CN112380430A

CN112380430A - Artificial intelligence based recommendation method and device and electronic equipment

Info

Publication number: CN112380430A
Application number: CN202011248238.8A
Authority: CN
Inventors: 张新宇
Original assignee: Tencent Technology Beijing Co Ltd
Current assignee: Tencent Technology Beijing Co Ltd
Priority date: 2020-11-10
Filing date: 2020-11-10
Publication date: 2021-02-19

Abstract

The application provides a recommendation method and device based on artificial intelligence, electronic equipment and a computer readable storage medium; the method comprises the following steps: acquiring the exposure times and the effective behavior times of each piece of information in the information set; determining expected behavior characteristics of corresponding information based on the exposure times and the effective behavior times, and determining uncertainty characteristics of the expected behavior characteristics based on the exposure times; establishing a first inverted index table of comprehensive expected behavior indexes including information based on the expected behavior characteristics and the uncertainty characteristics of each information; and sampling the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table, and executing recommendation operation based on the sampling processing result. Through the method and the device, the recommendation accuracy can be improved.

Description

Artificial intelligence based recommendation method and device and electronic equipment

Technical Field

The present application relates to artificial intelligence technologies, and in particular, to a 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.

Information recommendation is an important application of artificial intelligence, a recall stage in a recommendation system is a basis of the whole recommendation system, various efforts are made in the related art for recalling effective information in the recall stage to serve as a sequencing object of the recommendation system, for example, an interest-class recall mode is adopted, but in the process of implementing the embodiment of the application, an applicant finds that the information directly recalled from an inverted index and matched with a user portrait is difficult to effectively depict the positive influence of cold start information on user behaviors and the diverse interests of users, and further influences the precision of information recommendation.

Disclosure of Invention

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

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

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

acquiring the exposure times and the effective behavior times of each piece of information in the information set;

determining expected behavior characteristics corresponding to the information based on the exposure times and the effective behavior times, and determining uncertainty characteristics of the expected behavior characteristics based on the exposure times;

establishing a first inverted index table comprising a composite expected behavior index for each of the information based on the expected behavior signature and the uncertainty signature for the information;

and sampling the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table, and executing recommendation operation based on the sampling processing result.

In the foregoing solution, the adapting a plurality of candidate inverted index tables based on a scenario to which the recommendation request from the client belongs includes:

acquiring real-time data of the scene to extract real-time characteristics of the scene;

calling a neural network model based on the scene features to predict attributes of the scene;

the training sample of the neural network model comprises front-end scene data, and the marking data of the training sample comprises attributes of scenes corresponding to the front-end scene data;

determining an inverted index table adapted to the predicted attribute from the plurality of candidate inverted index tables.

In the foregoing solution, the sampling processing on the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table includes:

obtaining portrait tags from user portraits carried by a recommendation request of a client, and performing adaptation processing on a plurality of candidate inverted index tables based on a scene to which the recommendation request of the client belongs to obtain at least one candidate inverted index table;

wherein the candidate inverted index table is an inverted index table different from the first inverted index table in the plurality of candidate inverted index tables;

acquiring information corresponding to the label consistent with the image label from the first inverted index table to obtain an information set to be sampled;

acquiring the corresponding fraction of each information in the information set to be sampled in each candidate inverted index table;

performing fusion processing on the comprehensive expected behavior index of each information in the information set to be sampled and the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index;

and sampling the information set to be sampled based on the new comprehensive expected behavior index of each information in the information set to be sampled.

The embodiment of the application provides a recommendation device based on artificial intelligence, includes: .

The acquisition module is used for acquiring the exposure times and the effective behavior times of each piece of information in the information set;

a determining module, configured to determine an expected behavior feature corresponding to the information based on the exposure times and the effective behavior times, and determine an uncertainty feature of the expected behavior feature based on the exposure times;

an establishing module, configured to establish a first inverted index table including a comprehensive expected behavior index of each piece of information based on the expected behavior feature and the uncertainty feature of the piece of information;

and the recommending module is used for sampling the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table and executing recommending operation based on the sampling processing result.

In the foregoing solution, before obtaining the exposure times and the effective behavior times of each information in the information set, the obtaining module is further configured to: traversing the release time of each information in the information set according to a fixed time interval; and when the release time of the information is less than the release time threshold, removing the corresponding information from the information set.

In the foregoing solution, the determining module is further configured to:

determining an expected behavior feature corresponding to the information based on the following expected behavior feature functions:

wherein E is₁For the expected behavior feature, p is a ratio between the number of the effective behaviors and the number of the exposures, z is a confidence coefficient parameter, and n is the number of the exposures of the information.

In the foregoing solution, the determining module is further configured to:

determining an uncertainty feature corresponding to the expected behavior feature based on the following uncertainty feature functions:

wherein E is₂And T is the sum of the exposure times of all the information in the information set, and n is the exposure time of the information.

In the foregoing solution, the establishing module is further configured to: determining a first behavioral component that is positively correlated with the desired behavioral characteristic; determining a second behavior component that is positively correlated with the uncertainty feature; performing fusion processing on the first behavior component and the second behavior component to obtain a comprehensive expected behavior index of the information; and establishing the first inverted index table by taking the label of the information as a key and taking the information and the comprehensive expected behavior index corresponding to the information as values.

In the foregoing solution, the recommending module is further configured to: obtaining a portrait label in a user portrait from a recommendation request of a client; acquiring a plurality of information corresponding to the label consistent with the image label from the first inverted index table to form an information set to be sampled; and sampling the information set to be sampled based on the comprehensive expected behavior index of each information in the information set to be sampled.

In the foregoing solution, the recommending module is further configured to: performing sampling processing with the sampling frequency threshold, and performing the following processing in the process of each sampling processing: converting the comprehensive expected behavior index of each information in the information set to be sampled into a subinterval in a numerical interval; generating a random number corresponding to the numerical value interval through a random function; and determining the information corresponding to the subinterval in which the random number falls as the information obtained by sampling, and moving out of the information set to be sampled.

In the foregoing solution, the recommending module is further configured to: based on the scene to which the recommendation request of the client belongs, adapting a plurality of candidate inverted index tables; and when the first inverted index table is adapted to the multiple candidate inverted index tables, sampling the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table.

In the above solution, the multiple candidate inverted index tables further include a second inverted index table, and the second inverted index table includes a descending order based on the information creation time; the recommendation module is further configured to: and when the recommended scene corresponding to the recommendation request is a recommended scene with aging attribute, determining the second inverted index table as an inverted index table adapted to the recommendation request.

In the foregoing solution, the recommending module is further configured to: when the second reverse index table is adapted to the multiple candidate reverse index tables, updating the release time of multiple pieces of information ranked at the top in the second reverse index table; wherein, the update value used for updating the release time and the information posterior quality parameter of the information form a positive correlation relationship; and acquiring a plurality of pieces of information in the top order from the updated second inverted index table to execute recommendation operation.

In the above scheme, the multiple candidate inverted index tables further include a third inverted index table, and the third inverted index table includes a descending order based on the information posterior quality parameter; the recommendation module is further configured to: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with a quality attribute, determining the third inverted index table as an inverted index table adapted to the recommendation request.

In the foregoing solution, the recommending module is further configured to: when the third reverse index table is adapted to the multiple candidate reverse index tables, updating the posterior quality parameters of the information of the multiple pieces of information which are ranked at the front in the third reverse index table; wherein an update value for updating the information posterior quality parameter is negatively correlated with the number of exposures of the information; and acquiring a plurality of pieces of information in the top order from the updated third inverted index table to execute recommendation operation.

In the above solution, the multiple candidate inverted index tables further include a fourth inverted index table, and the fourth inverted index table includes a descending order based on the portrait similarity; the recommendation module is further configured to: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with interest attributes, determining the fourth inverted index table as an inverted index table adapted to the recommendation request.

In the foregoing solution, the recommending module is further configured to: when the fourth reverse index table is adapted to the multiple candidate reverse index tables, updating information portrait similarity of multiple pieces of information ranked in the fourth reverse index table; wherein, the updating value used for updating the image similarity is in negative correlation with the sorting position of the information; and acquiring a plurality of pieces of information in the top order from the updated fourth inverted index table to execute recommendation operation.

