Artificial intelligence-based recommendation methods, devices, and electronic equipment

Abstract

This application provides a recommendation method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence. The method includes: obtaining the exposure count and effective behavior count of each piece of information in an information set; determining the expected behavioral characteristics of the corresponding information based on the exposure count and effective behavior count, and determining the uncertainty characteristics of the expected behavioral characteristics based on the exposure count; establishing a first inverted index table including a comprehensive expected behavioral index for each piece of information based on the expected behavioral characteristics and uncertainty characteristics; sampling the information in the first inverted index table based on the comprehensive expected behavioral index of the information in the first inverted index table, and performing a recommendation operation based on the sampling results. This application can improve recommendation accuracy.

Description

Technical Field

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

[0002] Artificial intelligence (AI) is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0003] Information recommendation is an important application of artificial intelligence. The recall phase in a recommendation system is the foundation of the entire system. Various efforts have been made in related technologies to recall effective information as the ranking object of the recommendation system during the recall phase, such as using interest-based recall. However, in implementing the embodiments of this application, the applicant found that directly recalling information that matches the user profile from the inverted index is difficult to effectively characterize the positive impact of cold start information on user behavior and the diverse interests of users, thus affecting the accuracy of information recommendation. Summary of the Invention

[0004] This application provides an artificial intelligence-based recommendation method, apparatus, electronic device, and computer-readable storage medium, which can improve recommendation accuracy.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides an artificial intelligence-based recommendation method, including:

[0007] Get the number of exposures and the number of valid actions for each piece of information in the information set;

[0008] Based on the number of exposures and the number of valid actions, the expected behavioral characteristics corresponding to the information are determined, and the uncertainty characteristics of the expected behavioral characteristics are determined based on the number of exposures.

[0009] Based on the expected behavior characteristics and uncertainty characteristics of each piece of information, a first inverted index table is established that includes a comprehensive expected behavior index of the information.

[0010] 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 processed, and recommendation operations are performed based on the sampling and processing results.

[0011] In the above scheme, the adaptation of multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs includes:

[0012] Acquire real-time data of the scene to extract real-time features of the scene;

[0013] Based on the scene features, a neural network model is invoked to predict the attributes of the scene;

[0014] The training samples of the neural network model include front-end scene data, and the labeled data of the training samples include the attributes of the scene corresponding to the front-end scene data.

[0015] From the plurality of candidate inverted index tables, determine the inverted index table that matches the predicted attributes.

[0016] In the above scheme, the comprehensive expected behavior index based on the information in the first inverted index table, which samples and processes the information in the first inverted index table, includes:

[0017] Get profile tags from the user profile carried in the client's recommendation request, and perform adaptation processing on multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs, to obtain at least one candidate inverted index table;

[0018] Wherein, the candidate inverted index table is an inverted index table that is different from the first inverted index table among the plurality of candidate inverted index tables;

[0019] Retrieve information corresponding to tags that match the portrait tags from the first inverted index table to obtain a set of information to be sampled;

[0020] Obtain the corresponding score for each piece of information in the set of information to be sampled in each of the candidate inverted index tables;

[0021] The comprehensive expected behavior index of each piece of information in the set of information to be sampled is fused with the corresponding score in each of the other inverted index tables to obtain a new comprehensive expected behavior index.

[0022] Based on the new comprehensive expected behavior index of each piece of information in the set of information to be sampled, the set of information to be sampled is sampled.

[0023] This application provides an artificial intelligence-based recommendation device, including:

[0024] The acquisition module is used to acquire the number of exposures and the number of valid actions for each piece of information in the information set;

[0025] The determination module is used to determine the expected behavioral characteristics corresponding to the information based on the number of exposures and the number of effective behaviors, and to determine the uncertainty characteristics of the expected behavioral characteristics based on the number of exposures.

[0026] A module is established to create a first inverted index table of comprehensive expected behavior indicators that includes the information, based on the expected behavior characteristics and the uncertainty characteristics of each piece of information.

[0027] The recommendation module is used 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 recommendation operations based on the sampling results.

[0028] In the above scheme, before obtaining the exposure count and effective behavior count of each piece of information in the information set, the acquisition module is further configured to: traverse the publication time of each piece of information in the information set at fixed time intervals; when the publication time of the information is less than the publication time threshold, remove the corresponding information from the information set.

[0029] In the above scheme, the determining module is further configured to:

[0030] The expected behavior features corresponding to the information are determined based on the following expected behavior feature function:

[0031]

[0032] Wherein, E1 is the desired behavioral feature, p is the ratio between the number of effective behaviors and the number of exposures, z is the confidence parameter, and n is the number of times the information is exposed.

[0033] In the above scheme, the determining module is further configured to:

[0034] The uncertainty features corresponding to the desired behavior features are determined based on the following uncertainty feature function:

[0035]

[0036] Where E2 is the uncertainty feature of the expected behavior feature, T is the sum of the exposure times of all information in the information set, and n is the exposure times of the information.

[0037] In the above scheme, the establishment module is further configured to: determine a first behavioral component that is positively correlated with the expected behavioral feature; determine a second behavioral component that is positively correlated with the uncertainty feature; perform fusion processing on the first behavioral component and the second behavioral component to obtain a comprehensive expected behavioral index of the information; and establish a first inverted index table using the label of the information as the key and the information and the corresponding comprehensive expected behavioral index of the information as the value.

[0038] In the above scheme, the recommendation module is further configured to: obtain profile tags from the user profile in the recommendation request of the client; obtain multiple pieces of information corresponding to the tags that are consistent with the profile tags from the first inverted index table to form a set of information to be sampled; and perform sampling processing on the set of information to be sampled based on the comprehensive expected behavior index of each piece of information in the set of information to be sampled.

[0039] In the above scheme, the recommendation module is further configured to: perform sampling processing a number of times equal to the sampling number threshold, and perform the following processing in each sampling process: convert the comprehensive expected behavior index of each piece of information in the set of information to be sampled into a sub-interval within the numerical range; generate a random number corresponding to the numerical range through a random function; determine the information corresponding to the sub-interval in which the random number falls as the sampled information, and remove it from the set of information to be sampled.

[0040] In the above scheme, the recommendation module is further configured to: adapt multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs; when the first inverted index table is adapted from the multiple candidate inverted index tables, 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.

[0041] In the above scheme, the plurality of candidate inverted index tables also 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: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with time-sensitive attributes, determine the second inverted index table as the inverted index table adapted to the recommendation request.

[0042] In the above scheme, the recommendation module is further configured to: when the second inverted index table is adapted from the plurality of candidate inverted index tables, update the publication time of the plurality of information items ranked first in the second inverted index table; wherein the update value used for updating the publication time is positively correlated with the information posterior quality parameter of the information; and obtain the plurality of information items ranked first from the updated second inverted index table to perform the recommendation operation.

[0043] In the above scheme, the plurality of candidate inverted index tables also include a third inverted index table, and the third inverted index table includes a descending order based on the posterior quality parameter of the information; the recommendation module is further configured to: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with quality attributes, determine the third inverted index table as the inverted index table adapted to the recommendation request.

[0044] In the above scheme, the recommendation module is further configured to: when the third inverted index table is adapted from the multiple candidate inverted index tables, update the posterior quality parameters of the multiple pieces of information ranked at the top of the third inverted index table; wherein the update value used for updating the posterior quality parameters is negatively correlated with the exposure count of the information; and obtain the multiple pieces of information ranked at the top from the updated third inverted index table to perform the recommendation operation.

[0045] In the above scheme, the plurality of candidate inverted index tables also include a fourth inverted index table, and the fourth inverted index table includes a descending order based on image similarity; the recommendation module is further configured to: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with interest attributes, determine the fourth inverted index table as the inverted index table adapted to the recommendation request.

[0046] In the above scheme, the recommendation module is further configured to: when the fourth inverted index table is adapted from the multiple candidate inverted index tables, update the information profile similarity of multiple pieces of information ranked lower in the fourth inverted index table; wherein the update value used for profile similarity update is negatively correlated with the ranking position of the information; and obtain multiple pieces of information ranked higher from the updated fourth inverted index table to perform recommendation operation.

[0047] In the above scheme, the recommendation module is further configured to: acquire real-time data of the scene to extract real-time features of the scene; call a neural network model based on the scene features to predict the attributes of the scene; wherein the training samples of the neural network model include front-end scene data, and the labeled data of the training samples include the attributes of the scene corresponding to the front-end scene data; and determine an inverted index table that matches the predicted attributes from the plurality of candidate inverted index tables.

