Recommendation method, device, electronic device and storage medium based on artificial intelligence

Abstract

This invention provides an artificial intelligence-based recommendation method, apparatus, electronic device, and storage medium. The method includes: sampling the first candidate recommendation information set based on the click-through rate of each piece of information in the first candidate recommendation information set, and forming a second candidate recommendation information set from the sampled information; transferring the information with the highest click-through rate in the second candidate recommendation information set to a third candidate information set as the first piece of information in the third candidate information set, and continuing to transfer some information from the second candidate recommendation information set to the third candidate information set, wherein the semantic distance between adjacent transfers is greater than or equal to a semantic distance threshold; sorting the information in the third candidate recommendation information set, and performing a recommendation operation based on the information included in the sorting result. This invention enables fine-grained sorting results with low repetition rate and wide coverage.

Description

Technical Field

[0001] This invention relates to artificial intelligence technology, and more particularly to a recommendation method, apparatus, electronic device, and storage medium based on artificial intelligence. 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] The click-through rate prediction module of the recommendation system estimates and sorts the click-through rates of the input candidate recommendation information. It needs to estimate the click-through rates of hundreds to thousands of candidate recommendation information. Since there are a large number of these candidate recommendation information, it is difficult to display them all to the user. Therefore, a mixed sorting method is needed to extract a set of candidate recommendation information with a smaller number of candidate recommendation information to push to the user's mobile client.

[0004] In related technologies, the commonly used method is the mixed-rank recommendation method based on manual rules. This method selects a specified number of contents from the input set of candidate recommendation information, performs mixed-rank processing, and outputs them to the user. In recent years, mixed-rank schemes based on reinforcement learning and recurrent neural networks have also been proposed, but these schemes are relatively complex to implement and have not been widely adopted. Summary of the Invention

[0005] This invention provides an artificial intelligence-based recommendation method, apparatus, electronic device, and storage medium, which can efficiently and accurately perform comprehensive recommendations in various recommendation scenarios.

[0006] The technical solution of this invention is implemented as follows:

[0007] This invention provides an artificial intelligence-based recommendation method, comprising:

[0008] Based on the click rate of each piece of information in the first candidate recommendation information set, the first candidate recommendation information set is sampled, and the sampled information is used to form a second candidate recommendation information set.

[0009] The information with the highest click rate in the second candidate recommendation information set is transferred to the third candidate information set as the first information in the third candidate information set, and some information in the second candidate recommendation information set is transferred to the third candidate information set, and the semantic distance between two adjacent transferred information is greater than or equal to the semantic distance threshold.

[0010] The information in the third candidate recommendation information set is sorted, and a recommendation operation is performed based on the multiple pieces of information included in the sorting result.

[0011] This invention provides an artificial intelligence-based recommendation device, comprising:

[0012] The information sampling module is used to sample the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, and to form a second candidate recommendation information set from the sampled information.

[0013] The information transfer module is used to transfer the information with the highest click rate in the second candidate recommendation information set to the third candidate information set as the first information in the third candidate information set, and to continue to transfer some information in the second candidate recommendation information set to the third candidate information set, and the semantic distance between the information transferred in two adjacent transfers is greater than or equal to the semantic distance threshold.

[0014] The information sorting module is used to sort the information in the third candidate recommendation information set and perform recommendation operations based on the multiple pieces of information included in the sorting results.

[0015] In the above scheme, the AI-based recommendation device further includes:

[0016] The information generation module is configured to: acquire user characteristics of the object to be recommended, and acquire other objects to be recommended whose similarity to the user characteristics exceeds a user similarity threshold; wherein, the other objects to be recommended are users other than the object to be recommended in the recommendation log; acquire information corresponding to the other objects to be recommended in the candidate recommendation information database, and determine the click-through rate of the object to be recommended for the information; filter out multiple pieces of information whose click-through rate exceeds a click-through rate threshold from the candidate recommendation information database, and use the multiple pieces of information as candidate recommendation information corresponding to the object to be recommended, so as to form a first candidate recommendation information set.

[0017] In the above scheme, the AI-based recommendation device further includes:

[0018] The click-through rate (CTR) prediction module is further used to obtain candidate recommendation information features of the information; predict the CTR of the information based on the candidate recommendation information features and user features; wherein the CTR prediction model used for the prediction is trained using user features, candidate recommendation information features and actual CTR in the recommendation log as training samples, with the goal of minimizing the error between the actual CTR and the predicted CTR.

[0019] In the above scheme, the information sampling module is further configured to: when the number of information in the first candidate recommendation information set exceeds the threshold for activating the sampling function, determine the nonlinear sampling probability of the information based on the nonlinear sampling probability function, using the click rate of the information as a parameter in the nonlinear sampling probability function; perform random sampling processing on the information based on the nonlinear sampling probability of the information, and form the second candidate recommendation information set with the multiple information obtained from the sampling; when the number of information in the first candidate recommendation information set does not exceed the threshold for activating the sampling function, determine that all information in the first candidate recommendation information set belongs to the second candidate recommendation information set.

[0020] In the above scheme, the information sampling module is further configured to: determine the nonlinear sampling probability of the information according to the following nonlinear sampling probability function:

[0021] pro_sampling = 1 / (1+e) -prob_coef*(pctr-prob_bias) )

[0022] The pro_sampling is the non-linear sampling probability of the information, the prob_coef and the prob_bias are hyperparameters, e is the natural base, and pctr is the predicted click-through rate for the information.

[0023] In the above scheme, the information sampling module is further configured to: perform the following processing on each piece of information in the first candidate recommendation information set: generate a random number through a random function, and proportionally convert the nonlinear sampling probability of the information to the same numerical range as the random number; determine the information corresponding to the nonlinear sampling probability greater than the converted random number as the randomly sampled information, and transfer it from the first candidate recommendation information set to the second candidate recommendation information set.

[0024] In the above scheme, the information sampling module is further configured to: when the number of information in the second candidate recommendation information set is less than the sampling output number threshold, perform the following processing for each piece of information in the first candidate recommendation information set: assign the same sampling probability to each piece of information in the first candidate recommendation information set; randomly extract information from the first candidate recommendation information set according to the same sampling probability to transfer it to the second candidate recommendation information set, until the number of information in the second candidate recommendation information set reaches the sampling output number threshold.

[0025] In the above scheme, the information transfer module is further configured to: when the number of information in the second candidate recommendation information set exceeds the threshold for initiating the deduplication function, the number of information in the third candidate recommendation information set is less than the threshold for initiating the deduplication function, and the second candidate recommendation information set is not empty, perform the following processing on the information with the highest click-through rate in the second candidate recommendation information set: determine the semantic distance between the information with the highest click-through rate in the second candidate recommendation information set and the first information; when the semantic distance is greater than the semantic distance threshold, transfer the information with the highest click-through rate from the second candidate recommendation information set to the third candidate recommendation information set and update it to the first information in the third candidate recommendation information set; when the semantic distance is not greater than the semantic distance threshold, transfer the information with the highest click-through rate from the second candidate recommendation information set to the alternative recommendation information set.