In the foregoing solution, the recommending module is further configured to: acquiring real-time data of the scene to extract real-time characteristics of the scene; calling a neural network model based on the scene features to predict attributes of the scene; the training sample of the neural network model comprises front-end scene data, and the marking data of the training sample comprises attributes of scenes corresponding to the front-end scene data; determining an inverted index table adapted to the predicted attribute from the plurality of candidate inverted index tables.

In the foregoing solution, the recommending module is further configured to: obtaining portrait tags from user portraits carried by a recommendation request of a client, and performing adaptation processing on a plurality of candidate inverted index tables based on a scene to which the recommendation request of the client belongs to obtain at least one candidate inverted index table; wherein the candidate inverted index table is an inverted index table different from the first inverted index table in the plurality of candidate inverted index tables; acquiring information corresponding to the label consistent with the image label from the first inverted index table to obtain an information set to be sampled; acquiring the corresponding fraction of each information in the information set to be sampled in each candidate inverted index table; performing fusion processing on the comprehensive expected behavior index of each information in the information set to be sampled and the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index; and sampling the information set to be sampled based on the new comprehensive expected behavior index of each information in the information set to be sampled.

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 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 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:

based on the exposure times and the effective behavior times of the information, the expected behavior characteristics for user behavior prediction and the uncertainty characteristics corresponding to the expected behavior characteristics are specifically drawn for different information, the contribution of the information with different exposure times (such as cold start information and historical exposure information with different degrees) to the user behavior prediction is considered, meanwhile, through sampling processing, the diversity of information distribution and information distribution of middle and long tail interests is facilitated, the coverage range of the sampled information is wide, the information which the user is interested in can be better mined, the information recommendation precision of subsequent information recommendation is guaranteed, meanwhile, invalid recommendation is effectively avoided, and further computing resources related to recommendation logic in a server are saved.

Drawings

FIG. 1 is a schematic diagram of an architecture of an artificial intelligence based 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 recommendation methods provided by embodiments of the present application;

FIG. 4 is an overall architecture diagram of an artificial intelligence based recommendation method provided by an embodiment of the present application;

5A-5B are data flow diagrams of artificial intelligence based recommendation methods provided by embodiments of the application.

Detailed Description

In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described in further detail with reference to the attached 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 creative efforts shall 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, but rather the terms "first \ second \ third" are used to interchange specific orders or sequences, where appropriate, so as to enable the embodiments of the application described herein to 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) And (4) personalized recommendation, wherein existing and potential interest points of a user are quantitatively mined and expressed by utilizing the principle and method of a recommendation system, so that information which is interested in the user is recommended for the user.

2) For example, in an article system, each article in the forward index corresponds to multiple attributes, such as a title, a label, and a publisher, and the reverse index queries an article under the label with the attribute (the label or the title corresponding to the article) as a query condition, for example, according to the label.

3) The expected behavior feature is a historical average profit value, for example, a historical click rate (historical average profit value) of any information in a recommendation system, and the historical click rate can predict the information real click rate to some extent.

4) And the uncertainty characteristic is a value of a supremum of uncertainty representing the historical average profit value, for example, because the exposure times of some information in the recommendation system are small, the historical click rate of the information is not necessarily accurate to predict the actual click rate of the information, so that the expected behavior characteristic is corrected through the uncertainty characteristic.

5) The comprehensive expected behavior index is a value obtained by performing weighted calculation based on expected behavior characteristics and uncertainty characteristics, and can be used for predicting real click rates of information with different exposure degrees in a recommendation system.

In the related art, with the rapid development of information technology, various types of information are increased explosively, and users cannot obtain information really meaningful for themselves from massive information, so that personalized information recommendation becomes an important research subject.

In a recall module of a recommendation system, recall based on user interest is an important recall strategy, firstly, inverted arrangement is established for all information in an information pool, generally comprising primary classification inverted arrangement, secondary classification inverted arrangement, label inverted arrangement and the like, then, accumulated portrait is established for a user by using historical click behaviors of the user, when the user requests information (recommendation request), interest recall firstly pulls the user portrait, pulls corresponding information from an inverted index based on the portrait label, intercepts a plurality of pieces of pulled information, and finally recommends the information to the user through a sequencing module and a reordering module.

The recall based on interest classes firstly obtains interest points of a user from a user portrait, then pulls corresponding information from an inverted index, then intercepts K pieces of pulled information, sorts and scatters the information, and finally presents the information to the user, the establishment of the inverted index is the key of the recall, the information creation time of the information and the actual click rate of the information are mainly considered at present, and the interception algorithm of the pulled information can influence the quality and diversity of the recommended information, and the information is generally sorted and intercepted based on inverted scores or sorted and intercepted based on matching scores of the user portrait.

The index side in the recommendation system establishes the inverted ranks based on the information creation time or the information heat, the interest recall is usually performed by inverted-pulling based on the head interests of the user, and the pulled inverted ranks are subjected to head truncation based on inverted scores or matching scores of the information and the user, so that the following technical problems can be caused:

1. if the creation time of the information is considered to establish the inverted arrangement, only the publishing time of the information is used, the posterior quality of the information is not utilized, namely the posterior feedback of the information to the information by the user after the information is published is not utilized, the distribution of high-quality information is not facilitated, and the recommendation accuracy of the information is influenced, so that the experience of the user is influenced;

2. if the inverted arrangement is established by considering the click rate of the information, only the posterior quality of the information is used, the newly issued information is exposed less times, the confidence of the information quality score is low, the distribution of the cold start information is not facilitated, and the information with high click rate in the information issued historically is always positioned at the head of the inverted index, so that the information recommended to the user is all global hot information, the diversity of the recommended information is influenced, and the experience of the user is further influenced;

3. if a plurality of pieces of information with front sequencing are intercepted based on the inverted scores, the recommended information is made to depend too much on the calculation mode of the inverted scores, so that the information at the tail part of the inverted scores cannot be exposed effectively, and the diversity of the recommended information is not facilitated;

4. if the inverted information is scored based on the user portrait and a plurality of pieces of information with the front ranking are intercepted, the inverted information is easily influenced by the distribution of the user portrait, the head interests of most users have higher weight, the recommendation result is concentrated on the head interests of the users, the problem that the information recalled each time is single still exists, and the recommendation and recall information diversity of the long-tail interests of the users are influenced.

The embodiment of the application provides a recommendation method and device based on artificial intelligence, an electronic device and a computer-readable storage medium, which can take into account contributions of information (such as cold start information and different degrees of historical exposure information) of different exposure times to user behavior prediction, so as to improve recommendation accuracy.

The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, big data and an artificial intelligence platform, and the server may be directly or indirectly connected through wired or wireless communication, which is not limited herein.

An artificial intelligence cloud Service is also commonly referred to as AIaaS (AI as a Service, chinese). The method is a service mode of an artificial intelligence platform, and particularly, the AIaaS platform splits several types of common AI services and provides independent or packaged services at a cloud. This service model is similar to the one opened in an AI theme mall: all developers can access one or more artificial intelligence services provided by the platform through an API (application programming interface) interface, and some of the sophisticated developers can also use the AI framework and the AI infrastructure provided by the platform to deploy and operate and maintain the own dedicated cloud artificial intelligence services.

Referring to fig. 1, fig. 1 is a schematic diagram of an architecture of a recommendation system based on artificial intelligence provided in an embodiment of the present application, where the recommendation system may be used to support recommendation scenes of various information, such as an application scene for recommending news, an application scene for recommending commodities, an application scene for recommending videos, and the like, and according to different application scenes, the information may be news, video articles, pictures and texts, and the like, and may also be information related to products (e.g., real objects such as clothes, virtual articles such as game props). In the process that a user uses a client, a terminal 400 reports collected interactive behaviors (effective behavior times) and exposure states (exposure times) of the user aiming at information to a server 200, the server 200 can acquire the information from a database 500, establish a first inverted index table comprising comprehensive expected behavior indexes of the information based on the exposure times and the effective behavior times, store the first inverted index table into the database 500, respond to a recommendation request of the terminal 400, sample the information in the first inverted index table based on the comprehensive expected behavior indexes of the information in the first inverted index table in the database 500, execute recommendation operation based on a sampling processing result, and recommend the information in the sampling processing result to the terminal 400.