[0048] In the above scheme, the recommendation module is further configured to: obtain profile tags from the user profile carried in the client's recommendation request, and perform adaptation processing on multiple candidate inverted index tables based on the scenario to which the client's recommendation request 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 among the multiple candidate inverted index tables; obtain information corresponding to tags consistent with the profile tags from the first inverted index table to obtain a set of information to be sampled; obtain the corresponding score of each piece of information in the set of information to be sampled in each candidate inverted index table; fuse the comprehensive expected behavior index of each piece of information in the set of information to be sampled with the corresponding score in each of the other inverted index tables to obtain a new comprehensive expected behavior index; and perform sampling processing on the set of information to be sampled based on the new comprehensive expected behavior index of each piece of information in the set of information to be sampled.

[0049] This application provides an electronic device, including:

[0050] Memory, used to store executable instructions;

[0051] When the processor executes the executable instructions stored in the memory, it implements the AI-based recommended method provided in the embodiments of this application.

[0052] This application provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the AI-based recommendation method provided in this application.

[0053] The embodiments of this application have the following beneficial effects:

[0054] Based on the number of times information is exposed and the number of effective actions, the expected behavioral features and the corresponding uncertainty features of the expected behavioral features are characterized in a targeted manner for different information. This takes into account the contribution of information with different exposure times (such as cold start information and historical exposure information of different degrees) to user behavior prediction. At the same time, through sampling processing, it is beneficial to the distribution of information with mid-to-long-tail interests and the diversity of information distribution. As a result, the information obtained by sampling has a wide coverage and can better discover information that users are interested in, which ensures the accuracy of subsequent information recommendation and effectively avoids invalid recommendations, thereby saving the computing resources related to recommendation logic on the server. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the architecture of an artificial intelligence-based recommendation system provided in an embodiment of this application;

[0056] Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application;

[0057] Figures 3A-3D This is a flowchart illustrating the AI-based recommendation method provided in an embodiment of this application;

[0058] Figure 4 This is an overall architecture diagram of the AI-based recommendation method provided in the embodiments of this application;

[0059] Figures 5A-5B This is a data flow diagram of the AI-based recommendation method provided in the embodiments of this application. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0062] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0064] In the implementation of this application, the collection and processing of relevant data should be strictly in accordance with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0065] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0066] 1) Personalized recommendation: Using the principles and methods of recommendation systems, we quantitatively mine and describe users' existing and potential interests, and then recommend information that users are interested in.

[0067] 2) Inverted index: An inverted index is used to find records based on the value of an attribute. For example, in an article system, each article in the forward index corresponds to multiple attributes, such as title, tags, and publisher. An inverted index, on the other hand, uses the attribute (the tag or title corresponding to the article) as the query condition. For example, it can query articles under a certain tag.

[0068] 3) Expected behavioral features: Expected behavioral features are historical average revenue values. For example, in a recommendation system, this can be the historical click-through rate (historical average revenue value) of any information. The historical click-through rate can predict the actual click-through rate of information to a certain extent.

[0069] 4) Uncertainty features: Uncertainty features are the upper bound values ​​that characterize the uncertainty of historical average revenue values. For example, in a recommendation system, some information may have a low number of exposures, so its historical click-through rate may not accurately predict the actual click-through rate of the information. Therefore, the desired behavior features can be corrected by using uncertainty features.

[0070] 5) Comprehensive Expected Behavior Index: The comprehensive expected behavior index is a value calculated by weighting expected behavior characteristics and uncertainty characteristics. For example, in a recommendation system, it can be used to predict the actual click-through rate of information with different exposure levels.

[0071] In related technologies, with the rapid development of information technology, all kinds of information are growing explosively. Users cannot obtain information that is truly meaningful to them from the massive amount of information. Therefore, personalized information recommendation has become an important research topic. Recommendation systems generally include a profile module, a recall module, a ranking module, and a scattering module. The recall module is an important component of the recommendation system and can usually adopt collaborative filtering recall, geographic recall, interest-based recall, and operational recall, etc.

[0072] In the recall module of a recommendation system, user interest-based recall is an important recall strategy. It first builds an inverted index for all information in the information pool, which generally includes inverted indexes of primary categories, secondary categories, and tags. Then, it uses the user's historical click behavior to build a cumulative profile for the user. When the user requests information (recommendation request), interest-based recall first pulls the user profile, and then pulls the corresponding information from the inverted index based on the profile's tags. After extracting multiple articles from the pulled information, it goes through the sorting module and the re-sorting module, and finally recommends it to the user.

[0073] Interest-based recall first obtains the user's interests from the user profile, then pulls the corresponding information from the inverted index, and then extracts K pieces of information, sorts and shuffles them before finally presenting them to the user. The establishment of the inverted index is the key to this type of recall. Currently, the main considerations are the information creation time and the actual click-through rate of the information. Furthermore, the algorithm for extracting information will affect the quality and diversity of the recommended information. Usually, it is based on sorting and extracting based on the inverted index score or sorting and extracting based on the matching score with the user profile.

[0074] In recommendation systems, the index side builds an inverted list based on information creation time or information popularity. Interest-based retrieval typically uses the user's top interests as the basis for pulling the inverted list. The pulled inverted list is then truncated at the top based on the inverted list score or the information's matching score with the user. Therefore, the following technical problems may arise:

[0075] 1. If we consider the creation time of information to build an inverted index, we only use the publication time of the information and do not take advantage of the posterior quality of the information. In other words, we do not take advantage of the user's posterior feedback on the information after it is published. This is not conducive to the distribution of high-quality information, and affects the accuracy of information recommendation, thus affecting the user experience.

[0076] 2. If the inverted index is built based on the click-through rate of the information, only the posterior quality of the information is used. Newly published information has fewer exposures and lower information quality score confidence, which is not conducive to the distribution of information during the cold start. In addition, information with high click-through rates in the past is always at the head of the inverted index, resulting in the information recommended to users being all global hot information, which affects the diversity of recommended information and further affects the user experience.

[0077] 3. If multiple pieces of information at the top of the ranking are extracted based on the inverted score, the recommended information will rely too much on the calculation method of the inverted score, and the information at the bottom of the inverted list will not be effectively exposed, which is not conducive to the diversity of recommended information.

[0078] 4. If the inverted information is scored based on user profiles and the top-ranked information is extracted, it is easily affected by the distribution of user profiles. Most users' top interests have a large weight, which makes the recommendation results concentrated on the top interests of users. The problem of single information in each recall still exists, which affects the recommendation of users' long-tail interests and the diversity of recall information.

[0079] This application provides an artificial intelligence-based recommendation method, apparatus, electronic device, and computer-readable storage medium, which can take into account the contribution of information from different exposure times (such as cold start information and historical exposure information of different degrees) to user behavior prediction, thereby improving recommendation accuracy. The following describes an exemplary application of the electronic device provided in this application. The electronic device provided in this application can be implemented as a server. The following will describe an exemplary application when the device is implemented as a server.

[0080] A server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, and big data and artificial intelligence platforms. Servers can be connected directly or indirectly through wired or wireless communication methods, and this application does not impose any restrictions on this.

[0081] Artificial intelligence cloud services are generally also known as AIaaS (AI as a Service). This is currently a mainstream service model for artificial intelligence platforms. Specifically, AIaaS platforms break down several common AI services and provide them as independent or packaged services in the cloud. This service model is similar to opening an AI-themed marketplace: all developers can access and use one or more AI services provided by the platform through API interfaces. Some experienced developers can also use the AI ​​framework and AI infrastructure provided by the platform to deploy and maintain their own dedicated cloud AI services. In the AI-based recommendation method provided in this application embodiment, the recommendation system can be deployed and maintained through the AI ​​framework and AI infrastructure provided by the AI ​​cloud service.

[0082] See Figure 1 , Figure 1This is a schematic diagram of the architecture of an AI-based recommendation system provided in this application embodiment. The recommendation system can be used to support various information recommendation scenarios, such as application scenarios for recommending news, products, and videos. Depending on the application scenario, the information can be news, video articles, images and text, or information related to products (such as physical items like clothing, and virtual items like game props). During the user's use of the client, the terminal 400 reports the collected user interaction behavior (number of effective behaviors) and exposure status (number of exposures) to the server 200. The server 200 can obtain information from the database 500 and, based on the number of exposures and the number of effective behaviors, establish a first inverted index table including a comprehensive expected behavior index for the information, and store it in the database 500. In response to the recommendation request from the terminal 400, the server 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 performs a recommendation operation based on the sampling processing result, recommending the information in the sampling processing result to the terminal 400.