[0026] In the above scheme, the information transfer module is further configured to: transfer all information in the second candidate recommendation information set to the third candidate recommendation information set when the number of information in the second candidate recommendation information set does not exceed the threshold for starting the deduplication function.

[0027] In the above scheme, the information transfer module is further configured to: when the number of information in the third candidate recommendation information set is less than the threshold for initiating the deduplication function, and the second candidate recommendation information set is empty, perform the following processing on the substitute recommendation information set: assign the same sampling probability to each piece of information in the substitute recommendation information set; randomly extract information from the substitute recommendation information set according to the same sampling probability to transfer it to the third candidate recommendation information set, until the number of information in the third candidate recommendation information set reaches the threshold for initiating the deduplication function.

[0028] In the above scheme, the information sorting module is further configured to: remove the information with the highest click-through rate from the third candidate recommendation information set and determine it as the information ranked first; when the third candidate recommendation information set is not empty, perform the following processing for each piece of information in the third candidate recommendation information set: determine the ranking score of the information based on the click-through rate of the information in the third candidate recommendation information set and the semantic distance between it and the information ranked first; determine the information with the highest ranking score as the information ranked next and remove it from the third candidate recommendation information set.

[0029] In the above scheme, the information sorting module is further configured to: obtain the semantic feature distance between the information and the information ranked first; and determine the sorting score of the information based on a scoring function.

[0030] score=pctr A +w s *1 / ed AB ;

[0031] Wherein, the pctr A The click-through rate of the information, w s As a hyperparameter, the d AB The semantic distance between the information and the information ranked first.

[0032] This invention provides an electronic device, comprising:

[0033] Memory, used to store executable instructions;

[0034] The processor, when executing executable instructions stored in the memory, implements the AI-based recommendation method provided in this embodiment of the invention.

[0035] This invention provides a storage medium storing executable instructions that, when executed by a processor, implement the AI-based recommendation method provided in this invention.

[0036] The embodiments of the present invention have the following beneficial effects:

[0037] By transferring and sorting information based on click-through rate and semantic distance within a set, the resulting information not only has the characteristic of being inclined to be clicked, but also has semantic distance from each other, achieving a low repetition rate in recommendations. Recommending through multiple pieces of information can achieve broad coverage of points of interest, and does not rely on manually set rules, thus achieving efficient and fine-grained accurate recommendations. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of an AI-based hybrid layout module provided by related technologies;

[0039] Figure 2 This is an optional structural diagram of an artificial intelligence-based recommendation system 100 provided in an embodiment of the present invention;

[0040] Figure 3 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 the present invention;

[0041] Figures 4A-4E This is an optional flowchart illustrating an artificial intelligence-based recommendation method provided in an embodiment of the present invention.

[0042] Figure 5 This is a flowchart illustrating a practical application scenario of the recommendation method based on artificial intelligence provided in this embodiment of the invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] 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.

[0045] 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 the invention described herein can be implemented in an order other than that illustrated or described herein.

[0046] 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 invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0047] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be interpreted as follows.

[0048] 1) Candidate recommendation information: This refers to the items (candidate recommendation information) in the recommendation system. Items can refer to content that needs to be recommended, such as text, images, videos, and products.

[0049] 2) Primary Interests (Category L1): Represents a broad range of interests, such as "food", "economics", "military", etc. Recommendation systems typically include dozens of primary interests.

[0050] 3) Secondary interests (Category L2): These represent sub-interests under a specific primary interest. For example, the primary interest "Sports" includes multiple secondary interests such as "Basketball" and "Football". In recommendation systems, there are typically hundreds of secondary interests.

[0051] 4) Tags: Tags are used to label candidate recommendation information through automatic extraction or manual annotation. For example, "university", "Apple", "mobile phone". A candidate recommendation information can correspond to one or more tags. In a recommendation system, there are generally tens of thousands to hundreds of thousands of tags.

[0052] See Figure 1 , Figure 1 This is a schematic diagram of an AI-based mixed-ranking module provided by related technologies. The click-through rate (CTR) prediction module of the recommendation system estimates and ranks the CTR of the input candidate recommendation information. The CTR prediction module of the recommendation system in related technologies needs to estimate the CTR of hundreds to thousands of candidate recommendations. Some CTR prediction modules supporting news applications process an average of about 2000 candidate recommendations. This is a large number of candidates, making it difficult to display them all to the user. Therefore, a mixed-ranking method is needed. This method uses a predefined strategy to extract a small set of candidate recommendations, such as 10-20 recommendations, to push to the user's client. The mixed-ranking module of the recommendation system typically uses rules based on human experience, such as primary and secondary interests, tags, and CTR of the candidate recommendations, to generate a specified number of candidate recommendation sequences, based on the principle of diversity.

[0053] In related technologies, rule-based mixed recommendation methods are commonly used. These methods select a specified number of items from the input set of candidate recommendations, mix them, and output the results to the user. While some mixed recommendation schemes based on reinforcement learning (RL) and recurrent neural networks (RNN) have been proposed, these are complex to implement and not widely adopted. Rule-based mixed recommendation is more widely used. Assuming the input data format of the mixed recommendation module is (candidate recommendation ID, click-through rate, categoryL1, categoryL2, tag), typically several hundred to several thousand candidate recommendations are input and 10-20 candidate recommendations are output in each mixed recommendation. An example of rule-based mixed recommendation is as follows: Define the candidate recommendation set S output by the click-through rate prediction module; define the candidate recommendation set T output by the mixed recommendation module, initialized to empty; Rule 1: Select the candidate recommendation with the highest click-through rate from set S as the first item output by the mixed recommendation module. Rule 1: Add the candidate recommendation information to set T; delete the candidate recommendation information from set S. Rule 2: In set S, select the second candidate recommendation information and add it to set T according to the combination rules of primary interest, secondary interest and tag; delete the candidate recommendation information from set S. The specific combination rules can be formulated according to the needs of the application scenario. Rule 3: In set T generated by mixed sorting, the total number of candidate recommendation information for each primary interest cannot exceed m. Rule 4: In set T generated by mixed sorting, the total number of candidate recommendation information for each secondary interest cannot exceed n. Rule 5: In set T generated by mixed sorting, the total number of candidate recommendation information for each tag cannot exceed k. Rule 6: In the candidate recommendation information sequence generated by mixed sorting, two candidate recommendation information with the same secondary interest cannot be adjacent.

[0054] By using the aforementioned manual rules, corresponding mixed-rank output results can be generated and pushed to users. This method requires various rule combinations and tests based on extensive human experience for specific application scenarios, making the work quite tedious. When there are many rules, conflicts may exist between different rules, and optimization and improvement are not easy. As mentioned earlier, in recommendation systems of related technologies, manual rule-based mixed-ranking methods are usually used to achieve mixed-ranking of recommended content. Manual rule-based mixed-ranking methods require a large number of tedious rule combination attempts, the granularity of mixed-ranking is coarse and difficult to optimize. In addition, manual rule-based mixed-ranking methods need to be optimized for different application scenarios and do not have good adaptability. Mixed-ranking schemes based on reinforcement learning, RNN, and other methods have also been proposed, but these schemes are relatively complex and have not yet been widely used in the industry.

