medium storing executable instructions for implementing the artificial intelligence based information recommendation method of any one of claims 1 to 12 when executed by a processor.