[0083] The following describes the specific architecture of the recommendation system. In the recommendation system, terminal 400 connects to server 200 via network 300. Network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of both. The function of server 200 can be abstracted as establishing a first inverted index table and performing sampling processing based on the first inverted index table to obtain recalled information. Server 200 periodically maintains the first inverted index table in database 500 based on the interaction behavior (number of effective behaviors) and exposure status (number of exposures) reported by terminal 400. It receives recommendation requests from terminal 400. Based on the comprehensive expected behavior index of the information in the first inverted index table in database 500, server 200 samples the information in the first inverted index table and returns the information from the sampling processing results to terminal 400 for presentation.

[0084] See Figure 2 , Figure 2 This is a schematic diagram of the structure of a server 200 that applies an artificial intelligence-based recommendation method according to an embodiment of this application. Figure 2 The server 200 shown includes at least one processor 210, memory 250, and at least one network interface 220. The various components of server 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0085] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0086] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0087] Memory 250 may include volatile memory or non-volatile memory, or both. Non-volatile memory may be read-only memory (ROM), and volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0088] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0089] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0090] The network communication module 252 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220, such as Bluetooth, WiFi, and Universal Serial Bus (USB).

[0091] In some embodiments, the AI-based recommendation device provided in this application can be implemented in software. Figure 2 An AI-based recommendation device 255 stored in memory 250 is shown, which includes multiple modules. The modules can be software in the form of programs and plug-ins, including the following software modules: acquisition module 2551, determination module 2552, creation module 2553, and recommendation module 2554. These modules are logically related and can therefore be arbitrarily combined or further split according to the functions they implement. The functions of each module will be described below.

[0092] The AI-based recommendation method provided in this application will be described in conjunction with exemplary applications and implementations of the server provided in the embodiments of this application.

[0093] See Figure 3A , Figure 3A This is a flowchart illustrating the AI-based recommendation method provided in this application embodiment, which will be combined with... Figure 3A The steps shown are explained.

[0094] In step 101, the number of exposures and the number of valid actions for each piece of information in the information set are obtained.

[0095] In some embodiments, before obtaining the number of exposures and the number of valid actions for each piece of information in the information set, the following technical solution may also be performed: traversing the publication time of each piece of information in the information set at fixed time intervals; when the publication time of an information is less than the publication time threshold, the corresponding information is removed from the information set.

[0096] As an example, since the information set serves as the basis for information retrieval, it is necessary to ensure that the publication time of the information in the information set is within the publication time threshold. For example, if the current time is 8:00 AM on November 9, 2020, and the publication time threshold is 8:00 AM on November 6, 2020, if the publication time of the information is 7:00 AM on November 6, 2020, it indicates that the publication time of the information is less than the publication time threshold, and the information needs to be removed from the information set. The fixed time interval can be one day, half a day, or a more granular time. Through the above implementation method, it can be ensured that the information in the information set has timeliness, and that the information ultimately recommended to the user is all information that is meaningful in the time dimension.

[0097] In step 102, the expected behavioral characteristics of the corresponding information are determined based on the number of exposures and the number of valid behaviors, and the uncertainty characteristics of the expected behavioral characteristics are determined based on the number of exposures.

[0098] In some embodiments, determining the expected behavioral characteristics of the corresponding information based on the number of exposures and the number of valid behaviors in step 102 can be achieved through the following technical solution: determining the expected behavioral characteristics of the corresponding information based on the following expected behavioral characteristic function:

[0099]

[0100] Where E1 represents the desired behavioral feature, p represents the ratio between the number of effective behaviors and the number of exposures, z represents the confidence parameter, and n represents the number of times the information is exposed.

[0101] As an example, the number of valid behaviors can be click behaviors or more advanced user behaviors, such as user-information interaction behaviors, including forwarding behaviors, commenting behaviors, etc. z is the standard layer in the normal distribution. For example, z is 1.96, which has a 95% confidence level. The expected behavior features calculated by the above formula (1) can be understood as the historical average revenue of the recommendation system. The revenue in the recommendation system is usually an indicator related to user behavior, such as the historical average forwarding rate, the historical average collection rate, and the most common one is the historical average click rate.

[0102] In some embodiments, historical average click-through rate (CTR) can be directly used as the expected behavior feature, and then a comprehensive expected behavior index for each piece of information can be calculated based on the historical average CTR. However, during the implementation of this application's embodiments, the applicant found that simply using the historical average CTR as the expected behavior feature to predict the comprehensive expected behavior index at the initial recommendation stage is highly inaccurate. This is because CTR equals clicks / impressions, which only considers the ratio but not the sample size. With a small sample size, this ratio is inaccurate, while a larger sample size results in a more accurate ratio and better reflects the expected behavior. To reflect the true situation, for example, there are three ads: Ad A has 2 clicks and 10 impressions, Ad B has 20 clicks and 100 impressions, and Ad C has 200 clicks and 1000 impressions. Their click-through rates (CTRs) are all 0.2. However, it is clear from a confidence perspective that because Ad C has a larger sample size, its 0.2 CTR is more realistic and credible. To measure the impact of sample size on the confidence interval of CTR, the above formula (1) is introduced. Formula (1) is used to calculate the true CTR range at a certain confidence level (e.g., 95% confidence level). For ad A, with 2 clicks and 10 impressions, the 95% confidence interval for the click-through rate (CTR) is (0.07, 0.45). For ad B, with 20 clicks and 100 impressions, the 95% confidence interval for the CTR is (0.14, 0.27). For ad C, with 200 clicks and 1000 impressions, the 95% confidence interval for the CTR is (0.18, 0.22). In practical applications, the lowest value can be taken as the adjusted CTR. Therefore, the adjusted CTR for ad A is 0.07, and for ad B... The click-through rate after the correction was 0.14, and the click-through rate after the correction for ad C was 0.18. This is equivalent to applying a certain degree of attenuation and weighting to the click-through rate with insufficient sample size using the above formula (1). Through the above implementation method, it is ensured that even if the number of exposures is insufficient during the cold start process, a historical average click-through rate with high confidence (not the historical average click-through rate in statistics, but the corrected historical average click-through rate) can still be obtained as the expected behavior feature, thereby ensuring that the accurate comprehensive expected behavior index is obtained based on the expected behavior feature in the future, so as to improve the recommendation accuracy of the information obtained based on the comprehensive expected behavior index.

[0103] In some embodiments, the uncertainty feature of determining the desired behavioral feature based on the number of exposures in step 102 can be achieved by the following technical solution: determining the uncertainty feature corresponding to the desired behavioral feature based on the following uncertainty feature function:

[0104]

[0105] Where E2 represents the uncertainty of the expected behavioral characteristics, T represents the sum of the exposure times of all information in the information set, and n represents the exposure times of the information.

[0106] As an example, expected behavioral features can be understood as benefits derived from experience. During the recall phase, information with the highest expected behavioral features is selected for recall to obtain higher cumulative returns. However, since the returns are random, the probability estimate of the return for each piece of information is not accurate. The information with the highest actual return probability may not be the one with the highest expected behavioral feature. Therefore, it is necessary to explore less popular information. Exploring less popular information can bring uncertain benefits. The larger the value of the uncertainty feature of expected behavioral features, the larger the uncertain benefits brought by exploring less popular information. Randomly exploring different information can obtain a more accurate probability estimate of the return for each piece of information, thereby finding the truly optimal information. However, exploration requires recalling information whose current probability estimate of return is not high, which means that some opportunities to recall information with high historical average returns will be lost. Through the above implementation method, uncertain features can be obtained as a supplement to expected behavioral features, thereby increasing the chance of recalling information with low exposure or no exposure during the cold start process in actual recommendation scenarios, so as to explore the diverse interests of users.

[0107] As an example, the uncertainty feature of the expected behavior feature can be calculated based on the above formula (2). When n is small or T is relatively large, it indicates that the information is relatively unpopular. The less popular the information is, the smaller the expected behavior feature of the information is, and the larger the uncertainty feature of the expected behavior feature is. The expected behavior feature is modified to a large extent by the uncertainty feature of the expected behavior feature. When n is large or T is relatively small, it indicates that the information is relatively popular. The more popular the information is, the larger the expected behavior feature of the information is, and the smaller the uncertainty feature of the expected behavior feature is. That is, the expected behavior feature is modified to a small extent by the uncertainty feature of the expected behavior feature.

[0108] In step 103, based on the expected behavior characteristics and uncertainty characteristics of each piece of information, a first inverted index table including the comprehensive expected behavior index of the information is established.

[0109] In some embodiments, see Figure 3B , Figure 3B This is a flowchart illustrating the AI-based recommendation method provided in this application embodiment. In step 103, a first inverted index table is established based on the expected behavior characteristics and uncertainty characteristics of each piece of information. This can be achieved through steps 1031-1034.

[0110] In step 1031, a first behavioral component that is positively correlated with the desired behavioral characteristics is determined.

[0111] In step 1032, a second behavioral component that is positively correlated with the uncertainty feature is determined.