[0055] Therefore, to address the shortcomings of commonly used manual rule-based hybrid ranking methods in related technologies, an artificial intelligence-based recommendation method is proposed. The hybrid ranking algorithm used is simple and efficient. Compared with the commonly used manual rule-based methods in related technologies, the artificial intelligence-based recommendation method provided in this embodiment of the invention uses a nonlinear probability sampling function to sample candidate recommendation information, which can effectively explore user interests while ensuring recommendation effect. Based on probability sampling, the artificial intelligence-based recommendation method provided in this embodiment of the invention uses a method based on latent semantic similarity to automatically deduplicate and hybridize candidate recommendation information, which can effectively cover user interests at a finer granular level and can better adapt to different application scenarios.

[0056] This invention provides an artificial intelligence-based recommendation method, apparatus, electronic device, and storage medium, capable of achieving fine-grained ranking results with low repetition rate and wide coverage. The following describes exemplary applications of the electronic device provided in this invention. The electronic device can be implemented as various types of user terminals such as laptops, tablets, desktop computers, set-top boxes, and mobile devices (e.g., mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable gaming devices), or as a server. The following describes exemplary applications when the device is implemented as a server.

[0057] See Figure 2 , Figure 2This is an optional architecture diagram of an AI-based recommendation system 100 provided in this embodiment of the invention. This recommendation system can support application scenarios such as recommending news, recommending products, and recommending videos. Depending on the application scenario, candidate recommendation information can be news, actual products, etc. During user interaction with the client, the terminal 400 reports the collected user interaction behavior regarding candidate recommendation information to the server 200 as training sample data, along with the corresponding user profile and user characteristics. The training sample data comes from the behavioral data of different users reported by various terminals. Based on this behavioral data, a click-through rate prediction model is trained. The user profile and user characteristics are obtained from feedback from the terminal corresponding to a specific user. The click-through rate prediction model predicts the click-through rate of candidate recommendation information based on the acquired user profile and user characteristics corresponding to that user, and performs mixed sorting and recommendation based on the prediction results. Therefore, it can recommend information that meets user interests. Given the content of the points, the diversity of recommended content can be achieved through mixed sorting. The specific architecture of the recommendation system is described below: Terminal 400 connects to server 200 through network 300. Network 300 can be a wide area network, a local area network, or a combination of both. Server 200 contains an information generation module, a click-through rate prediction module, and a mixed sorting module. The information generation module uses various recommendation algorithms to quickly filter out content relevant to the user from the candidate recommendation information database 500 and submits it to the click-through rate prediction module. The click-through rate prediction module uses a ranking model to predict and rank the click-through rate of the content, and then submits it to the mixed sorting module. The mixed sorting module selects a specified number of contents from the output of the click-through rate prediction module and rearranges them to recommend to the user, which is then presented on the user's terminal 400. Logs containing data such as content exposure and click-through rate are reported to the recommendation system in real time as training samples for generating real-time user features and real-time content features, and are used to train the click-through rate prediction model in the click-through rate prediction module.

[0058] See Figure 3 , Figure 3 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 the present invention. Figure 3 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 3 The general labeled all buses as Bus System 240.

[0059] 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.

[0060] 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.

[0061] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this embodiment is intended to include any suitable type of memory.

[0062] 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.

[0063] 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;

[0064] 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, Wi-Fi, and Universal Serial Bus (USB).

[0065] In some embodiments, the AI-based recommendation device provided in this invention can be implemented in software. Figure 3 An AI-based recommendation device 255 stored in memory 250 is shown. It can be software in the form of programs and plug-ins, including the following software modules: information generation module 2551, click-through rate prediction module 2552, information sampling module 2553, information transfer module 2554, and information sorting module 2555. These modules are logically related and can therefore be arbitrarily combined or further divided according to the functions they implement. The functions of each module will be described below.

[0066] In other embodiments, the AI-based recommendation device provided in this invention can be implemented in hardware. As an example, the AI-based recommendation device provided in this invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the AI-based recommendation method provided in this invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0067] The artificial intelligence-based recommendation method provided by the embodiments of the present invention will be described in conjunction with exemplary applications and implementations of the recommendation system provided in the embodiments of the present invention.

[0068] See Figure 4A , Figure 4A This is an optional flowchart illustrating an artificial intelligence-based recommendation method provided in an embodiment of the present invention, which will be combined with... Figure 4A Steps 101-103 shown are explained.

[0069] In step 101, the server samples the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, and forms a second candidate recommendation information set from the sampled information.

[0070] In some embodiments, the information in the first candidate recommendation information set is the candidate recommendation information output by the click-through rate prediction module. The information in the first candidate recommendation information set is sampled using nonlinear probability to generate a new candidate recommendation information set, which is the second candidate recommendation information set. The nonlinear sampling probability function is based on the click-through rate of the candidate recommendation information. The higher the click-through rate of the candidate recommendation information, the greater the probability of it being sampled. The lower the click-through rate of the candidate recommendation information, the less likely it is to be sampled. Therefore, most of the candidate recommendation information with high click-through rates can be sampled into the new set to ensure the online recommendation effect. Candidate recommendation information in the middle and lower parts of the click-through rate ranking also has a certain probability of being sampled into the new set to achieve effective exploration of user interests.

[0071] In some embodiments, before sampling the first candidate recommendation information set based on the click-through rate of each piece of information in the first candidate recommendation information set, the following technical solution may be implemented: obtaining the user characteristics of the object to be recommended, and obtaining other objects to be recommended whose similarity to the user characteristics exceeds the user similarity threshold; wherein, the other objects to be recommended are users other than the objects to be recommended in the recommendation log; obtaining information corresponding to other objects to be recommended in the candidate recommendation information database, and determining the click-through rate of the object to be recommended for the information; filtering out multiple pieces of information whose click-through rate exceeds the click-through rate threshold from the candidate recommendation information database, and using the multiple pieces of information as candidate recommendation information for the corresponding object to be recommended, so as to form the first candidate recommendation information set.

[0072] In some embodiments, the data volume of the recommendation system is in the millions to tens of millions. Therefore, it is necessary to find a batch of content corresponding to user interests and obtain a set of candidate recommendation information with a data volume of thousands to tens of thousands. This process is called the recall process. The information of the corresponding user is the candidate recommendation information of the corresponding user. The user's interaction behavior log with the candidate recommendation information is divided into two types: rating matrix and exposure click rate. The basic idea of ​​the recall process is to first determine the user similarity, find multiple users closest to the user to be recommended, and then predict the user's rating of the candidate recommendation information based on the ratings of these users. User similarity is the similarity of historical rating behavior, which can also indicate similar interests. The Pearson correlation coefficient is used to determine the similarity between two users. The value of the Pearson correlation coefficient ranges from -1 (strong negative correlation) to 1 (strong positive correlation). The calculation of the Pearson correlation coefficient takes into account the fact that different users have different rating standards. Some users habitually give high scores, while others habitually give low scores.