A specific architecture of the recommendation system is described below, in which the terminal 400 is connected to the server 200 through the network 300, the network 300 may be a wide area network or a local area network, or a combination of the two, and functions of the server 200 may be abstracted to establish the first inverted index table and perform sampling processing based on the first inverted index table to obtain recalled information. The server 200 regularly maintains the first inverted index table in the database 500 based on the interactive behavior (effective behavior times) and the exposure state (exposure times) aiming at the information reported by the terminal 400, receives the recommendation request of the terminal 400, and the server 200 samples the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table in the database 500 and returns the information in the sampling processing result to the terminal 400 for presentation.

Referring to fig. 2, fig. 2 is a schematic structural diagram of a server 200 applying an artificial intelligence based recommendation method according to an embodiment of the present application, where the server 200 shown in fig. 2 includes: at least one processor 210, memory 250, and at least one network interface 220. The various components in server 200 are coupled together by a bus system 240. It is understood that the bus system 240 is used to enable communications among the 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, or the like, wherein the general purpose Processor may be a microprocessor or any conventional Processor, or the like.

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 herein is intended to comprise any suitable type of memory.

In some embodiments, memory 250 is capable of storing data, examples of which include programs, modules, and data structures, or a subset or superset thereof, to support various operations, 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 recommendation apparatus provided by the embodiments of the present application may be implemented in software, and fig. 2 illustrates an artificial intelligence based recommendation apparatus 255 stored in a memory 250, which includes a plurality of modules, which may be software in the form of programs and plug-ins, and includes the following software modules: an obtaining module 2551, a determining module 2552, a building module 2553 and a recommending module 2554, which are logical and thus can be arbitrarily combined or further split according to the implemented functions, and the functions of the respective modules will be described below.

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

Referring to fig. 3A, fig. 3A is a schematic flowchart of an artificial intelligence based recommendation method provided in an embodiment of the present application, and will be described with reference to the steps shown in fig. 3A.

In step 101, the number of exposures and the number of active actions for each information in the set of information is obtained.

In some embodiments, before acquiring the exposure times and the effective behavior times of each information in the information set, the following technical solutions may also be performed: traversing the release time of each information in the information set according to a fixed time interval; and when the release time of the information is less than the release time threshold, removing the corresponding information from the information set.

As an example, since the information set is used as a basis for information recall, it is necessary to ensure that the release time of the information in the information set is within a release time threshold, for example, the current time is 8 o ' clock at 11 month, 9 th and 8 o ' clock at 11 month, 6 th and 6 th at 2020 year, and if the release time of the information is 7 o ' clock at 11 month, 6 th and 7 th at 2020 year, the release time of the characterizing information is less than the release time threshold, the information needs to be removed from the information set, the fixed time interval may be one day or half day, or time with finer granularity.

In step 102, the expected behavior feature of the corresponding information is determined based on the exposure times and the effective behavior times, and the uncertainty feature of the expected behavior feature is determined based on the exposure times.

In some embodiments, the step 102 of determining the expected behavior characteristics of the corresponding information based on the number of exposures and the number of valid behaviors may be implemented by the following technical solutions: determining an expected behavior feature of the corresponding information based on the following expected behavior feature functions:

wherein E is₁For the expected behavior feature, p is the ratio between the number of valid behaviors and the number of exposures, z is the confidence parameter, and n is the number of exposures for the information.

By way of example, the effective behavior number may be click behavior, or may be a more advanced user behavior, such as an interaction behavior between a user and information, including a forwarding behavior, a comment behavior, and the like, z is a standard level of normal distribution, for example, z is 1.96 with a 95% confidence, and the expected behavior characteristics calculated by the above formula (1) may be understood as a historical average benefit of the recommendation system, and the benefit in the recommendation system is usually an index related to the user behavior, such as a historical average forwarding rate, a historical average collection rate, and most commonly, a historical average click rate.

In some embodiments, the historical average click rate may be directly used as the expected behavior feature, and then the comprehensive expected behavior index for each piece of information may be calculated based on the historical average click rate, and the applicant finds that, in implementing the embodiments of the present application, it is very inaccurate to predict the comprehensive expected behavior index by simply using the historical average click rate as the expected behavior feature at the beginning of recommendation, because the click rate is equal to the click times/exposure times, i.e. only the relation of the proportion is considered, but the sample number size is not considered, because the sample number is small, the proportion is actually inaccurate, and the proportion is more accurate when the sample number is larger, the true situation is reflected more, for example, there are three advertisements: the click number of the advertisement A is 2, the exposure number is 10, the click rate of the advertisement B is 20, the exposure number is 100, the click rate of the advertisement C is 200, the exposure number is 1000, and the click rates of the advertisement B and the advertisement C are all 0.2, but obviously, from the confidence level, the click rate of the advertisement C is 0.2, which is more real and more reliable, in order to measure the influence of the sample number on the confidence interval of the click rate, the formula (1) is introduced, the real click rate range under a certain confidence level (for example, the confidence level of the click rate is 95%) is calculated by using the formula (1), the click number is 2, the exposure number is 10, the confidence interval of the click rate with 95% is (0.07,0.45), for the advertisement B, the click number is 20, the exposure number is 100, the confidence interval of the click rate with 95% is (0.14,0.27), for the advertisement C, the number of clicks was 200, the number of exposures was 1000, the confidence interval for the click rate with 95% confidence was (0.18,0.22), in practical application, the lowest value can be taken as the corrected click rate, that is, the corrected click rate of the advertisement A is 0.07, the corrected click rate of the advertisement B is 0.14, the corrected click rate of the advertisement C is 0.18, this is equivalent to giving a certain weight to the click rate with insufficient number of samples through the above formula (1), and by the above embodiment, it is ensured that during cold start, the historical average click rate with higher confidence (not the statistical historical average click rate, but the corrected historical average click rate) can be obtained as the expected behavior characteristics even if the exposure times are insufficient, therefore, accurate comprehensive expected behavior indexes are obtained based on the expected behavior characteristics subsequently, and the recommendation accuracy of information obtained based on the comprehensive expected behavior indexes is improved.

In some embodiments, the step 102 of determining the uncertainty characteristic of the expected behavior characteristic based on the number of exposures may be implemented by the following technical solutions: determining an uncertainty feature corresponding to the expected behavior feature based on the following uncertainty feature functions:

wherein E is₂T is the sum of the exposure times of all the information in the information set, and n is the exposure time of the information.

By way of example, the expected behavior feature may be understood as an empirically obtained benefit, and the recall stage selects the information with the highest expected behavior feature to recall to obtain a higher cumulative reward, but because the reward is random, the probability of the reward of each piece of information is not accurately estimated, and perhaps the information with the highest true probability of the reward is not the information with the highest expected behavior feature, so it is necessary to search cold information, where searching cold information may bring uncertainty benefits, the uncertainty features of the expected behavior feature have a larger value, the uncertainty benefits brought by the search of cold information are larger, and different pieces of information are randomly searched to obtain a more accurate probability estimate of the reward of each piece of information, so as to find the true optimal piece of information, but to recall the information with the lowest probability estimate of the current reward, meaning that some chance of recalling information with high historical average profit is lost, the above embodiment can acquire uncertainty characteristics as a supplement to expected behavior characteristics, so as to improve the chance of recalling information with low exposure rate or not exposed in the cold starting process in the actual recommendation scene, so as to excavate diversified interests of users.

As an example, the uncertainty characteristic of the expected behavior feature may be calculated based on the above formula (2), when n is smaller or T is relatively larger, the information is characterized as being cooler information, the cooler the information is, the smaller the expected behavior feature of the information is, the larger the uncertainty characteristic of the expected behavior feature is, the expected behavior feature is corrected to a greater extent by the uncertainty characteristic of the expected behavior feature, when n is larger or T is relatively smaller, the information is characterized as being hotter information, the hotter the information is, the larger the expected behavior feature is, the smaller the uncertainty characteristic of the expected behavior feature is, that is, the expected behavior feature is corrected to a lesser extent by the uncertainty characteristic of the expected behavior feature.

In step 103, a first inverted index table including a composite expected behavior index of the information is established based on the expected behavior signature and the uncertainty signature of each information.