[0112] In step 1033, the first behavioral component and the second behavioral component are fused to obtain a comprehensive expected behavioral index of the information.

[0113] As an example, if the first inverted index is built solely based on the first row component, then the overall expected behavior index in the first inverted index is obtained entirely based on a greedy strategy. It is prone to getting trapped in local extrema, for example, always recalling a few very similar pieces of information with good historical performance. The significance of the second row component lies in the fact that if there is too little knowledge about a niche piece of information, then its expected behavior feature (historical average return, such as historical average click-through rate) has a very low confidence level at this time, and the uncertainty is very high, resulting in a large confidence interval. That is, it is not believed that the expected behavior feature at this time is its true average return, so it is necessary to select this information to obtain more exploration results. Therefore, the second row component can be understood as an indicator that measures how much knowledge is available about a certain piece of information. The less knowledge is available, the larger the second row component is. By integrating the second row component with the first row component, it can be understood that when there is insufficient knowledge about a piece of information, that information still has a chance to be selected, even if the average return of this information is very low.

[0114] In step 1034, a first inverted index table is established using the information's tag as the key and the information and its corresponding comprehensive expected behavior index as the value.

[0115] As an example, suppose there are three pieces of information, information A, information B, and information C. Information A, information B, and information C all have 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 can be a key-value pair, for example, the key is label 1, and the values ​​are information A (0.5), information B (0.45), and information C (0.3). In the subsequent online process, retrieval can be based on the first inverted index table.

[0116] As an example, steps 101-103 can be completed online or offline. The first inverted index table is updated at fixed update intervals to ensure that the indicator data used for recall based on the first inverted index table in subsequent online processes is valid, thereby ensuring the recommendation accuracy of the recommendation system.

[0117] 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 processed, and a recommendation operation is performed based on the sampling and processing results.

[0118] As an example, performing recommendation operations based on sampling processing results is actually making recommendations based on the information obtained from sampling. Recommendation operations can include sorting operations, reordering operations, and shuffling operations, etc.

[0119] In some embodiments, see Figure 3C , Figure 3C This is a flowchart illustrating the AI-based recommendation method provided in this application embodiment. In step 104, the information in the first inverted index table is sampled based on the comprehensive expected behavior index of the information in the first inverted index table, which can be implemented through steps 1041-1043.

[0120] In step 1041, the profile tags in the user profile are obtained from the client's recommendation request.

[0121] In step 1042, multiple pieces of information corresponding to the same tags as the portrait tags are obtained from the first inverted index table to form a set of information to be sampled.

[0122] As an example, the client's recommendation request carries a user profile, which includes multiple profile tags. Profile tags can represent a user's interests and social attributes, such as pregnant woman, mother and baby, etc. The first inverted index table stores multiple tags, each with corresponding information. For example, if the first inverted index table also has a mother and baby tag, then multiple information corresponding to the tag (mother and baby) that matches the profile tag (mother and baby) will be retrieved from the first inverted index table to form a set of information to be sampled. Here, "matching" can be understood as completely matching, or if the matching degree of two tags is greater than the matching degree threshold, then they are considered to be matching by default. For example, although "pregnant woman" does not completely match "mother and baby", if the matching degree is greater than the matching degree threshold, then when the profile tag is "pregnant woman", multiple information corresponding to the tag "mother and baby" will also be retrieved.

[0123] In step 1043, the set of information to be sampled is sampled based on the comprehensive expected behavior index of each piece of information in the set of information to be sampled.

[0124] In some embodiments, the sampling process in step 1043, which involves sampling information based on the comprehensive expected behavior index of each piece of information in the set of information to be sampled, can be achieved through the following technical solution: performing sampling processing a number of times equal to the sampling number threshold, and performing the following processing in each sampling process: converting the comprehensive expected behavior index of each piece of information in the set of information to be sampled into a sub-interval within a numerical range; generating a random number for the corresponding numerical range using a random function; determining the information corresponding to the sub-interval into which the random number falls as the sampled information, and removing it from the set of information to be sampled.

[0125] As an example, the sampling threshold is the number of pieces of information that ultimately need to be recalled. For instance, if the number of pieces of information to be recalled is 5000, then 5000 samplings without replacement will be performed. Taking one sampling without replacement as an example, assuming there are three pieces of information in the set to be sampled: information A, information B, and information C, and the comprehensive expected behavior indices of information A, information B, and information C are 0.5, 0.3, and 0.2, respectively, then for the comprehensive expected behavior index of each piece of information in the set to be sampled, a corresponding sub-interval is determined within a certain numerical range. Then the sampling probability of information A is 0.5 / (0.5+0.3+0.2). Assuming the numerical range is 0-1, then the comprehensive expected behavior index of information A, 0.5, is transformed into the sub-interval 0-0.5, and the comprehensive expected behavior index of information B, 0.3, is transformed into the sub-interval 0-0.5. The range is 0.5-0.8. The comprehensive expected behavior index of information C (0.2) is transformed into a sub-range of 0.8-1. A random number is generated within the range of 0-1 using a random function. If the generated random number is 0.6, it falls within the sub-range of 0.5-0.8 derived from the comprehensive expected behavior index of information B (0.3). Information B is then identified as the sampled information and removed from the set of information to be sampled. For the next sampling, information A and information C are sampled, with the sampling probability of information A being 0.5 / (0.5+0.2). Assuming the range remains 0-1, the comprehensive expected behavior index of information A (0.5) is transformed into a sub-range of 0-5 / 7, and the comprehensive expected behavior index of information C (0.2) is transformed into a sub-range of 5 / 7-1. Random numbers are then generated again for sampling. This sampling method allows information without a high comprehensive expected behavior index to be recalled, thereby uncovering non-head-level mid-to-long-tail interests and satisfying the diversity of interests in the recommendation process.

[0126] In some embodiments, see Figure 3D , Figure 3D This is a flowchart illustrating the recommendation method based on artificial intelligence provided in this application embodiment. In step 104, the information in the first inverted index table is sampled based on the comprehensive expected behavior index of the information in the first inverted index table, which can be implemented through steps 1044-1045.

[0127] In step 1044, multiple candidate inverted index tables are adapted based on the scenario to which the client's recommendation request belongs.

[0128] In step 1045, when the first inverted index table is adapted from multiple candidate inverted index tables, the information in the first inverted index table is sampled based on the comprehensive expected behavior index of the information in the first inverted index table.

[0129] In some embodiments, in addition to the first inverted index table, the recommendation system also has inverted index tables built based on other criteria. For example, there is a second inverted index table built based on descending order of information creation time, a third inverted index table built based on descending order of information posterior quality parameters, and a fourth inverted index table built based on descending order of profile similarity. Different inverted index tables are suitable for different scenarios. For example, in a cold start recommendation scenario or a scenario that needs to explore diverse user interests, the first inverted index table can be adapted from multiple candidate inverted index tables. Then, 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 processed.

[0130] In some embodiments, the appropriate inverted index table can be selected based on the data completeness or timeliness of the information in the database. For example, for data of three dimensions: publication time, post-hoc quality, and profile data, the inverted index of the corresponding method is selected for retrieval based on the data with the highest completeness.

[0131] In some embodiments, the multiple candidate inverted index tables also include a second inverted index table, and the second inverted index table includes a descending order based on the information creation time; the adaptation of multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs in step 1044 can be achieved through the following technical solution: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with time-sensitive attributes, the second inverted index table is determined as the inverted index table adapted to the recommendation request.

[0132] In some embodiments, when a second inverted index table is adapted from multiple candidate inverted index tables, the publication time of multiple pieces of information ranked at the top of the second inverted index table is updated; wherein the update value used for updating the publication time is positively correlated with the posterior quality parameter of the information; multiple pieces of information ranked at the top are obtained from the updated second inverted index table to perform a recommendation operation.

[0133] As an example, when the recommendation request corresponds to a recommendation scenario with a time-sensitive attribute, the second inverted index table is determined as the inverted index table adapted to the recommendation request. A recommendation scenario with a time-sensitive attribute could be a morning news recommendation scenario. Typically, in morning news, it's necessary to ensure that all information is up-to-date. Therefore, when the second inverted index table is matched among multiple candidate inverted index tables, the publication time of the top-ranked information in the second inverted index table is updated, similar to the first inverted index table. This includes key-value pairs, where the key is the tag, and the value is the information and its publication time. For example, if the profile tag is "maternal and infant," then the top-ranked information (with a set proportion or quantity) corresponding to the "maternal and infant" tag is retrieved, and a posterior feedback correction value (update value) is introduced for their publication time. The publication time (composed of the year, month, day, hour, and minute fields)... The new publication time is obtained by combining the combination of the two factors and the posterior feedback correction value (the quantified value of the posterior feedback). The updated value used for updating the publication time is positively correlated with the posterior quality parameter of the information. Here, the posterior feedback correction value is positive, which is equivalent to correcting the publication time of high-quality and earlier information to a larger value, so as to prevent high-quality information from sinking due to slightly older publication. Since the publication time of the header information is corrected, multiple pieces of information ranked at the top are obtained from the updated second inverted index table to perform the recommendation operation. Since the quality has been corrected, multiple pieces of information ranked at the top (with a set proportion or number of pieces of information) can be directly extracted to perform the recommendation operation, or the above-mentioned sampling without replacement method can be adopted, using the corrected publication time as the above-mentioned comprehensive expected behavior index, to perform sampling without replacement, and then perform the recommendation operation based on the sampling results.