[0073] In some embodiments, before sampling the first candidate recommendation information set based on the click-through rate of each piece of information in the first candidate recommendation information set, the following technical solution may be performed: obtaining candidate recommendation information features of the candidate recommendation information; predicting the click-through rate of the candidate recommendation information based on the candidate recommendation information features and user features; wherein, the click-through rate prediction model used for prediction is trained with user features, candidate recommendation information features and actual click-through rates in the recommendation log as training samples, with the goal of minimizing the error between the actual click-through rate and the predicted click-through rate.

[0074] In some embodiments, a click-through rate (CTR) prediction model is used to predict the CTR of candidate recommendations. The training samples for this model are obtained through real-time data reporting. User behavior logs can be used to obtain the interaction behavior between users and candidate recommendations. This interaction behavior can be one or more of the following: clicking, saving, forwarding, or commenting. Since saving, forwarding, and commenting are all based on clicks, the CTR is used as the primary indicator. User characteristics, the information characteristics of candidate recommendations, and the actual CTR data of candidate recommendations are used as training samples to positively propagate the CTR prediction model. The specific process is as follows: through clicks... The convolutional layer in the click-through rate (CTR) prediction model acquires the user features of the user to be recommended and the information features of the candidate recommendation information. The user features and information features are fused, and the fused features are reduced in dimensionality by the pooling layer of the CTR prediction model. Then, the features are mapped by the fully connected layer to obtain the predicted CTR result. The predicted result and the actual CTR are substituted into the loss function. The parameter values ​​of the CTR prediction model are updated in reverse according to the result of the loss function to minimize the convergence of the loss function. Finally, the CTR prediction model obtained when the loss function minimizes the convergence is used as the CTR prediction model in the CTR prediction module to predict the click-through rate of the user to be recommended for the recalled candidate recommendation information.

[0075] See Figure 4B , Figure 4B This is an optional flowchart illustrating an AI-based recommendation method provided in an embodiment of the present invention. In step 101, the server samples the first candidate recommendation information set based on the click-through rate of each piece of information in the first candidate recommendation information set. Figure 4B Steps 1011-1013 shown are implemented.

[0076] In step 1011, when the number of information in the first candidate recommendation information set exceeds the threshold for the number of times the sampling function is activated, the server determines the nonlinear sampling probability of the candidate recommendation information based on the nonlinear sampling probability function, using the click rate of the candidate recommendation information as a parameter in the nonlinear sampling probability function.

[0077] In some embodiments, in step 1011, the click rate of the candidate recommendation information is used as a parameter in the nonlinear sampling probability function. The server determines the nonlinear sampling probability of the candidate recommendation information based on the nonlinear sampling probability function. This can be achieved through the following technical solution: the nonlinear sampling probability of the candidate recommendation information is determined according to the following nonlinear sampling probability function (1):

[0078] pro_sampling = 1 / (1+e) -prob_coef*(pctr-prob_bias) (1)

[0079] Where pro_sampling is the non-linear sampling probability of candidate recommendation information, prob_coef and prob_bias are hyperparameters, e is the natural base, and pctr is the click-through rate predicted for candidate recommendation information. The non-linear sampling probability function here is not limited to the above formula. Different coefficients can be set before the base and logarithm to adjust for different data scenarios.

[0080] In some embodiments, the nonlinear sampling probability function samples candidate recommendation information based on the click-through rate. The higher the click-through rate of candidate recommendation information, the greater the probability of it being sampled; the lower the click-through rate of candidate recommendation information, the less likely it is to be sampled. Therefore, most candidate recommendation information with high click-through rates can be sampled into the new set to ensure the effectiveness of online recommendations. Candidate recommendation information with click-through rates in the middle and lower parts also have a certain probability of being sampled into the new set, thus realizing the effective exploration of user interests.

[0081] In step 1012, based on the nonlinear sampling probability of the candidate recommendation information, the server performs random sampling processing on the candidate recommendation information and forms a second candidate recommendation information set with the multiple sampled candidate recommendation information.

[0082] In some embodiments, step 1012, which involves randomly sampling the candidate recommendation information based on the nonlinear sampling probability of the candidate recommendation information and forming a second candidate recommendation information set with the sampled candidate recommendation information, can be implemented using the following technical solution: for each piece of information in the first candidate recommendation information set, the following processing is performed: a random number is generated using a random function, and the nonlinear sampling probability of the information and the random number are proportionally converted to the same numerical range; the candidate recommendation information corresponding to the nonlinear sampling probability greater than the converted random number is determined as the randomly sampled candidate recommendation information and transferred from the first candidate recommendation information set to the second candidate recommendation information set.

[0083] In some embodiments, a random number generation function is invoked to generate a random number. The random number and the nonlinear sampling probability of each piece of information are then commutated to the same numerical range. For example, if the random number generation function generates a random number within the range of 0-100, the nonlinear sampling probability is commutated to 0-100. When the random number generated by the random function is 20 within the range of 0-100 and the nonlinear sampling probability is 40%, the nonlinear sampling probability is commutated to 40. Since the nonlinear sampling probability is greater than the random number 20, the candidate recommendation information corresponding to this nonlinear sampling probability is determined as the candidate recommendation information obtained by random sampling. This candidate recommendation information is then transferred from the first candidate recommendation information set to the second candidate recommendation information set. This transfer means that the candidate recommendation information is deleted from the first candidate recommendation information set and added to the second candidate recommendation information set.

[0084] In step 1013, when the number of information in the first candidate recommendation information set does not exceed the threshold for the number of times the sampling function is activated, the server determines that all candidate recommendation information in the first candidate recommendation information set belongs to the second candidate recommendation information set.

[0085] In some embodiments, when the number of information in the first candidate recommendation information set is less than the threshold for the number of times the sampling function is initiated, that is, the number of recalled information is already insufficient, so there is no need to perform sampling. For example, when the threshold for the number of times the sampling function is initiated is 10,000, but the number of recalled information is less than 10,000, all candidate recommendation information in the first candidate recommendation information set is directly assigned to the second candidate recommendation information set, which is equivalent to skipping the sampling process and directly proceeding to the deduplication process.

[0086] In some embodiments, when the number of candidate recommendation information in the second candidate recommendation information set is less than the sampling output number threshold, the following processing is performed for each piece of information in the first candidate recommendation information set: assign the same sampling probability to each candidate recommendation information in the first candidate recommendation information set; randomly select candidate recommendation information in the first candidate recommendation information set according to the same sampling probability to transfer it to the second candidate recommendation information set until the number of information in the second candidate recommendation information set reaches the sampling output number threshold.