In some embodiments, referring to fig. 3B, fig. 3B is a schematic flowchart of a recommendation method based on artificial intelligence provided in an embodiment of the present application, and in step 103, a first inverted index table including a comprehensive expected behavior index of information is established based on an expected behavior feature and an uncertainty feature of each information, which may be implemented through steps 1031-1034.

In step 1031, a first behavior component that is positively correlated with the desired behavior feature is determined.

In step 1032, a second behavior component that is positively correlated to the uncertainty feature is determined.

In step 1033, the first behavior component and the second behavior component are fused to obtain a comprehensive expected behavior index of the information.

By way of example, if the first inverted index table is built based only on the first behavior component, the overall expected behavior index in the first inverted index table is based entirely on a greedy strategy, which is easily trapped in local extrema, e.g., several pieces of information with better historical performance are always recalled, the second behavior component has the significance that if one cold information is known too little, the confidence of its expected behavior characteristic (historical average profit, e.g., historical average click rate) is low at this time, the uncertainty is high, the confidence interval is large, i.e., the expected behavior characteristic at this time is not believed to be its true average profit, so this information needs to be selected to obtain more exploration results, and thus the second behavior component can be understood as an index measuring how much is known about a certain information, the less knowledge is the larger the second behavior component is, fusing the second behavior component on the basis of the first behavior component may be understood as when knowledge about a piece of information is not sufficient, the piece of information has the opportunity to be selected even if the average benefit of this piece of information is low.

In step 1034, a first inverted index table is established with the tag of the information as a key and the information and the comprehensive expected behavior index of the corresponding information as values.

As an example, assuming that there are 3 pieces of information, information a, information B, and information C, each having a label 1, the comprehensive expected behavior indexes of information a, information B, and information C are 0.5, 0.45, and 0.3, respectively, any row in the first inverted index table may be a key value pair, for example, the key is label 1, the value is information a (0.5), information B (0.45), and information C (0.3), and a subsequent online process may be recalled based on the first inverted index table.

For example, the process of step 101-103 may be completed online or offline, and the first inverted index table is updated according to a fixed update time interval, so that the index data according to which the recall is performed based on the first inverted index table in the subsequent online process is ensured to be valid, and the recommendation accuracy of the recommendation system is ensured.

In step 104, sampling processing is performed on the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table, and a recommendation operation is executed based on the sampling processing result.

As an example, performing the recommendation operation based on the sampling processing result is actually recommending for the sampled information, and the recommendation operation may include a sorting operation, a reordering operation, a scattering operation, and the like.

In some embodiments, referring to fig. 3C, fig. 3C is a schematic flowchart of a recommendation method based on artificial intelligence provided in the embodiment of the present application, and in step 104, based on the comprehensive expected behavior index of the information in the first inverted index table, the information in the first inverted index table is sampled, which may be implemented by step 1041-1043.

In step 1041, a portrait label in the user portrait is obtained from the recommendation request from the client.

In step 1042, a plurality of information corresponding to the label consistent with the image label is obtained from the first inverted index table to form an information set to be sampled.

By way of example, the recommendation request of the client carries a user portrait, the user portrait includes a plurality of portrait tags, the portrait tags may represent interests of the user, the portrait tags may also represent social attributes of the user, and the like, for example, a pregnant woman, a mother and a baby, and the like, the first inverted index table stores a plurality of tags, each tag has a corresponding plurality of information, for example, the first inverted index table also has a mother and baby tag, a plurality of information corresponding to the tag (mother and baby) consistent with the portrait tag (mother and baby) is obtained from the first inverted index table to form an information set to be sampled, where the consistency may be understood as complete consistency, or the matching degree of the two tags is greater than a matching degree threshold value and is default to be consistent, for example, when the matching degree of the "pregnant woman" is not completely consistent with the "mother and baby", but the matching degree is greater than the matching degree threshold value, the tag is a portrait, a plurality of information corresponding to the 'mother and infant' label can be acquired.

In step 1043, the information set to be sampled is sampled based on the comprehensive expected behavior index of each information in the information set to be sampled.

In some embodiments, in step 1043, the information is sampled based on the comprehensive expected behavior index of each information in the information set to be sampled, and the following technical solution may be implemented: performing sampling processing with the sampling frequency threshold, and performing the following processing in the process of each sampling processing: converting the comprehensive expected behavior index of each information in the information set to be sampled into a subinterval in a numerical interval; generating random numbers corresponding to the numerical value intervals through a random function; and determining the information corresponding to the subinterval in which the random number falls as the information obtained by sampling, and moving out of the information set to be sampled.

As an example, the sampling time threshold is the number of information that needs to be recalled finally, for example, when the number of information that needs to be recalled is 5000, 5000 times of non-playback sampling is performed, and taking the case of performing one time of non-playback sampling, assuming that there are three pieces of information in the information set to be sampled, the comprehensive expected behavior indexes of the information a, the information B, and the information C are 0.5, 0.3, and 0.2, respectively, for the comprehensive expected behavior index of each piece of information in the information set to be sampled, a corresponding subinterval is determined within a certain numerical interval, the sampling probability of the information a is 0.5/(0.5+0.3+0.2), assuming that the numerical interval is 0 to 1, the comprehensive expected behavior index of the information a 0.5 is converted into the subinterval 0 to 0.5, the comprehensive expected behavior index of the information B0.3 is converted into the subinterval 0.5 to 0.8, and the comprehensive expected behavior index of the information C0.2 is converted into the subinterval 0.8 to 1, generating a random number of a numerical range of 0-1 through a random function, if the generated random number is 0.6, determining the information B as sampled information and moving the sampled information B out of an information set to be sampled, and sampling the information A and the information C when the next sampling is performed, wherein the sampling probability of the information A is 0.5/(0.5+0.2), and if the numerical range is still 0-1, converting the comprehensive expected behavior index 0.5 of the information A into the sub-range of 0-5/7, converting the comprehensive expected behavior index 0.2 of the information C into the sub-range of 5/7-1, and then continuously generating the random number for sampling. Through the sampling implementation mode, information without high comprehensive expected behavior indexes can be recalled, and therefore non-head middle and long tail interests are mined, and interest diversity in the recommendation process is met.

In some embodiments, referring to fig. 3D, fig. 3D is a schematic flowchart of a recommendation method based on artificial intelligence provided in the embodiment of the present application, and in step 104, based on the comprehensive expected behavior index of the information in the first inverted index table, the information in the first inverted index table is sampled, which may be implemented in step 1044-1045.

In step 1044, the adaptation of the multiple candidate inverted index tables is performed based on the scene to which the recommendation request of the client belongs.

In step 1045, when the first inverted index table is adapted to the multiple candidate inverted index tables, sampling the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table.

In some embodiments, in addition to the first inverted index table, there are other inverted index tables established according to other criteria in the recommendation system, for example, a second inverted index table established based on descending order of information creation time, a third inverted index table established based on descending order of information posterior quality parameters, and a fourth inverted index table established based on descending order of image similarity, where different inverted index tables are applicable to different scenes, for example, a recommended scene for cold start or a scene requiring mining of diverse interests of a user, and may be adapted to the first inverted index table from among multiple candidate inverted index tables, so that information in the first inverted index table is sampled based on a comprehensive expected behavior index of the information in the first inverted index table.

In some embodiments, the adapted inverted index table may also be selected according to the data completeness or timeliness of the information of the database, for example: and releasing data of three dimensions of time, posterior quality and portrait data, and selecting the reverse index in a corresponding mode for recall if the integrity of the data of the information is the highest.

In some embodiments, the plurality of candidate inverted index tables further comprises a second inverted index table, and the second inverted index table comprises a descending order based on the information creation time; in step 1044, based on the scene to which the recommendation request of the client belongs, the adaptation of the multiple candidate inverted index tables is performed, which can be implemented by the following technical scheme: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with aging attribute, determining the second inverted index table as the inverted index table adapted to the recommendation request.

In some embodiments, when a second inverted index table is adapted to the plurality of candidate inverted index tables, performing publication time update on a plurality of information ranked in the second inverted index table; wherein, the updating value used for updating the release time and the information posterior quality parameter of the information form a positive correlation relationship; and acquiring a plurality of pieces of information which are ranked at the top from the updated second inverted index table to execute the recommendation operation.