[0134] By combining time and quality through the above implementation method, and under the premise of satisfying time sensitivity, information that has been published for a slightly longer period of time is subject to a certain degree of weight adjustment, thereby ensuring the recommendation quality in time-sensitive scenarios.

[0135] In some embodiments, the multiple candidate inverted index tables also include a third inverted index table, and the third inverted index table includes a descending order based on the posterior quality parameter of the information; the adaptation of multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs in step 1044 can be achieved through the following technical solution: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with quality attributes, the third inverted index table is determined as the inverted index table adapted to the recommendation request.

[0136] In some embodiments, when a third inverted index table is adapted from multiple candidate inverted index tables, the posterior quality parameters of the top-ranked information in the third inverted index table are updated; wherein, the update value used for updating the posterior quality parameters is negatively correlated with the number of times the information is exposed; the top-ranked information is obtained from the updated third inverted index table to perform a recommendation operation.

[0137] As an example, the posterior quality parameters of information can be parameters used to measure information quality based on user feedback, such as click-through rate, forwarding rate, and collection rate, or parameters used to measure information quality based on a combination of click-through rate, forwarding rate, collection rate, etc.

[0138] As an example, when the recommendation request corresponds to a recommendation scenario with quality attributes, the third inverted index table is determined as the inverted index table adapted to the recommendation request. A recommendation scenario with quality attributes can be a trending search scenario. Typically, in trending searches, it's necessary to ensure that all information is the most popular information. Therefore, when the third inverted index table is adapted from multiple candidate inverted index tables, the posterior quality parameters of the top-ranked information in the third inverted index table are updated. Similar to the first inverted index table, this includes key-value pairs, where the key is the tag, and the value is the information and its posterior quality parameters. For example, if the profile tag is "maternal and infant," then the top-ranked information (with a set proportion or quantity) with the highest posterior quality parameters among all information corresponding to the "maternal and infant" tag are retrieved. An exposure count correction value (update value) is introduced for their posterior quality parameters. The new posterior quality parameter is obtained by adding the exposure count correction value (the normalized value of the exposure count) to the posterior quality parameter. Here, the exposure count correction value is negative. The update value used to update the posterior quality parameter is negatively correlated with the exposure count of the information. This is equivalent to reducing the ranking advantage gained by earlier published information due to accumulated exposure. Since the posterior quality parameter of the header information has been corrected, multiple pieces of information ranked at the top are retrieved from the updated third inverted index table to perform the recommendation operation. Since the exposure count has been corrected, multiple pieces of information ranked at the top (with a set proportion or number of pieces of information) can be directly extracted to perform the recommendation operation. Alternatively, the above-mentioned sampling without replacement method can be adopted, using the corrected posterior quality parameter as the above-mentioned comprehensive expected behavior index for sampling without replacement, and then performing the recommendation operation based on the sampling results.

[0139] By combining exposure and quality through the above implementation method, under the premise of satisfying quality sensitivity, information with high posterior quality parameters due to excessive exposure is subject to a certain degree of weight correction (lowering), thereby ensuring the recommendation quality in quality-sensitive scenarios.

[0140] In some embodiments, the multiple candidate inverted index tables also include a fourth inverted index table, and the fourth inverted index table includes a descending order based on image similarity; in step 1044, the adaptation of multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs can be achieved through the following technical solution: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with interest attributes, the fourth inverted index table is determined as the inverted index table adapted to the recommendation request.

[0141] In some embodiments, when a fourth inverted index table is adapted from multiple candidate inverted index tables, the information profile similarity of multiple pieces of information ranked lower in the fourth inverted index table is updated; wherein, the update value used for profile similarity update is negatively correlated with the ranking position of the information; multiple pieces of information ranked higher are obtained from the updated fourth inverted index table to perform a recommendation operation.

[0142] As an example, when the recommendation request corresponds to a recommendation scenario with interest attributes, the fourth inverted index table is determined as the inverted index table adapted to the recommendation request. A recommendation scenario with quality attributes could be a short video recommendation scenario. Typically, in short video recommendations, it's necessary to ensure that the information best matches the user's interests. Therefore, when the fourth inverted index table is matched among multiple candidate inverted index tables, the information profile similarity of the lower-ranked information in the fourth inverted index table is updated. Similar to the first inverted index table, this includes key-value pairs, where the key is the tag and the value is the information and its information profile similarity. For example, if the profile tag is "maternal and infant," then the lower-ranked information (with a set proportion or number of items) corresponding to the "maternal and infant" tag is retrieved, and a similarity correction value (update value) is introduced for their information profile similarity. The information profile similarity plus the similarity correction value yields... The new information profile similarity is used to update the profile similarity value, which is negatively correlated with the information's ranking position (it can be an update value that shows a decreasing trend in the long tail; here, the similarity correction value is positive, although the similarity has increased, the increase shows a decreasing trend in the long tail). This is equivalent to mitigating the ranking disadvantage of information in the long tail that always sinks due to low similarity. Since the information profile similarity of the tail information has been corrected, multiple pieces of information ranked at the top are obtained from the updated fourth inverted index table to perform recommendation operations. Since the exposure count has been corrected, multiple pieces of information ranked at the top (with a set proportion or number of pieces of information) can be directly extracted to perform recommendation operations, or the above-mentioned sampling without replacement method can be adopted, using the corrected information profile similarity as the above-mentioned comprehensive expected behavior index for sampling without replacement, and then performing recommendation operations based on the sampling results.

[0143] By combining profile similarity with ranking position through the above implementation method, the profile similarity of mid-to-long-tail information is positively corrected under the premise of satisfying interest similarity, thereby ensuring the diversity of information recommendation in interest-sensitive scenarios.

[0144] In some embodiments, the adaptation of multiple candidate inverted index tables based on the scene to which the client's recommendation request belongs in step 1044 can be achieved through the following technical solution: obtaining real-time data of the scene to extract real-time features of the scene; calling a neural network model based on the scene features to predict the attributes of the scene; wherein, the training samples of the neural network model include front-end scene data, and the labeled data of the training samples include the attributes of the scene corresponding to the front-end scene data; determining the inverted index table that matches the predicted attributes from multiple candidate inverted index tables.

[0145] As an example, an inverted index table that matches the predicted attributes can also be determined from multiple candidate inverted index tables using a neural network model. The training samples of the neural network model include front-end scene data, and the labeled data of the training samples include the attributes of the scene corresponding to the front-end scene data. The inverted index table can be determined quickly and efficiently through the neural network model, achieving a good balance between recommendation quality and recommendation efficiency.

[0146] In some embodiments, the sampling processing of information in the first inverted index table based on the comprehensive expected behavior index of information in the first inverted index table in step 104 can be achieved through the following technical solution: obtaining profile tags from the user profile carried by the client's recommendation request, and performing adaptation processing on multiple candidate inverted index tables based on the scenario to which the client's recommendation request 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 among multiple candidate inverted index tables; obtaining information corresponding to tags consistent with the profile tags from the first inverted index table to obtain a set of information to be sampled; obtaining the corresponding score of each piece of information in the set of information to be sampled in each candidate inverted index table; fusing the comprehensive expected behavior index of each piece of information in the set of information to be sampled with the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index; and sampling processing of the set of information to be sampled based on the new comprehensive expected behavior index of each piece of information in the set of information to be sampled.