[0087] In some embodiments, the following technical solution can also be implemented: when the number of candidate recommendation information in the second candidate recommendation information set is less than the sampling output number threshold, where the sampling output number threshold is determined based on the sampling rate, for example, if the number of information in the first candidate recommendation information set is 100 and the sampling rate is 50%, then the sampling output number threshold is 50. That is, when the number in the second candidate recommendation information set is less than 50, the following processing is performed on each piece of information in the first candidate recommendation information set. Here, a portion of the first candidate recommendation information has been transferred to the second candidate recommendation information set, and each piece of information in the current first candidate recommendation information set will continue to be sampled. This continued sampling can be done by directly performing random sampling. The process involves assigning the same sampling probability to each candidate recommendation in the first candidate recommendation information set; randomly selecting candidate recommendation information from the first candidate recommendation information set according to the same sampling probability and transferring it to the second candidate recommendation information set, until the number of information in the second candidate recommendation information set reaches the sampling output number threshold. The specific process is as follows: the random number generation function is called again to generate random numbers, assigning the same probability to each candidate recommendation information, and converting the probability and the random number to the same numerical range. Candidate recommendation information with a probability greater than the random number is transferred to the second candidate recommendation information set. The above process is repeated until the number of candidate recommendation information in the second candidate recommendation information set reaches the sampling output number threshold.

[0088] In some embodiments, each piece of information in the current first candidate recommendation information set is sampled again. This sampling can still be based on random sampling based on nonlinear sampling probability. The specific process is as follows: the random number generation function is still called to generate random numbers. According to the nonlinear sampling probability function, a nonlinear sampling probability is generated for each candidate recommendation information. The probability and the random number are still converted to the same numerical range. The candidate recommendation information corresponding to the probability greater than the random number is transferred to the second candidate recommendation information set. The above process is repeated until the number of candidate recommendation information in the second candidate recommendation information set reaches the sampling output number threshold.

[0089] In step 102, the server transfers the information with the highest click rate in the second candidate recommendation information set to the third candidate information set as the first information in the third candidate information set, and continues to transfer some information in the second candidate recommendation information set to the third candidate information set, and the semantic distance between the information transferred in two adjacent transfers is greater than or equal to the semantic distance threshold.

[0090] See Figure 4C , Figure 4CThis is an optional flowchart of an AI-based recommendation method provided in an embodiment of the present invention. In step 102, some information from the second candidate recommendation information set is transferred to the third candidate information set, and the semantic distance between the information transferred in two adjacent transfers is greater than or equal to the semantic distance threshold. Figure 4C Steps 1021-1023 shown are implemented.

[0091] When the number of candidate recommendations in the second candidate recommendation set exceeds the threshold for triggering the deduplication function, and the number of candidate recommendations in the third candidate recommendation set is less than the threshold for triggering the deduplication function, and the second candidate recommendation set is not empty, the deduplication function threshold is the number of information in the third candidate recommendation set, which is the number of information to be obtained after deduplication. When the number of information in the second candidate recommendation set exceeds the number of information in the third candidate recommendation set, it means that deduplication needs to be performed on the second candidate recommendation set. When the number of candidate recommendations in the third candidate recommendation set is less than the threshold for triggering the deduplication function, it means that the number of information in the third candidate recommendation set has not reached the preset value, and deduplication needs to be continued on the second candidate recommendation set. When the above conditions are met, the following processing is performed:

[0092] In step 1021, the server determines the semantic distance between the candidate recommendation with the highest click-through rate in the second candidate recommendation information set and the first information.

[0093] In some embodiments, the semantic distance is obtained based on the semantic features of the two pieces of information. These semantic features can be the latent semantic vectors of the candidate recommendation information, and the format of the latent semantic vectors of the candidate recommendation information is [a1, a2, a3, ..., a...]. n ], a i It is a floating-point number. The length of the latent semantic vector of the candidate recommendation information is specified during matrix factorization or training of the recommendation algorithm based on deep neural networks. In engineering implementation, the latent semantic vector of the candidate recommendation information is synchronized to the server of the mixed sorting module and stored in a hash table in the server memory. Thus, during mixed sorting, the identifier of the candidate recommendation information can be used to quickly look up the corresponding latent semantic vector in the hash table. The semantic distance between two pieces of information is determined by formula (2):

[0094]

[0095] Where, d AB Let a be the semantic distance between candidate recommendation information a and candidate recommendation information b. i and b iThese are the latent semantic vectors of candidate recommendation information a and candidate recommendation information b, respectively.

[0096] In step 1022, when the semantic distance is greater than the semantic distance threshold, the server transfers the candidate recommendation information with the highest click-through rate from the second candidate recommendation information set to the third candidate recommendation information set, and updates it to the first information in the third candidate recommendation information set.

[0097] In some embodiments, a semantic distance greater than a semantic distance threshold indicates a low similarity between two candidate recommendations, meaning that the two candidate recommendations may belong to different primary interests, different secondary interests, or different tags. Here, the candidate recommendation with the highest click-through rate is the information in the current second candidate recommendation set. When the semantic distance is greater than the semantic distance threshold, the server removes the candidate recommendation with the highest click-through rate from the second candidate recommendation set and adds it to the third candidate recommendation set, updating it to the first information in the third candidate recommendation set. The server then continues to determine the semantic distance between the candidate recommendation with the highest click-through rate in the second candidate recommendation set and the first information just updated to the third candidate recommendation set, in order to continue filtering out information from the second candidate recommendation set that is dissimilar to the information previously transferred to the third candidate recommendation set and transferring it to the third candidate recommendation set. Here, similarity and dissimilarity are defined by the semantic distance threshold. When the semantic distance between two pieces of information is greater than the semantic distance threshold, the two pieces of information are determined to be dissimilar. When the semantic distance between two pieces of information is not greater than the semantic distance threshold, the two pieces of information are determined to be similar.

[0098] In step 1023, when the semantic distance is not greater than the semantic distance threshold, the server transfers the candidate recommendation information with the highest click-through rate from the second candidate recommendation information set to the substitute recommendation information set.

[0099] In some embodiments, a semantic distance not greater than a semantic distance threshold indicates that the similarity between two candidate recommendation information is high, meaning that the two candidate recommendation information can belong to the same primary interest, the same secondary interest, or the same tag. Here, the candidate recommendation information with the highest click-through rate is the information in the current second candidate recommendation information set. When the semantic distance is not greater than the semantic distance threshold, the server will delete the candidate recommendation information with the highest click-through rate from the second candidate recommendation information set and add it to the backup recommendation information set. Here, the backup recommendation information set is a standby set.

[0100] In some embodiments, the following technical solution may also be implemented: when the number of information in the third candidate recommendation information set is less than the threshold for starting the deduplication function, and the second candidate recommendation information set is empty, the following processing is performed on the substitute recommendation information set: assign the same sampling probability to each piece of information in the substitute recommendation information set; randomly extract information from the substitute recommendation information set according to the same sampling probability to transfer it to the third candidate recommendation information set until the number of candidate recommendation information in the third candidate recommendation information set reaches the threshold for starting the deduplication function.