As an example, when the recommendation scene corresponding to the recommendation request is a recommendation scene with aging attribute, the second inverted index table is determined as the inverted index table adapted to the recommendation request, the recommendation scene with aging attribute may be a recommendation scene of morning news, and it is usually required to ensure that all information is the latest information in morning news, therefore, when the second inverted index table is adapted to the plurality of candidate inverted index tables, the distribution time update is performed on a plurality of pieces of information ranked earlier in the second inverted index table, which includes a key value pair, the key is a label, the value is the information and the distribution time of the information, for example, if the label is a mother, a plurality of pieces of information (information with set proportion or quantity) with earlier distribution time in the used information corresponding to the mother-baby label are obtained, posterior feedback correction values (update values) are introduced for their distribution time, the release time (from the combination of the year, month, day time division fields) plus the posterior feedback correction value (the quantization value of the posterior feedback) to obtain a new release time, the updated value for updating the release time is in positive correlation with the information posterior quality parameter of the information, the posterior feedback correction value is positive, which is equivalent to the release time of the high-quality and early information is corrected to a larger value, the high-quality information is prevented from sinking for a long time of release, the release time of the header information is corrected, so a plurality of pieces of information in the front of the sequence are obtained from the updated second inverted index table to execute the recommendation operation, the quality correction is already carried out, so a plurality of pieces of information in the front of the sequence (information with set proportion or quantity) can be directly intercepted to execute the recommendation operation, or the mode without the back sampling is adopted, the corrected release time is used as the comprehensive expected behavior index, and performing non-playback sampling, and then executing recommended operation based on a sampling result.

By combining time and quality through the embodiment, on the premise of meeting time sensitivity, weight correction is carried out to a certain extent on information which is issued for a long time, so that recommendation quality in a time-sensitive scene is ensured.

In some embodiments, the plurality of candidate inverted index tables further comprises a third inverted index table, and the third inverted index table comprises a descending ordering based on the information a posteriori quality parameter; in step 1044, based on the scene to which the recommendation request of the client belongs, the adaptation of the multiple candidate inverted index tables is performed, which can be implemented by the following technical scheme: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with a quality attribute, determining the third inverted index table as the inverted index table adapted to the recommendation request.

In some embodiments, when a third inverted index table is adapted to the plurality of candidate inverted index tables, performing information posterior quality parameter update on a plurality of information ranked in the third inverted index table; wherein, the updating value used for updating the posterior quality parameter of the information is negatively correlated with the exposure times of the information; and acquiring a plurality of pieces of information which are ranked at the top from the updated third inverted index table to execute the recommendation operation.

As an example, the information posterior quality parameter may be a parameter for measuring information quality fed back by a user, such as a click rate, a forwarding rate, a collection rate, or a parameter for measuring information quality obtained based on a combination of the click rate, the forwarding rate, the collection rate, and the like.

As an example, when the recommended scene corresponding to the recommendation request is a recommended scene with a quality attribute, the third inverted index table is determined as an inverted index table adapted to the recommendation request, the recommended scene with the quality attribute may be a recommended scene of a hot search, and it is generally required to ensure that all information is hottest information in the hot search, therefore, when the third inverted index table is adapted to the candidate inverted index tables, the posterior information quality parameter update is performed on the information of the information in the third inverted index table, similar to the first inverted index table, the posterior information quality parameter update is performed, the posterior information quality parameter update includes a key value pair, the key is a label, the value is an information and an information posterior quality parameter of the information, for example, if the image label is a mother and baby, a plurality of information (information of set proportion or quantity) in front of the posterior information quality parameter in the used information corresponding to the mother and baby label are obtained, the exposure number correction value (update value) is introduced for their posterior information quality parameters, adding an exposure correction value (normalized value of the exposure) to the information posterior quality parameter to obtain a new information posterior quality parameter, wherein the exposure correction value is negative, the update value for updating the information posterior quality parameter is negatively related to the exposure of the information, which is equivalent to pull down the ranking advantage of the earlier released information due to the accumulated exposure, and since the information posterior quality parameter of the header information is corrected, a plurality of pieces of information in the front ranking are obtained from the updated third inverted index table to execute the recommendation operation, since the exposure correction has been performed, a plurality of pieces of information in the front ranking (information with set proportion or quantity) can be directly intercepted to execute the recommendation operation, or the corrected information posterior quality parameter is taken as the comprehensive expected behavior index in the manner of no back sampling, and performing non-playback sampling, and then executing recommended operation based on a sampling result.

By combining exposure and quality through the embodiment, on the premise of meeting the requirement of quality sensitivity, the information with higher posterior quality parameters of the information due to excessive exposure is subjected to certain weight correction (pull-down), so that the recommendation quality in a quality sensitive scene is ensured.

In some embodiments, the plurality of candidate inverted index tables further includes a fourth inverted index table, and the fourth inverted index table includes a descending ordering based on portrait similarity; in step 1044, based on the scene to which the recommendation request of the client belongs, the adaptation of the multiple candidate inverted index tables is performed, which can be implemented by the following technical scheme: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with the interest attribute, determining the fourth inverted index table as the inverted index table adapted to the recommendation request.

In some embodiments, when a fourth inverted index table is adapted among the plurality of candidate inverted index tables, performing information portrait similarity updating on a plurality of information ranked in the fourth inverted index table; wherein, the updating value used for updating the image similarity is in negative correlation with the sorting position of the information; and acquiring a plurality of pieces of information which are ranked at the top from the updated fourth inverted index table to execute the recommendation operation.

As an example, when the recommendation scene corresponding to the recommendation request is a recommendation scene with an interest attribute, the fourth inverted index table is determined as an inverted index table adapted to the recommendation request, and the recommendation scene with a quality attribute may be a recommendation scene of a small video, and it is usually necessary to ensure that information best meets the interest of the user in small video recommendation, so when the fourth inverted index table is adapted to the candidate inverted index tables, information representation similarity updating is performed on a plurality of pieces of information ranked next in the fourth inverted index table, similar to the first inverted index table, including a key value pair, where the key is a label and the value is information and information representation similarity of the information, for example, the representation label is a mother and then a plurality of pieces of information (information with set proportion or number) with next information representation similarity in used information corresponding to the mother and baby label are obtained, a similarity correction value (updated value) is introduced for their information representation similarity, the information image similarity plus the similarity correction value obtains a new information image similarity, the update value for updating the image similarity is in a negative correlation relationship with the sorting position of the information (may be an update value with a decay trend in the long tail, where the similarity correction value is a positive number, although the similarity is increased, the decay trend in the long tail is increased), which is equivalent to generally reducing the sorting disadvantage that the information of the long tail part always sinks because of low similarity, since the information image similarity of the tail information is corrected, a plurality of pieces of information in the front sorting are obtained from the updated fourth inverted index table to execute the recommendation operation, since the correction of the exposure times has been performed, a plurality of pieces of information in the front sorting (information of a set proportion or quantity) can be directly intercepted, the recommendation operation is executed, or the above-mentioned way of non-resampling is adopted, and taking the corrected information portrait similarity as the comprehensive expected behavior index, performing non-playback sampling, and executing recommendation operation based on a sampling result.

By combining the portrait similarity with the sequencing position in the embodiment, the portrait similarity of the middle and long tail information is corrected in the forward direction on the premise of meeting the similarity of the interests, so that the diversity of information recommendation in the interest sensitive scene is ensured.

In some embodiments, in step 1044, based on a scene to which a recommendation request of a client belongs, the adaptation of the multiple candidate inverted index tables may be implemented by the following technical solutions: acquiring real-time data of a scene to extract real-time characteristics of the scene; calling a neural network model based on the scene characteristics to predict the attributes of the scene; the training sample of the neural network model comprises front-end scene data, and the marking data of the training sample comprises attributes of scenes corresponding to the front-end scene data; an inverted index table that is adapted to the predicted attribute is determined from the plurality of candidate inverted index tables.

As an example, an inverted index table adapted to the predicted attribute may be determined from a plurality of candidate inverted index tables through a neural network model, a training sample of the neural network model includes front-end scene data, and labeling data of the training sample includes attributes of a scene corresponding to the front-end scene data.