[0147] As an example, user profile tags are obtained from the user profile carried in the client's recommendation request. Based on the scenario to which the client's recommendation request belongs, multiple candidate inverted index tables are adapted to obtain at least one candidate inverted index table. The candidate inverted index table is an inverted index table that is different from the first inverted index table among the multiple candidate inverted index tables. Information corresponding to tags that are consistent with the user profile tags is obtained from the first inverted index table to obtain a set of information to be sampled, such as information A and information B. The corresponding score of each piece of information in the set of information to be sampled is obtained in each candidate inverted index table. For example, the score corresponding to the publication time of information A in the second inverted index table and the score corresponding to the quality of information A in the third inverted index table are obtained. The comprehensive expected behavior index of each piece of information in the set of information to be sampled is fused with the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index. The comprehensive expected behavior index of information A is fused with the above two scores to obtain the comprehensive expected behavior index. A similar processing process exists for information B. Finally, based on the new comprehensive expected behavior index of each piece of information in the set of information to be sampled, the set of information to be sampled is sampled. The above implementation method can take into account recommendation scenarios with multiple attributes, realize effective information recommendation, and enable the recommendation system to be reused in different recommendation scenarios, thereby improving recommendation efficiency and saving the background computing resources occupied by the recommendation system.

[0148] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0149] The AI-based recommendation method provided in this application can be applied to various recommendation scenarios, such as e-commerce recommendation scenarios and news recommendation scenarios. The following description uses a news recommendation system as an example. This application provides an AI-based recommendation method that differs from the inverted index retrieval method based on interest-based recommendation algorithms used in related technologies. This application provides an AI-based recommendation method that calculates a comprehensive expected behavior index for each piece of information still in the information pool. An inverted index (first inverted index table) is then built at the tag granularity based on the comprehensive expected behavior index. The comprehensive expected behavior index considers the click-through rate and freshness (exposure count) of the information, which is beneficial for effectively exploring low-exposure information while utilizing effective behaviors, such as user click behavior and the feedback data. Unlike the candidate information interception schemes in related technologies, in the AI-based recommendation method provided in this application, tags based on user profiles are pulled from the Redis database. Information is retrieved, and the final information is extracted based on the comprehensive expected behavior index as the sampling probability for all retrieved information. Since the comprehensive expected behavior index is used as the probability, it is beneficial to take into account the information click-through rate and cold start information exploration. In the AI-based recommendation method provided in this application embodiment, the weight of the profile is avoided to be applied to a single piece of information. This makes information with similar click-through rates and exposure corresponding to the user's head interests and mid-to-long-tail interests have similar probabilities to be sampled as recall information. This makes mid-to-long-tail information have higher exposure, increases the probability of information corresponding to the user's long-tail interests being recommended, explores the user's diverse interests, and thus improves the user experience. Furthermore, since the probability sampling is based on the comprehensive expected behavior index, each piece of information has an exposure probability, making the information recommended each time have a certain difference, which further improves the diversity of the recommended information. Finally, the recalled information is recommended to the user after being processed by the sorting module and the re-sorting module.

[0150] See Figure 4 , Figure 4This application provides a system architecture diagram for an AI-based recommendation method. The news recommendation system includes four main modules: a user profiling module, a recall module, a ranking module, and a scattering module. The user profiling module calculates the user's long-term and short-term interests based on the user's historical behavior, thus providing basic information for the recall and ranking modules. The recall module mines information of interest to the user from different angles from massive amounts of information. The ranking module predicts the click-through rate of the recalled information based on user information and basic information. To avoid recommending overly simplistic information to the user, the scattering module collects the information after the ranking module according to pre-set restrictions and finally presents it to the user. The user browses the information recommended by the client and clicks and reads the thumbnails of the information titles and illustrations according to their interests. The recall algorithm used by the recall module is a filter between the information pool and the information that the user may be potentially interested in, providing basic data for the subsequent ranking and scattering modules. An inaccurate recall algorithm will result in the recommended information not including the user's potential interests, severely limiting the final recommendation effect and ultimately leading to a poor user experience.

[0151] The recall module in related technologies uses interest-based algorithms, which retrieve corresponding information from the inverted index based on user profiles accumulated from historical user behavior. Unlike other types of recall algorithms, interest-based algorithms, due to the use of users' accumulated historical behavior, have advantages such as high accuracy, stable representation of user interests, and strong interpretability. The recall module usually sets a large recall volume for interest-based recalls to enter the ranking module. However, the establishment of the inverted index in interest-based recall strategies in related technologies mainly depends on the information release time or information click rate. This causes some information to occupy the top of the inverted index for a long time, which is not conducive to the cold start of information and the distribution of less popular information. The recall strategy mainly uses the user's head interests as the basis for inverted index retrieval, which will result in insufficient recall of users' mid-to-long-tail interests. The current algorithm for extracting recalled information mainly relies on the inverted index score or information and matching degree score (the matching degree score between information and user profile). This will cause some information that is not of real interest to occupy the top of the recall for a long time without producing actual recommendation effect. This not only wastes the computing resources of the ranking and scattering modules, but also fails to effectively explore users' potential interests, seriously affecting the user experience.

[0152] To address these issues, this application provides an artificial intelligence-based recommendation method, the processing flow of which in the entire recommendation system is as follows: Figure 4As shown, this is mainly applied to the recall module. It establishes a first inverted index table with tag granularity for information by comprehensively considering expected behavioral indicators, and stores this table offline. When a user requests information (recommendation request), the recall module retrieves information from the first inverted index table corresponding to the user profile's tags and samples it. The sampled information is the recalled information. After the sorting module predicts the click-through rate, the scattering module collects information and allocates positions, ultimately recommending the information to the user. The user browses and clicks according to their preferences. Actual click data is reported to the recommendation system for algorithm updates and data accumulation across different modules, thereby continuously improving the user experience.

[0153] See Figure 5A , Figure 5A This is a data flow diagram of the AI-based recommendation method provided in this application embodiment. In the offline part, the information still in the information pool is first obtained, and then the current exposure and cumulative click rate corresponding to the information are retrieved from the database. The comprehensive expected behavior index of each information is calculated using formula (3). Then, an information inverted index (first inverted index table) is established at the tag granularity, and the information inverted index is stored in the Redis database with the tag as the key for later use. Wilson(PV, Clk) is the range of historical average revenue based on the number of exposures and clicks of a certain information, usually taking the lower bound of the range. T is the number of times all information is exposed, PV is the number of times the corresponding information is exposed, and α and β are hyperparameters.

[0154]

[0155] See Figure 5B , Figure 5B This is a data flow diagram of the AI-based recommendation method provided in this application embodiment. In the online calculation part, when a user makes a request (recall request), the recall module first retrieves the tags from the user's historical cumulative profile (calculated based on historical click information), and then retrieves the corresponding information from the Redis database based on the tags in the profile. In order to capture the information of the number of requests, a sampling algorithm without replacement is used to sample the comprehensive expected behavior index of the information. Finally, the sampled information is returned to the user.

[0156] In a recommendation system, how to quickly and effectively distribute new information is an important issue. According to the central limit theorem, if the number of trials for a piece of information is large enough, the average revenue represents the real revenue (using the click rate of the information as the historical average revenue). Formula (4) can be used to calculate the comprehensive expected behavior index of each piece of information.

[0157]

[0158] First, the statistical average return for each piece of information is used to replace the actual return, in formula (4). This represents the historical average return (expected behavioral characteristics) of this information, and is achieved through... The upper bound of the uncertainty of the historical average return (the uncertainty of the expected behavioral characteristics) is represented by T, which is the number of times all information is exposed, and PV is the number of times the corresponding information is exposed. It can be found that when this information has been explored (exposed) many times, its return is equal to the historical average return. When the number of explorations (exposures) is small, the upper bound of the uncertainty of the historical average return will be large, so there should be a greater probability to explore the information.

[0159] In the AI-based recommendation method provided in this application embodiment, a comprehensive expected behavior index is calculated for each piece of information. The click-through rate of the information is used as the historical average revenue. In order to balance the historical average revenue and the uncertain revenue, two hyperparameters are used to control the ratio of the two. After experimentation, 0.9 and 0.1 were used as parameters for historical average revenue and uncertain revenue, respectively. The click-through rate is very inaccurate at the beginning, while the confidence of 10 clicks for 100 exposures and 1000 clicks for 10,000 exposures is very high. Therefore, in the AI-based recommendation method provided in this application embodiment, the concept of Wilson interval is introduced. The meaning of Wilson interval is the range of the actual click-through rate under a certain confidence level. The Wilson correction formula is shown in formula (5).

[0160]

[0161] Where p is the probability of a click, also known as the click-through rate; n is the total number of samples, i.e., the number of times each piece of information is exposed; z is the standard layer in the normal distribution. For example, if z is 1.96, the click-through rate has a 95% confidence level.

[0162] Next, an inverted index was established based on the comprehensive expected behavior index. The key of the inverted index is the tag, and the value is the corresponding information and the comprehensive expected behavior index, as shown in Table 1. The key is the Tag id, for example, Tag: 5844784, and the value is the corresponding list of information. Each information is accompanied by its own comprehensive expected behavior index, and the information is sorted in descending order of the comprehensive expected behavior index, such as A5, A6, and A7. The top 50 pieces of information of the comprehensive expected behavior index of each tag are extracted as the inverted index information of that tag, which helps to save storage space and discard inferior information. In order to enhance the efficiency of online recall, this inverted index information is stored in the Redis database for later use.