[0101] In some embodiments, the threshold for initiating the deduplication function is actually a pre-set number of elements that should be in the third candidate recommendation information set after deduplication. When the number of candidate recommendation information in the second candidate recommendation information set does not exceed the threshold for initiating the deduplication function, all candidate recommendation information in the second candidate recommendation information set is transferred to the third candidate recommendation information set, which is equivalent to omitting the deduplication process. When the number of information in the third candidate recommendation information set is less than the threshold for initiating the deduplication function and the second candidate recommendation information set is empty, it indicates that the number of information in the third candidate recommendation information set has not reached the pre-set number, but the second candidate recommendation information set is already an empty set. At this time, the information in the substitute recommendation information set is transferred to the third candidate recommendation information set.

[0102] In some embodiments, the process of randomly selecting information from the backup recommendation information set with the same sampling probability and transferring it to the third candidate recommendation information set until the number of candidate recommendation information in the third candidate recommendation information set reaches the threshold for triggering the deduplication function can refer to the implementation of the sampling process described above. That is, assigning the same sampling probability to each candidate recommendation information in the backup candidate recommendation information set; randomly selecting candidate recommendation information from the backup candidate recommendation information set with the same sampling probability and transferring it to the third candidate recommendation information set until the number of information in the third candidate recommendation information set reaches the threshold for triggering the deduplication function. The specific process is as follows: still calling the random number generation function to generate random numbers, assigning the same probability to each candidate recommendation information, still converting the probability and the random number to the same numerical range, transferring the candidate recommendation information corresponding to the probability greater than the random number to the third candidate recommendation information set, and continuously repeating the above process until the number of candidate recommendation information in the third candidate recommendation information set reaches the threshold for triggering the deduplication function.

[0103] By adopting the above method, we can quickly achieve deduplication of candidate recommendation information and efficiently filter out combinations of candidate recommendation information covering different user interests at a finer granularity to optimize the recommendation effect. The semantic distance value determines the output result of deduplication, and an appropriate value needs to be selected according to the specific application scenario.

[0104] In step 103, the server sorts the information in the third candidate recommendation information set and performs recommendation operations based on the multiple pieces of information included in the sorting results.

[0105] See Figure 4D , Figure 4D This is an optional flowchart of an artificial intelligence-based recommendation method provided in an embodiment of the present invention. In step 103, the information in the third candidate recommendation information set is sorted, which can be achieved through... Figure 4D Steps 1031-1032 shown are implemented.

[0106] In step 1031, the server removes the candidate recommendation with the highest click rate from the third candidate recommendation information set and determines it as the candidate recommendation information ranked first.

[0107] In step 1032, when the third candidate recommendation information set is not empty, the server performs the following processing for each piece of information in the third candidate recommendation information set:

[0108] In step 10321, the server determines the ranking score of the candidate recommendation information based on the click-through rate of the candidate recommendation information in the third candidate recommendation information set and the semantic distance between the candidate recommendation information in the third candidate recommendation information set and the candidate recommendation information ranked first.

[0109] The semantic distance here is the semantic distance between the candidate recommendation information in the third candidate recommendation information set and the candidate recommendation information ranked first, and the candidate recommendation information in the third candidate recommendation information set refers to each candidate recommendation information in the third candidate recommendation information set.

[0110] In some embodiments, the ranking score of the candidate recommendation information in step 10321, based on the click-through rate of the candidate recommendation information in the third candidate recommendation information set and the semantic distance between the candidate recommendation information in the third candidate recommendation information set and the candidate recommendation information ranked first, can be achieved through the following technical solution: obtaining the semantic feature distance between the candidate recommendation information and the candidate recommendation information ranked first; determining the ranking score of the candidate recommendation information based on the scoring function (3):

[0111] score=pctr A +w s *1 / ed AB (3)

[0112] Among them, pctr A For the click-through rate of candidate recommendation information, w s For hyperparameters, dAB This represents the semantic distance between candidate recommendations and the candidate recommendation ranked first.

[0113] Here is pctr A The coefficient between the base and the base can also be adjusted to adapt to different data scenarios.

[0114] In step 10322, the server determines the candidate recommendation with the highest ranking score as the next candidate recommendation and removes it from the third candidate recommendation set.

[0115] In some embodiments, the candidate recommendation with the highest click rate in the third candidate recommendation information set is first removed from the third candidate recommendation information set and placed at the top. Then, the semantic distance between each piece of information in the current third candidate recommendation information set and the previously removed top-ranked information is determined. The semantic distance and the click rate of each piece of information in the current third candidate recommendation information set are substituted into the scoring function (3) to obtain the score of each piece of information in the current third candidate recommendation information set. The information with the highest score is removed from the third candidate recommendation information set and placed at the next position as the candidate recommendation information. This process is repeated until the third candidate recommendation information set is empty, thus determining the ranking position of each piece of information in each third candidate recommendation information set.

[0116] See Figure 4E , Figure 4E This is an overall interactive flowchart of the recommendation method based on artificial intelligence provided in the embodiments of the present invention, which will be explained below in conjunction with steps 201-208.

[0117] In step 201, the terminal reports the interactive behavior of the user to be recommended in response to the candidate recommendation information to the server as training sample data and user characteristics of the user to be recommended.

[0118] The training sample data here comes from the behavioral data of different users to be recommended reported by various terminals. The click-through rate prediction model is trained based on this behavioral data. The user features are obtained from the terminals of the users to be recommended.

[0119] In step 202, the server obtains multiple candidate recommendation information from the candidate recommendation information database based on user characteristics, forming a first candidate recommendation information set.

[0120] In step 203, the server predicts the click-through rate of candidate recommendation information based on user characteristics using a click-through rate prediction model.

[0121] In step 204, the server samples the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, and forms a second candidate recommendation information set from the sampled information.

[0122] In step 205, the server transfers the information with the highest click rate in the second candidate recommendation information set to the third candidate information set as the first information in the third candidate information set, and continues to transfer some information in the second candidate recommendation information set to the third candidate information set, and the semantic distance between the information transferred in two adjacent transfers is greater than or equal to the semantic distance threshold.

[0123] In step 206, the server sorts the information in the third candidate recommendation information set.

[0124] In step 207, the server performs a recommendation operation to the terminal based on multiple pieces of information included in the sorting results.

[0125] In step 208, the terminal presents the received information.

[0126] The following will describe an exemplary application of the embodiments of the present invention in the news recommendation scenario.

[0127] See Figure 5 , Figure 5 This is a flowchart illustrating a practical application scenario of the AI-based recommendation method provided in this invention. The recommendation system for the news client includes candidate content generation, click-through rate prediction, and mixed ranking components, etc. Figure 5 As shown, the candidate content generation part uses different recommendation algorithms, such as item-to-item (I2I), user-based collaborative filtering (UCF), and item-based collaborative filtering (ICF), to quickly filter 1,000-2,000 user-related content items from the content database and submit them to the click-through rate (CTR) prediction part. The CTR prediction part uses a CTR prediction ordering model to predict and rank the content CTRs and then submits them to the blending part. The blending part selects a specified number of items from the ranked content according to manual rules and rearranges them for recommendation to users. In the news client, user logs such as content exposure and clicks are reported to the recommendation system's log system in real time. The user logs reported by the client can be used to generate real-time user and content features and model training samples.