In some embodiments, in step 104, based on the comprehensive expected behavior index of the information in the first inverted index table, the information in the first inverted index table is sampled, and the following technical solutions may be implemented: obtaining portrait tags from user portraits carried by a recommendation request of a client, and performing adaptation processing on a plurality of candidate inverted index tables based on a scene to which the recommendation request of the client belongs to obtain at least one candidate inverted index table; the candidate reverse index table is a reverse index table different from the first reverse index table in the multiple candidate reverse index tables; acquiring information corresponding to the label consistent with the image label from the first inverted index table to obtain an information set to be sampled; acquiring the corresponding fraction of each information in the information set to be sampled in each candidate inverted index table; fusing the comprehensive expected behavior index of each information in the information set to be sampled with the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index; and sampling the information set to be sampled based on the new comprehensive expected behavior index of each information in the information set to be sampled.

As an example, a portrait label is obtained from a user portrait carried by a recommendation request of a client, and adaptation processing of a plurality of candidate inverted index tables is performed based on a scene to which the recommendation request of the client belongs to obtain at least one candidate inverted index table; the candidate reverse index table is a reverse index table different from the first reverse index table in the multiple candidate reverse index tables; acquiring information corresponding to a label consistent with an image label from a first inverted index table, acquiring an information set to be sampled, including information A and information B for example, acquiring a corresponding score of each information in the information set to be sampled in each candidate inverted index table, such as a score corresponding to the release time of the information A in a second inverted index table and a score corresponding to the quality of the information A in a third inverted index table, and fusing a comprehensive expected behavior index of each information in the information set to be sampled with the corresponding score in each other inverted index table to acquire a new comprehensive expected behavior index; and finally, sampling the information set to be sampled based on the new comprehensive expected behavior index of each information in the information set to be sampled. By the embodiment, recommendation scenes with multiple attributes can be considered, effective recommendation of information is achieved, the recommendation system can be reused in different recommendation scenes, recommendation efficiency is improved, and background computing resources occupied by the recommendation system are saved.

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

The artificial intelligence based recommendation method provided by the embodiment of the application can be applied to various recommendation scenes, such as e-commerce recommendation scenes and news recommendation scenes, and is described below by taking a news recommendation system as an example, the artificial intelligence based recommendation method provided by the embodiment of the application is different from the recommendation method based on interest analogy adopted in the related art for recalling, the artificial intelligence based recommendation method provided by the embodiment of the application respectively calculates comprehensive expected behavior indexes aiming at information still in an information pool, establishes an inverted index (a first inverted index table) at a tag granularity according to the comprehensive expected behavior indexes, considers the click rate of the information and the freshness degree (exposure times) of the information by the comprehensive expected behavior indexes, is beneficial to 8 effectively exploring low-exposure information while utilizing effective behaviors, such as user click behaviors and fed back data, different from a candidate information interception scheme in the related technology, in the recommendation method based on artificial intelligence provided by the embodiment of the application, information is pulled from a Redis database based on a label of a user portrait, and final information interception is performed on all pulled information by taking a comprehensive expected behavior index as a sampling probability, and because the comprehensive expected behavior index is taken as the probability, information click rate and cold start information exploration are both considered, in the recommendation method based on artificial intelligence provided by the embodiment of the application, the weight of the portrait is prevented from being applied to a single piece of information, so that the information of head interest and tail interest of a user, which correspond to close click rate and exposure, has similar probability and is sampled as recall information, so that the tail information of the user has higher exposure, and the recommendation probability of the information of the tail interest of the user is increased, the method has the advantages that diversified interests of users are mined, so that user experience is improved, probability sampling is carried out based on comprehensive expected behavior indexes, so that each piece of information has exposure probability, the information recommended each time has certain difference, diversity of recommended information is further improved, and finally recalled information is recommended to the users after being processed by the sequencing module and the reordering module.

Referring to fig. 4, fig. 4 is a system architecture diagram of a recommendation method based on artificial intelligence provided in an embodiment of the present application, where the news recommendation system includes 4 main modules: the system comprises a user portrait module, a recall module, a sorting module and a scattering module, wherein the user portrait module calculates long-term interest and short-term interest of a user based on historical behaviors of the user so as to provide basic information for the recall module and the sorting module, the recall module excavates information interesting to the user from mass information at different angles, the sorting module predicts click rate of recalled information based on the user information and the information basic information, in order to avoid that the information content recommended to the user is too single, the scattering module collects the information after passing through the sorting module according to preset limiting rules and finally presents the information to the user, the user browses the information recommended by a client, clicks and reads thumbnails of seen information titles and information insets according to own interests, wherein a recall algorithm used by the recall module is a filter from an information pool to potential information of interest of the user, basic data are provided for a subsequent sorting module and a subsequent scattering module, and an inaccurate recall algorithm can lead recommended information not to contain potential interest of a user, seriously limit the final recommendation effect and finally lead to poor user experience.

The recall algorithm adopted by the recall module in the related art is based on an interest-class algorithm, corresponding information is pulled from an information reverse index based on a user portrait accumulated by user historical behaviors, the interest-class algorithm is different from other types of recall algorithms, the interest-class algorithm has the advantages of high accuracy, stable representation of user interests, strong interpretability and the like because of using the accumulated historical behaviors of users, a large recall amount is usually set for interest-class recall in the recall module to enter the ranking module, but the establishment of interest-based recall strategy reverse in the related art mainly depends on information publishing time or information click rate, so that some information occupies a reverse head part for a long time and is not beneficial to cold start of the information and distribution of cold door information, the recall strategy mainly performs reverse pull based on the head interests of the users, which causes medium-tail recall shortage of the users, the interception algorithm for the recall information mainly depends on the inverted score or the score of the information and the matching degree (the score of the matching degree between the information and the user portrait), which causes that some information which is not really interesting to the user occupies the recall head for a long time and cannot generate the actual recommendation effect, thereby not only wasting the calculation resources of the sequencing module and the scattering module, but also not well mining the potential interest of the user and seriously influencing the experience effect of the user.

In order to solve the problems, an embodiment of the present application provides a recommendation method based on artificial intelligence, a processing flow in the entire recommendation system is as shown in fig. 4, and the recommendation method is mainly applied to a recall module, a first inverted index table of tag granularity of information is established by integrating expected behavior indexes, the first inverted index table is stored offline, when a user requests information (recommendation request), the recall module pulls information of a tag corresponding to a user portrait in the first inverted index table and performs sampling selection, the sampled information is recalled information, click rate prediction is performed by a sorting module, information collection and position distribution are performed by a scattering module, and finally information is recommended to the user, the user browses and clicks and reads according to own preference, real click data is reported to the recommendation system to perform algorithm update and data accumulation of different modules, thereby continuously improving the user experience.

Referring to fig. 5A, fig. 5A is a data flow diagram of a recommendation method based on artificial intelligence provided in an embodiment of the present application, in an offline portion, first obtain information still in an information pool, then pull a current exposure amount and an accumulated click rate corresponding to the information from a database, calculate a comprehensive expected behavior index of each information by using formula (3), then establish information inversion (a first inverted index table) at a tag granularity, and store the information inversion in a redis database with a tag as a key for standby. Wilson (PV, Clk) is a section of historical average profit based on the number of exposures and the number of clicks for certain information, and usually takes the lower bound of the section. T is the number of times all information is exposed, PV is the number of times corresponding information is exposed, and α and β are hyper-parameters.

Referring to fig. 5B, fig. 5B is a data flow diagram of a recommendation method based on artificial intelligence provided in an embodiment of the present application, in an online computing portion, when a user requests (recall request), a recall module first pulls a label in a historical accumulated portrait of the user (obtained by computation based on information clicked in history), then pulls corresponding information from a redis database to be inverted based on the label in the portrait, in order to intercept information of the number of requests, a sampling algorithm without playback is adopted to sample by using an overall expected behavior index of the information, and finally, the information obtained by sampling is returned to the user.

According to the central limit theorem, if the experiment times of one piece of information are enough, the average income represents the real income (the click rate of the information is used as the historical average income), and the comprehensive expected behavior index of each piece of information can be calculated by using a formula (4).

First, the statistical average profit per information is used instead of the real profit, in equation (4)

Represents the historical average benefit (expected behavior characteristics) of the information, and passes

Representing the upper uncertainty bound of the historical average benefit (uncertainty characteristic of the desired behavior characteristic), T being the number of times all information has been exposed, and PV being the number of times the corresponding information has been exposed, it can be seen that when this information has been explored (exposed) many times, his benefit is equal to the historical average benefit, when the number of explorations (exposures) is low, the upper uncertainty bound of the historical average benefit will be greater, and there should be a greater probability to explore the information.