[0163]

[0164] Table 1 First Inverted Index Table

[0165] When a user requests data (a recommendation request), the system first retrieves the user's profile data and extracts their profile tags, which represent the user's accumulated historical interests. Then, it connects to the Redis database and retrieves the corresponding inverted index information. The amount of information retrieved at this stage is generally higher than the number of items requested in this request. A sampling algorithm without replacement is used to sample the retrieved information, directly using the comprehensive expected behavior index as the sampling probability. This eliminates the need to consider the weight of the user profile, giving information about the user's long-tail interests a greater probability of exposure. This is beneficial for the distribution of information about users' niche interests and also for the development of new user interests.

[0166] The recalled information is sent to the sorting module. The sorting module predicts the click-through rate for each piece of information based on user characteristics, information characteristics, and contextual information (current time, etc.). The information sorted based on the click-through rate is sent to the scattering module. The scattering module selects information and assigns positions based on information classification, media type (video, text, etc.), etc. Finally, the information is transmitted to the user's client for display.

[0167] In the AI-based recommendation method provided in this application embodiment, the historical average return and uncertainty of each piece of information are used as the upper bound of the predicted return (comprehensive expected behavior index) for a single piece of information. The method determines which information to explore based on the predicted return. As the number of experiments increases, the uncertain return continuously decreases, eventually converging to the true value, i.e., the average historical return. Based on the law of large numbers, when the number of experiments is sufficiently large, the average return converges to the true return. Based on this, this application embodiment provides an AI-based recommendation method, the main process of which is as follows: First, all currently available information is retrieved, and the historical cumulative click-through rate and historical exposure count of the corresponding information are obtained. Based on the comprehensive expected behavior index... The algorithm calculates the comprehensive expected behavior index for each piece of information. Then, it uses the comprehensive expected behavior index to build a tag-granular inverted index, which stores the information identifier and the corresponding comprehensive expected behavior index. The inverted index information is stored in a Redis database and is updated offline every half hour. When a user requests a recall, the algorithm pulls the user's tag-granular profile and then retrieves the information corresponding to these tags from the Redis database as candidate information for this recall. The candidates are sampled to obtain multiple pieces of information from this request as recall data. Here, the sampling algorithm uses the comprehensive expected behavior index as the sampling probability to mine more long-tail information. The recall information is then processed by a sorting module and a shuffling module before finally being presented to the user.

[0168] The AI-based recommendation method provided in this application is applied to a news recommendation system. It uses a comprehensive expected behavior index to build an inverted ranking at the tag granularity. The inverted ranking is retrieved online based on profile tags, and then sampled and returned using a sampling algorithm without replacement. The inverted ranking uses the comprehensive expected behavior index as the sorting criterion, which is beneficial for information cold start and fully utilizing historical exposure information. The sampling algorithm facilitates the distribution of information with mid-to-long-tail interests and the diversity of information distributed each time. The recalled information is finally presented to the user after passing through a sorting module and a reordering module, resulting in a wider coverage of recalled information, extensive mining of information of interest to users, and improved user experience.

[0169] The following continues to describe the exemplary structure of the AI-based recommendation device 255 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2 As shown, the software modules in the AI-based recommendation device 255 stored in the memory 250 may include: an acquisition module 2551, used to acquire the exposure count and effective behavior count of each piece of information in the information set; a determination module 2552, used to determine the expected behavior characteristics of the corresponding information based on the exposure count and effective behavior count, and to determine the uncertainty characteristics of the expected behavior characteristics based on the exposure count; an establishment module 2553, used to establish a first inverted index table including the comprehensive expected behavior index of the information based on the expected behavior characteristics and uncertainty characteristics of each piece of information; and a recommendation module 2554, used 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 to perform a recommendation operation based on the sampling result.

[0170] In some embodiments, before obtaining the number of exposures and the number of valid actions for each piece of information in the information set, the acquisition module 2551 is further configured to: traverse the publication time of each piece of information in the information set at fixed time intervals; and remove the corresponding information from the information set when the publication time of the information is less than the publication time threshold.

[0171] In some embodiments, the determining module 2552 is further configured to: determine the expected behavior features of the corresponding information based on the following expected behavior feature function: Where E1 represents the desired behavioral feature, p represents the ratio between the number of effective behaviors and the number of exposures, z represents the confidence parameter, and n represents the number of times the information is exposed.

[0172] In some embodiments, the determining module 2552 is further configured to: determine the uncertainty feature corresponding to the desired behavior feature based on the following uncertainty feature function: Where E2 represents the uncertainty of the expected behavioral characteristics, T represents the sum of the exposure times of all information in the information set, and n represents the exposure times of the information.

[0173] In some embodiments, the establishment module 2553 is further configured to: determine a first behavioral component that is positively correlated with the expected behavioral characteristics; determine a second behavioral component that is positively correlated with the uncertainty characteristics; perform fusion processing on the first behavioral component and the second behavioral component to obtain a comprehensive expected behavioral index of the information; and establish a first inverted index table with the information's label as the key and the information and the corresponding comprehensive expected behavioral index of the information as the value.

[0174] In some embodiments, the recommendation module 2554 is further configured to: obtain profile tags from the user profile in the recommendation request from the client; obtain multiple pieces of information corresponding to the tags that are consistent with the profile tags from the first inverted index table to form a set of information to be sampled; and perform sampling processing on the set of information to be sampled based on the comprehensive expected behavior index of each piece of information in the set of information to be sampled.

[0175] In some embodiments, the recommendation module 2554 is further configured to: perform sampling processing for a number of times equal to the sampling number threshold, and perform the following processing in each sampling process: convert the comprehensive expected behavior index of each piece of information in the set of information to be sampled into a sub-interval within the numerical range; generate a random number for the corresponding numerical range using a random function; determine the information corresponding to the sub-interval into which the random number falls as the sampled information, and remove it from the set of information to be sampled.

[0176] In some embodiments, the recommendation module 2554 is further configured to: adapt multiple candidate inverted index tables based on the scenario to which the client's recommendation request belongs; when the first inverted index table is adapted from multiple candidate inverted index tables, 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.

[0177] In some embodiments, 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 2554 is further configured to: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with a time-sensitive attribute, determine the second inverted index table as the inverted index table adapted to the recommendation request.

[0178] In some embodiments, the recommendation module 2554 is further configured to: when adapting to the second inverted index table from multiple candidate inverted index tables, update the publication time of multiple pieces of information ranked at the top of the second inverted index table; wherein the update value used for updating the publication time is positively correlated with the posterior quality parameter of the information; and obtain the multiple pieces of information ranked at the top from the updated second inverted index table to perform a recommendation operation.

[0179] In some embodiments, 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 posterior quality parameter of the information; the recommendation module 2554 is further configured to: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with quality attributes, determine the third inverted index table as the inverted index table adapted to the recommendation request.

[0180] In some embodiments, the recommendation module 2554 is further configured to: when adapting to a third inverted index table from multiple candidate inverted index tables, update the posterior quality parameters of multiple pieces of information ranked at the top of the third inverted index table; wherein the update value used for updating the posterior quality parameters is negatively correlated with the number of times the information is exposed; and obtain the multiple pieces of information ranked at the top from the updated third inverted index table to perform a recommendation operation.

[0181] In some embodiments, 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 image similarity; the recommendation module 2554 is further configured to: when the recommendation scenario corresponding to the recommendation request is a recommendation scenario with interest attributes, determine the fourth inverted index table as the inverted index table adapted to the recommendation request.

[0182] In some embodiments, the recommendation module 2554 is further configured to: when adapting to the fourth inverted index table from multiple candidate inverted index tables, update the information profile similarity of multiple pieces of information ranked lower in the fourth inverted index table; wherein the update value used for profile similarity update is negatively correlated with the ranking position of the information; and obtain multiple pieces of information ranked higher from the updated fourth inverted index table to perform the recommendation operation.

[0183] In some embodiments, the recommendation module 2554 is further configured to: acquire real-time data of the scene to extract real-time features of the scene; call a neural network model based on the scene features to predict the attributes of the scene; wherein the training samples of the neural network model include front-end scene data, and the labeled data of the training samples include the attributes of the scene corresponding to the front-end scene data; and determine an inverted index table that matches the predicted attributes from multiple candidate inverted index tables.