[0128] This invention provides an artificial intelligence-based recommendation method, the core of which is a hybrid ranking recommendation algorithm based on probability sampling and similarity. It includes nonlinear probability sampling, a similarity deduplication algorithm based on the latent vector of candidate recommendation information, and the generation of hybrid ranking results. Each part will be described in detail below.

[0129] In nonlinear probability sampling, the candidate recommendation information output by the click-through rate prediction module is first sampled nonlinearly to generate a new set of candidate recommendation information. The nonlinear sampling probability function is based on the click-through rate of the candidate recommendation information, as shown in Formula 1. The higher the click-through rate of the candidate recommendation information, the greater the probability of it being sampled, and the lower the click-through rate of the candidate recommendation information, the less likely it is to be sampled. Therefore, most of the candidate recommendation information with high click-through rates can be sampled into the new set to ensure the online recommendation effect. Candidate recommendation information with click-through rates in the middle and lower parts also have a certain probability of being sampled into the new set, realizing the effective exploration of user interests.

[0130] pro_sampling = 1 / (1+e) -prob_coef*(pctr-prob_bias) (1)

[0131] Among them, prob_coef and prob_bias are hyperparameters, and appropriate values ​​need to be selected according to specific application requirements. By using different combinations of hyperparameters, the proportion of candidate recommendation information with different click rates in the newly generated set can be adjusted.

[0132] Assume the input candidate recommendation information set for the click-through rate prediction module is input_set, and the number of candidate recommendation information is C. input The minimum number of candidate recommendation information C required to activate the probability sampling function. min The probability sampling ratio is not less than r; the output set of probability sampling is defined as output_set, and the specific steps of probability sampling are as follows:

[0133] Step 1: Compare the number C of candidate recommendation information in the input set input_set. input and C min If C input >C min If the probability sampling function is activated, then the probability sampling function will be activated; otherwise, the probability sampling function will not be activated, and the current probability sampling will be completed.

[0134] Step 2: For the candidate recommendation information contained in the input_set, calculate the probability of it being non-linearly sampled based on the click rate of the candidate recommendation information using Formula 1.

[0135] Step 3: Based on the probability of each candidate recommendation being sampled, a random method is used to determine whether it should be sampled. If the candidate recommendation is sampled, it is added to the output set output_set and removed from the input_set.

[0136] Step 4: Calculate the number C of candidate recommendation information in the output_set. output If C output <C i nput *r, then randomly select C from the set input_set input *r–C output Each candidate recommendation is added to the set output_set. If C output Greater than or equal to C input *r, complete this probability sampling.

[0137] The following describes the similarity deduplication process based on the latent semantic vectors of candidate recommendation information. After completing the nonlinear probability sampling of candidate recommendation information, a new set of candidate recommendation information will be generated. In order to perform fine-grained adaptive deduplication of candidate recommendation information and better cover users' interests, a similarity calculation method based on the latent semantics of candidate recommendation information will be used. The latent vectors of candidate recommendation information will be obtained directly from matrix factorization or DNN-based recommendation algorithms. The format of the latent semantic vectors of candidate recommendation information is [a1, a2, a3, ..., an], where ai is a floating-point number. The length of the latent semantic vectors of candidate recommendation information is specified during the training of matrix factorization or DNN-based recommendation algorithms. In the engineering implementation, the latent semantic vectors of candidate recommendation information are synchronized to the server of the mixed sorting module and stored in a hash table in the server's memory. Thus, during mixed sorting, the identifier of the candidate recommendation information can be used to quickly look up the corresponding latent semantic vector in the hash table.

[0138] The adaptive deduplication method based on the latent semantic vectors of candidate recommendation information is as follows: Define an input set S, initialize the candidate recommendation information set output by nonlinear probability sampling, and the number of elements in the set is N; define an output set T, initialized as an empty set, and the number of candidate recommendation information in set T cannot exceed M; define a substitute recommendation information set R, initialized as empty; define the formula for calculating the latent vector distance between candidate recommendation information A and candidate recommendation information B, as shown in Formula 2:

[0139]

[0140] Where, d AB Let a be the semantic distance between candidate recommendation information a and candidate recommendation information b. i and b i These are the latent semantic vectors of candidate recommendation information a and candidate recommendation information b, respectively.

[0141] Define the semantic distance threshold d of the candidate recommendation information. As a hyperparameter, it can be adjusted according to specific requirements. The specific processing process is as follows:

[0142] Step 1: If the number of elements in set S is less than or equal to M, put all the elements of set S into set T, clear set S, and stop the processing flow. If the number of elements in set S is greater than M, go to Step 2.

[0143] Step 2: Select the candidate recommendation information with the highest click-through rate in set S (marked as S1), add it to set T as the first element (the first piece of information) of set T, marked as T1. Delete S1 from set S, and the original S2 ranked second in the click-through rate ranking becomes the new S1.

[0144] Step 3: Use formula 2 to calculate the distance between S1 and T1. If the distance between the two is less than or equal to d, then delete S1 from set S, add it to set R, and then calculate the distance between the new S1 and T1, and iterate in turn. If the distance between the two is greater than d, then add S1 to set T, and delete S1 from set S.

[0145] Step 4: Assume that set T already contains i elements (i < M), there are j elements remaining in set S, and set R contains N - i - j elements. When calculating the (i + 1)-th element of set T, calculate the distance between Ti and S1. If the distance between the two is less than or equal to d, delete S1 from set S, add it to set R, and then calculate the distance between the new S1 and Ti, and iterate in turn. If the distance between the two is greater than d, then add S1 to set T as the (i + 1)-th element, and delete S1 from set S.

[0146] Step 5: If the number of elements in set T is less than M and the elements in set S are not empty, repeat Step 4 iteratively. If the number of elements in set T is equal to M, stop the processing flow. If the elements in set S are empty and the number of elements in T is less than M, go to Step 6. count <M, and the elements in set S are not empty, repeat Step 4 iteratively. If the number of elements in set T is less than M and the elements in set S are not empty, repeat Step 4 iteratively. If the number of elements in set T is equal to M, stop the processing flow. If the elements in set S are empty and the number of elements in T is less than M, go to Step 6. count equals M, stop the processing flow. If the elements in set S are empty and the number of elements in T is less than M, go to Step 6.

[0147] Step 6: Randomly select M - T count pieces of candidate recommendation information from each candidate recommendation information in set R with the same probability, put them into set T, and stop the processing flow.

[0148] By adopting the above method, the similarity de-duplication of the input candidate recommendation information can be quickly achieved, and the candidate recommendation information combinations covering different user interest points can be efficiently screened at a finer granularity to optimize the recommendation effect. Among them, the value of the hyperparameter d determines the output result of the mixed arrangement, and a suitable value needs to be selected according to the specific application scenario.