In the artificial intelligence based recommendation method provided by the embodiment of the application, a comprehensive expected behavior index is calculated for each piece of information, the click rate of the information is used as the historical average income, in order to balance the historical average income and the uncertainty income, two super parameters are used for controlling the proportion of the two super parameters, through experiments, 0.9 and 0.1 are used as parameters of the historical average income and the uncertainty income respectively, the click rate is very inaccurate at the beginning, the confidence coefficients of 10 clicks in 100 times of exposure and 1000 clicks in 10000 times of exposure are very large, so that in the artificial intelligence based recommendation method provided by the embodiment of the application, the concept of a Wilson (Wils on) interval is introduced, the meaning of the Wilson interval refers to the real click rate range under a certain confidence coefficient, and a Wi lson correction formula is shown as a formula (5);

wherein p is the probability of click, namely the click rate; n is the total number of samples, namely the exposure times of each piece of information; z is the standard layer of normal fractal, e.g., z is 1.96, and click-through has 95% confidence.

Then, an inverted information is established based on the comprehensive expected behavior index, the inverted key is a label, the value is the corresponding information and the comprehensive expected behavior index, as shown in table 1, the key is Tag id, for example, Tag: 5844784, the value is the corresponding information list, the information is accompanied with the respective comprehensive expected behavior index, the ordering is arranged according to the descending order of the comprehensive expected behavior index, such as A5, A6 and A7, the first 50 information of the comprehensive expected behavior index of each label is intercepted as the inverted information of the label, which is beneficial to saving the storage space and discarding the inferior information, and the inverted information is stored in the reds database for standby in order to enhance the efficiency of online recall.

TABLE 1 first inverted index Table

When a user requests data (recommendation request), the portrait data of the user is firstly pulled, and the portrait label of the user is analyzed, which represents the historical interest accumulated by the user. And then connecting a redis database, pulling corresponding inverted information, wherein the quantity of the pulled information is generally higher than that of the information requested at this time, sampling the pulled information by adopting a non-replacement sampling algorithm, and directly adopting a comprehensive expected behavior index as a sampling probability, so that the weight of the portrait of the user can be not considered, the information of the middle and tail interests of the user is exposed with a higher probability, the information distribution of the small and medium interests of the user is facilitated, and the development of new interests of the user is facilitated.

The recalled information is sent to a sorting module, the sorting module predicts the click rate of each information according to the user characteristics, the information characteristics and the context information (current time and the like), the information sorted based on the click rate is sent to a scattering module, the scattering module selects the information and distributes the position according to the information such as the classification of the information and the media type (video, image and text and the like), and finally the information is transmitted to a user client for display.

In the artificial intelligence based recommendation method provided by the embodiment of the application, the historical average profit and uncertainty of each piece of information are used as the upper bound of the prediction profit (comprehensive expected behavior index) of a single piece of information, which piece of information to explore at this time is determined based on the prediction profit, the uncertainty profit is continuously reduced with the increase of the number of experiments, and finally the uncertainty profit converges to the true value, that is, the average historical profit is based on the law of majority, and when the number of experiments is enough, the average profit converges to the true profit, and the artificial intelligence based recommendation method provided by the embodiment of the application has the following main processes: firstly, pulling all current information which is still time-efficient, acquiring historical accumulated click rate and historical exposure times of corresponding information, calculating comprehensive expected behavior indexes of each information based on a comprehensive expected behavior index calculation formula, then establishing label granularity inverted arrangement by using the comprehensive expected behavior indexes, storing information identifications and corresponding comprehensive expected behavior indexes in the inverted arrangement, storing inverted information in a redis database, updating the inverted arrangement once in an off-line updating mode for half an hour, pulling images of label granularity of a user when the user requests to recall, and then acquiring information corresponding to the labels from the redis database as candidate information recalled at this time; the candidate is sampled to obtain a plurality of pieces of information of the request as recall data, the sampling algorithm takes the comprehensive expected behavior index as the probability of sampling to mine more long-tail information, and the recall information is finally presented to the user through the sorting module and the scattering module.

The artificial intelligence-based recommendation method is applied to a news recommendation system, inverted arrangement is built on the label granularity by using comprehensive expected behavior indexes of information, the inverted arrangement is pulled on line based on portrait labels, then sampling and returning are performed by adopting a non-replacement sampling algorithm, the comprehensive expected behavior indexes are used as a sequencing basis for inverted arrangement, cold start of information and full utilization of historical exposure information are facilitated, the sampling algorithm is adopted, distribution of information of middle and long tail interests and diversity of information distributed at each time are facilitated, recalled information is finally presented to a user through a sequencing module and a rearrangement module, the recalled information coverage range is wider, information of interest of the user is widely mined, and user experience is improved.

Continuing with the exemplary structure of the artificial intelligence based recommendation device 255 provided by the embodiments of the present application as software modules, in some embodiments, as shown in fig. 2, the software modules stored in the artificial intelligence based recommendation device 255 of the memory 250 may include: an obtaining module 2551, configured to obtain the number of exposures and the number of effective behaviors of each piece of information in the information set; a determining module 2552, configured to determine an expected behavior feature of the corresponding information based on the exposure times and the effective behavior times, and determine an uncertainty feature of the expected behavior feature based on the exposure times; an establishing module 2553, configured to establish a first inverted index table including a comprehensive expected behavior index of the information based on the expected behavior feature and the uncertainty feature of each information; and the recommending module 2554 is configured to sample the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table, and perform a recommending operation based on a sampling processing result.

In some embodiments, before obtaining the number of exposures and the number of valid actions for each information in the set of information, the obtaining module 2551 is further configured to: traversing the release time of each information in the information set according to a fixed time interval; and when the release time of the information is less than the release time threshold, removing the corresponding information from the information set.

In some embodiments, determination module 2552 is further configured to: determining an expected behavior feature of the corresponding information based on the following expected behavior feature functions:

In some embodiments, determination module 2552 is further configured to: determining an uncertainty feature corresponding to the expected behavior feature based on the following uncertainty feature functions:

In some embodiments, the establishing module 2553 is further configured to: determining a first behavioral component that is positively correlated with the desired behavioral characteristic; determining a second behavior component that is positively correlated with the uncertainty feature; performing fusion processing on the first behavior component and the second behavior component to obtain a comprehensive expected behavior index of the information; and establishing a first inverted index table by taking the label of the information as a key and taking the information and the comprehensive expected behavior index of the corresponding information as values.

In some embodiments, recommendation module 2554 is further configured to: obtaining a portrait label in a user portrait from a recommendation request of a client; acquiring a plurality of information corresponding to the label consistent with the image label from the first inverted index table to form an information set to be sampled; and sampling the information set to be sampled based on the comprehensive expected behavior index of each information in the information set to be sampled.

In some embodiments, recommendation module 2554 is further configured to: performing sampling processing with the sampling frequency threshold, and performing the following processing in the process of each sampling processing: converting the comprehensive expected behavior index of each information in the information set to be sampled into a subinterval in a numerical interval; generating random numbers corresponding to the numerical value intervals through a random function; and determining the information corresponding to the subinterval in which the random number falls as the information obtained by sampling, and moving out of the information set to be sampled.

In some embodiments, recommendation module 2554 is further configured to: based on the scene to which the recommendation request of the client belongs, adapting a plurality of candidate inverted index tables; when the first inverted index table is adapted to the multiple candidate inverted index tables, the information in the first inverted index table is sampled and processed based on the comprehensive expected behavior index of the information in the first inverted index table.

In some embodiments, the plurality of candidate inverted index tables further comprises a second inverted index table, and the second inverted index table comprises a descending order based on the information creation time; a recommendation module 2554, further configured to: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with aging attribute, determining the second inverted index table as the inverted index table adapted to the recommendation request.

In some embodiments, recommendation module 2554 is further configured to: when a second inverted index table is adapted to the multiple candidate inverted index tables, updating the release time of multiple pieces of information which are ranked in the second inverted index table in the front; wherein, the updating value used for updating the release time and the information posterior quality parameter of the information form a positive correlation relationship; and acquiring a plurality of pieces of information which are ranked at the top from the updated second inverted index table to execute the recommendation operation.