[0184] In some embodiments, the recommendation module 2554 is further configured to: obtain profile tags from the user profile carried in the recommendation request from the client, and perform adaptation processing on multiple candidate inverted index tables based on the scenario to which the recommendation request 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 among the multiple candidate inverted index tables; obtain information corresponding to tags consistent with the profile tags from the first inverted index table to obtain a set of information to be sampled; obtain the corresponding score of each piece of information in the set of information to be sampled in each candidate inverted index table; fuse the comprehensive expected behavior index of each piece of information in the set of information to be sampled with the corresponding score in each other inverted index table to obtain a new comprehensive expected behavior index; and perform sampling processing on the set of information to be sampled based on the new comprehensive expected behavior index of each piece of information in the set of information to be sampled.

[0185] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the artificial intelligence-based recommended method described above in this application.

[0186] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to execute the AI-based recommendation method provided in this application, for example... Figures 3A-3D The AI-based recommendation method is shown.

[0187] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EP ROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0188] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0189] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0190] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0191] In summary, by using the exposure frequency and effective behavior frequency of information in the embodiments of this application, the expected behavioral features and the corresponding uncertainty features of the expected behavioral features for user behavior prediction are characterized in a targeted manner for different information. This takes into account the contribution of information with different exposure frequencies (such as cold start information and historical exposure information of different degrees) to user behavior prediction. At the same time, through sampling processing, it is beneficial to the distribution of information with mid-to-long-tail interests and the diversity of information distribution. As a result, the information obtained by sampling has a wide coverage and can better mine information that users are interested in, ensuring the accuracy of information recommendation in subsequent information recommendation, while effectively avoiding invalid recommendations, thereby saving the computing resources related to recommendation logic on the server.

[0192] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. An artificial intelligence-based recommendation method, characterized in that, The method includes: Get the number of exposures and the number of valid actions for each piece of information in the information set; Based on the number of exposures, the number of effective behaviors, the ratio of the number of effective behaviors to the number of exposures, and the confidence parameter, the expected behavioral characteristics corresponding to the information are determined, and based on the number of exposures and the sum of the number of exposures, the uncertainty characteristics of the expected behavioral characteristics are determined. Based on the expected behavior characteristics and uncertainty characteristics of each piece of information, a comprehensive expected behavior index is obtained; Multiple pieces of information corresponding to tags that match the user profile tags are obtained from the first inverted index table to form a set of information to be sampled; wherein, the first inverted index table uses the tags of the information as keys, and the information and the comprehensive expected behavior index corresponding to the information as values; The comprehensive expected behavior index of each piece of information in the set of information to be sampled is converted into a sub-interval within the numerical range, and the information corresponding to the sub-interval into which the random number falls is determined as the recall information; The recommendation operation is performed based on the recall information.

2. The method according to claim 1, characterized in that, Before obtaining the exposure count and effective behavior count for each piece of information in the information set, the method further includes: The publication time of each piece of information in the information set is iterated over at fixed time intervals. When the publication time of the information is less than the publication time threshold, the corresponding information is removed from the information set.

3. The method according to claim 1, characterized in that, The method further includes: Determine the first behavioral component that is positively correlated with the desired behavioral characteristics; Determine a second behavioral component that is positively correlated with the aforementioned uncertainty feature; The first behavioral component and the second behavioral component are fused to obtain a comprehensive expected behavioral index of the information. Using the tags of the information as keys and the information and the corresponding comprehensive expected behavior index as values, a first inverted index table is established.

4. The method according to claim 1, characterized in that, The step of converting the comprehensive expected behavior index of each piece of information in the set of information to be sampled into a sub-interval within a numerical range, and determining the information corresponding to the sub-interval into which the random number falls as the recall information, includes: Perform sampling processing a number of times equal to a sampling threshold, and perform the following processing during each sampling process: The comprehensive expected behavior index of each piece of information in the set of information to be sampled is converted into a sub-interval within the numerical range; Random numbers corresponding to the specified numerical range are generated using a random function; The information corresponding to the sub-interval into which the random number falls is determined as the recall information and removed from the set of information to be sampled.

5. The method according to claim 1, characterized in that, Before retrieving multiple pieces of information corresponding to tags that match the user profile tags from the first inverted index table to form a set of information to be sampled, the method further includes: Based on the scenario to which the client's recommendation request belongs, multiple candidate inverted index tables are adapted; When the first inverted index table is matched in the multiple candidate inverted index tables, the process proceeds to obtaining multiple pieces of information corresponding to the tags that are consistent with the user profile's profile tags from the first inverted index table, in order to form a set of information to be sampled.

6. The method according to claim 5, characterized in that, The plurality of candidate inverted index tables also include a second inverted index table, and the second inverted index table includes a descending sort based on the information creation time; The adaptation of multiple candidate inverted index tables based on the scenario to which the client-side recommendation request belongs includes: When the recommendation request corresponds to a recommendation scenario with a time-sensitive attribute, the second inverted index table is determined as the inverted index table that is adapted to the recommendation request.

7. The method according to claim 6, characterized in that, The method further includes: When the second inverted index table is matched to the multiple candidate inverted index tables, the publication time of the multiple pieces of information that are ranked first in the second inverted index table is updated. The update value used for updating the publication time is positively correlated with the information posterior quality parameter of the information. Retrieve the top-ranked information from the updated second inverted index table to perform a recommendation operation.

8. The method according to claim 5, characterized in that, The plurality of candidate inverted index tables also include a third inverted index table, and the third inverted index table includes a descending sort based on the posterior quality parameter of the information. The adaptation of multiple candidate inverted index tables based on the scenario to which the client-side recommendation request belongs includes: When the recommendation scenario corresponding to the recommendation request is a recommendation scenario with quality attributes, the third inverted index table is determined as the inverted index table that is adapted to the recommendation request.

9. The method according to claim 8, characterized in that, The method further includes: When the third inverted index table is matched to the multiple candidate inverted index tables, the information posterior quality parameters of the multiple information items ranked first in the third inverted index table are updated. Among them, the update value used for updating the information posterior quality parameter is negatively correlated with the number of times the information is exposed; The top-ranked information is retrieved from the updated third inverted index table to perform a recommendation operation.

10. The method according to claim 5, characterized in that, The plurality of candidate inverted index tables also include a fourth inverted index table, and the fourth inverted index table includes a descending sort based on image similarity; The adaptation of multiple candidate inverted index tables based on the scenario to which the client-side recommendation request belongs includes: When the recommendation request corresponds to a recommendation scenario with interest attributes, the fourth inverted index table is determined as the inverted index table that is adapted to the recommendation request.

11. The method according to claim 10, characterized in that, The method further includes: When the fourth inverted index table is matched to the multiple candidate inverted index tables, the information profile similarity is updated for multiple pieces of information that are ranked later in the fourth inverted index table. Among them, the update value used for updating the image similarity is negatively correlated with the sorting position of the information; The top-ranked information is retrieved from the updated fourth inverted index table to perform a recommendation operation.

12. A recommendation device based on artificial intelligence, characterized in that, The device includes: The acquisition module is used to acquire the number of exposures and the number of valid actions for each piece of information in the information set; The determination module is used to determine the expected behavioral characteristics corresponding to the information based on the number of exposures, the number of effective behaviors, the ratio of the number of effective behaviors to the number of exposures, and the confidence parameter, and to determine the uncertainty characteristics of the expected behavioral characteristics based on the number of exposures and the sum of the number of exposures. The recall module is used to retrieve multiple pieces of information corresponding to tags that are consistent with the user profile's profile tags from the first inverted index table to form a set of information to be sampled; wherein, the first inverted index table uses the tags of the information as keys, and the information and the comprehensive expected behavior index corresponding to the information as values; the comprehensive expected behavior index of each piece of information in the set of information to be sampled is converted into a sub-interval within a numerical range, and the information corresponding to the sub-interval in which the random number falls is determined as the recall information; The recommendation module is used to perform recommendation operations based on recall information.

13. The apparatus according to claim 12, characterized in that, The acquisition module is also used for: Before obtaining the exposure count and effective behavior count of each piece of information in the information set, if the publication time of the information is less than the publication time threshold, the corresponding information is removed from the information set.

14. The apparatus as claimed in claim 12, characterized in that, The device further includes: A module is established to determine a first behavioral component that is positively correlated with the desired behavioral feature; and to determine a second behavioral component that is positively correlated with the uncertainty feature. The first behavioral component and the second behavioral component are fused to obtain a comprehensive expected behavioral index of the information. Using the tags of the information as keys and the information and the corresponding comprehensive expected behavior index as values, a first inverted index table is established.

15. An electronic device, characterized in that, include: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the AI-based recommended method as described in any one of claims 1 to 11.

16. A computer-readable storage medium, characterized in that, It stores executable instructions for implementing the AI-based recommendation method as described in any one of claims 1 to 11 when executed by a processor.

17. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, the method of any one of claims 1 to 11 is implemented.

Images