In some embodiments, the plurality of candidate inverted index tables further comprises a third inverted index table, and the third inverted index table comprises a descending ordering based on the information a posteriori quality parameter; a recommendation module 2554, further configured to: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with a quality attribute, determining the third inverted index table as the inverted index table adapted to the recommendation request.

In some embodiments, recommendation module 2554 is further configured to: when a third inverted index table is adapted to the multiple candidate inverted index tables, updating the posterior quality parameters of the information of the multiple pieces of information which are ranked in the third inverted index table in the front; wherein, the updating value used for updating the posterior quality parameter of the information is negatively correlated with the exposure times of the information; and acquiring a plurality of pieces of information which are ranked at the top from the updated third inverted index table to execute the recommendation operation.

In some embodiments, the plurality of candidate inverted index tables further includes a fourth inverted index table, and the fourth inverted index table includes a descending ordering based on portrait similarity; a recommendation module 2554, further configured to: and when the recommendation scene corresponding to the recommendation request is a recommendation scene with the interest attribute, determining the fourth inverted index table as the inverted index table adapted to the recommendation request.

In some embodiments, recommendation module 2554 is further configured to: when a fourth inverted index table is adapted to the multiple candidate inverted index tables, updating the information portrait similarity of multiple pieces of information ranked backwards in the fourth inverted index table; wherein, the updating value used for updating the image similarity is in negative correlation with the sorting position of the information; and acquiring a plurality of pieces of information which are ranked at the top from the updated fourth inverted index table to execute the recommendation operation.

In some embodiments, recommendation module 2554 is further configured to: acquiring real-time data of a scene to extract real-time characteristics of the scene; calling a neural network model based on the scene characteristics to predict the attributes of the scene; the training sample of the neural network model comprises front-end scene data, and the marking data of the training sample comprises attributes of scenes corresponding to the front-end scene data; an inverted index table that is adapted to the predicted attribute is determined from the plurality of candidate inverted index tables.

In some embodiments, recommendation module 2554 is further configured to: obtaining portrait tags from user portraits carried by a recommendation request of a client, and performing adaptation processing on a plurality of candidate inverted index tables based on a scene to which the recommendation request belongs to obtain at least one candidate inverted index table; the candidate reverse index table is a reverse index table different from the first reverse index table in the multiple candidate reverse index tables; acquiring information corresponding to the label consistent with the image label from the first inverted index table to obtain an information set to be sampled; acquiring the corresponding fraction of each information in the information set to be sampled in each candidate inverted index table; fusing the comprehensive expected behavior index of each information in the information set to be sampled with the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index; and sampling the information set to be sampled based on the new comprehensive expected behavior index of each information in the information set to be sampled.

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 recommendation method according to the embodiment of the present application.

Embodiments of the present application provide a computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform artificial intelligence based recommendation methods provided by embodiments of the present application, for example, artificial intelligence based recommendation methods as illustrated 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, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, and 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).

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

In summary, according to the exposure times and the effective behavior times of the information in the embodiment of the application, the expected behavior characteristics for user behavior prediction and the uncertainty characteristics corresponding to the expected behavior characteristics are specifically depicted for different information, the contribution of the information with different exposure times (such as cold start information and historical exposure information with different degrees) to the user behavior prediction is considered, and meanwhile, through sampling processing, the information distribution and the diversity of information distribution of middle and long tail interests are facilitated, so that the coverage range of the sampled information is wide, the information of interest of the user can be better mined, the information recommendation precision of subsequent information recommendation is ensured, meanwhile, invalid recommendation is effectively avoided, and further, the calculation resources related to recommendation logic in the server are saved.

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 artificial intelligence based recommendation method, characterized in that the method comprises:

2. The method of claim 1, wherein prior to obtaining the number of exposures and the number of active actions for each information in the set of information, the method further comprises:

traversing the release time of each information in the information set according to a fixed time interval;

and when the release time of the information is less than the release time threshold, removing the corresponding information from the information set.

3. The method of claim 1, wherein said building a first inverted index table comprising a composite expected behavior index for each of said information based on said expected behavior signature and said uncertainty signature of said information comprises:

determining a first behavioral component that is positively correlated with the desired behavioral characteristic;

determining a second behavior component that is positively correlated with the uncertainty feature;

performing fusion processing on the first behavior component and the second behavior component to obtain a comprehensive expected behavior index of the information;

and establishing the first inverted index table by taking the label of the information as a key and taking the information and the comprehensive expected behavior index corresponding to the information as values.

4. The method according to claim 1, wherein the sampling the information in the first inverted index table based on the integrated expected behavior index of the information in the first inverted index table comprises:

obtaining a portrait label in a user portrait from a recommendation request of a client;

acquiring a plurality of information corresponding to the label consistent with the image label from the first inverted index table to form an information set to be sampled;

and sampling the information set to be sampled based on the comprehensive expected behavior index of each information in the information set to be sampled.

5. The method according to claim 4, wherein the sampling the information based on the comprehensive expected behavior index of each information in the information set to be sampled comprises:

performing sampling processing with the sampling frequency threshold, and performing the following processing in the process of each sampling processing:

converting the comprehensive expected behavior index of each information in the information set to be sampled into a subinterval in a numerical interval;

generating a random number corresponding to the numerical value interval through a random function;

and determining the information corresponding to the subinterval in which the random number falls as the information obtained by sampling, and moving out of the information set to be sampled.

6. The method according to claim 1, wherein the sampling the information in the first inverted index table based on the integrated expected behavior index of the information in the first inverted index table comprises:

based on the scene to which the recommendation request of the client belongs, adapting a plurality of candidate inverted index tables;

and when the first inverted index table is adapted to the multiple candidate inverted index tables, sampling the information in the first inverted index table based on the comprehensive expected behavior index of the information in the first inverted index table.

7. The method of claim 6,

the plurality of candidate inverted index tables further comprises a second inverted index table, and the second inverted index table comprises a descending order based on information creation time;

the adaptation of the multiple candidate inverted index tables is performed based on the scene to which the recommendation request of the client belongs, and comprises the following steps:

and when the recommended scene corresponding to the recommendation request is a recommended scene with aging attribute, determining the second inverted index table as an inverted index table adapted to the recommendation request.

8. The method of claim 7, further comprising:

when the second reverse index table is adapted to the multiple candidate reverse index tables, updating the release time of multiple pieces of information ranked at the top in the second reverse index table;

wherein, the update value used for updating the release time and the information posterior quality parameter of the information form a positive correlation relationship;

and acquiring a plurality of pieces of information in the top order from the updated second inverted index table to execute recommendation operation.

9. The method of claim 6,

the plurality of candidate inverted index tables further comprise a third inverted index table, and the third inverted index table comprises a descending order based on the information posterior quality parameter;

and when the recommendation scene corresponding to the recommendation request is a recommendation scene with a quality attribute, determining the third inverted index table as an inverted index table adapted to the recommendation request.

10. The method of claim 9, further comprising:

when the third reverse index table is adapted to the multiple candidate reverse index tables, updating the posterior quality parameters of the information of the multiple pieces of information which are ranked at the front in the third reverse index table;

wherein an update value for updating the information posterior quality parameter is negatively correlated with the number of exposures of the information;

and acquiring a plurality of pieces of information in the top order from the updated third inverted index table to execute recommendation operation.

11. The method of claim 6,

the plurality of candidate inverted index tables further comprises a fourth inverted index table, and the fourth inverted index table comprises descending order based on portrait similarity;

and when the recommendation scene corresponding to the recommendation request is a recommendation scene with interest attributes, determining the fourth inverted index table as an inverted index table adapted to the recommendation request.

12. The method of claim 11, further comprising:

when the fourth reverse index table is adapted to the multiple candidate reverse index tables, updating information portrait similarity of multiple pieces of information ranked in the fourth reverse index table;

wherein, the updating value used for updating the image similarity is in negative correlation with the sorting position of the information;

and acquiring a plurality of pieces of information in the top order from the updated fourth inverted index table to execute recommendation operation.

13. An artificial intelligence based recommendation apparatus, the apparatus comprising:

14. An electronic device, comprising:

a memory for storing executable instructions;

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

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