[0149] The following describes the sorting process. After generating set T, the elements in T are sorted according to dimensions such as the click-through rate of candidate recommendations and the similarity between candidate recommendations. The candidate recommendation with the highest click-through rate in set T is selected as the output sequence S. output The first candidate recommendation is selected. Set T removes this candidate recommendation. Assuming the output sequence already contains i candidate recommendations and set T contains Mi elements, the method for calculating the (i+1)th element of the output sequence is as follows: Using each candidate recommendation in set T (denoted as A) and the i-th candidate recommendation in the output sequence (denoted as B), calculate the score using Formula 3. Select the candidate recommendation with the highest score as the (i+1)-th candidate recommendation in the output sequence, and remove this candidate recommendation from set T. This process is iterated in this way:

[0150] score=pctr A +w s *1 / ed AB (3)

[0151] Among them, pctr A For the click-through rate of candidate recommendation information, w s For hyperparameters, d AB The semantic distance between candidate recommendations and the top-ranked candidate recommendation needs to be selected according to specific requirements. Using the AI-based recommendation method provided in this invention, news items are recommended based on the ranking results. Since scores and click-through rates are positively correlated, the recommended news items align with the user's interests and fall within the scope of information the user is interested in. Furthermore, before ranking based on scores, deduplication has been performed based on the semantic distance threshold. Therefore, the recommended news items do not have strong similarities; for example, the recommended news items may belong to different primary or secondary interests, thus achieving diversified recommendations based on user interests.

[0152] The AI-based recommendation method provided in this invention is an adaptive mixed-rank recommendation algorithm based on probability sampling and similarity. Compared with the mixed-ranking method based on manual rules commonly used in related technologies, its innovation lies in two aspects: First, it uses a nonlinear probability sampling function to sample the candidate recommendation information sequence input by the ranking module, which can better explore user interests while ensuring recommendation effect; Second, it uses a method based on latent semantic similarity to automatically deduplicate and mix-rank the candidate recommendation information set obtained by probability sampling, which can better cover user interests at a fine-grained level. Compared with the mixed-ranking recommendation method based on manual rules, the AI-based recommendation method provided in this invention not only has the above advantages, but also has the characteristics of being simple and efficient, not requiring cumbersome rule combinations and trials, and being adaptable to different application scenarios.

[0153] The AI-based recommendation method provided in this invention is a hybrid ranking solution based on probability sampling and similarity. It can achieve fine-grained and efficient automatic hybrid ranking recommendations. Compared with the hybrid ranking methods based on manual rules commonly used in related technologies, the AI-based recommendation method provided in this invention is simple, efficient, and can avoid a large amount of manual rule combination and trial work. It is suitable for different application scenarios. It was launched on the recommendation channel of a news client and A / B testing was conducted. Compared with hybrid ranking solutions in related technologies, its click-through rate and other indicators are positive.

[0154] The following description continues to illustrate the exemplary structure of the AI-based recommendation device 255 provided in the embodiments of the present invention as a software module. In some embodiments, such as... Figure 3 As shown, the software modules stored in the AI-based recommendation method apparatus 255 in the memory 250 may include: an information sampling module 2553, used to sample the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, and form a second candidate recommendation information set from the sampled information; an information transfer module 2554, used to transfer the information with the highest click rate in the second candidate recommendation information set to the third candidate information set as the first piece of information in the third candidate information set, and continue to transfer some information in the second candidate recommendation information set to the third candidate information set, wherein the semantic distance between adjacent transferred information is greater than or equal to a semantic distance threshold; and an information sorting module 2555, used to sort the information in the third candidate recommendation information set, and perform recommendation operations based on the multiple pieces of information included in the sorting results.

[0155] In the above scheme, the AI-based recommendation device 255 further includes: an information generation module 2551, used to: acquire user characteristics of the object to be recommended, and acquire other objects to be recommended whose similarity to the user characteristics exceeds a user similarity threshold; wherein, other objects to be recommended are users other than the objects to be recommended in the recommendation log; acquire information corresponding to other objects to be recommended in the candidate recommendation information database, and determine the click rate of the object to be recommended for the information; filter out multiple pieces of information whose click rate exceeds the click rate threshold from the candidate recommendation information database, and use the multiple pieces of information as candidate recommendation information for the corresponding object to be recommended, so as to form a first set of candidate recommendation information.

[0156] In the above scheme, the AI-based recommendation device 255 further includes: a click-through rate prediction module 2552, which is also used to obtain candidate recommendation information features of candidate recommendation information; predict the click-through rate of candidate recommendation information based on candidate recommendation information features and user features; wherein, the click-through rate prediction model used for prediction is trained with user features, candidate recommendation information features and actual click-through rate in recommendation log as training samples, and with the goal of minimizing the error between actual click-through rate and predicted click-through rate.

[0157] In the above scheme, the information sampling module 2553 is further configured to: when the number of information in the first candidate recommendation information set exceeds the threshold for the number of times the sampling function is activated, determine the nonlinear sampling probability of the candidate recommendation information based on the nonlinear sampling probability function, using the click rate of the candidate recommendation information as a parameter in the nonlinear sampling probability function; perform random sampling processing on the candidate recommendation information based on the nonlinear sampling probability of the candidate recommendation information, and form a second candidate recommendation information set with the multiple candidate recommendation information obtained from the sampling; when the number of information in the first candidate recommendation information set does not exceed the threshold for the number of times the sampling function is activated, determine that all candidate recommendation information in the first candidate recommendation information set belongs to the second candidate recommendation information set.

[0158] In the above scheme, the information sampling module 2553 is further used to: determine the nonlinear sampling probability of the candidate recommendation information according to the following nonlinear sampling probability function:

[0159] pro_sampling=1 / (1+e -prob_coef*(pctr-prob_bias) )

[0160] pro_sampling is the non-linear sampling probability of candidate recommendation information, prob_coef and prob_bias are hyperparameters, e is the natural base, and pctr is the predicted click-through rate for candidate recommendation information.

[0161] In the above scheme, the information sampling module 2553 is further configured to: perform the following processing on each piece of information in the first candidate recommendation information set: generate a random number through a random function, and proportionally convert the nonlinear sampling probability of the information to the same numerical range as the random number; determine the candidate recommendation information corresponding to the nonlinear sampling probability greater than the converted random number as the candidate recommendation information obtained by random sampling, and transfer it from the first candidate recommendation information set to the second candidate recommendation information set.

[0162] In the above scheme, the information sampling module 2553 is further configured to: when the number of candidate recommendation information in the second candidate recommendation information set is less than the sampling output number threshold, perform the following processing for each piece of information in the first candidate recommendation information set: assign the same sampling probability to each candidate recommendation information in the first candidate recommendation information set; randomly extract candidate recommendation information in the first candidate recommendation information set according to the same sampling probability to transfer it to the second candidate recommendation information set until the number of information in the second candidate recommendation information set reaches the sampling output number threshold.

[0163] In the above scheme, the information transfer module 2554 is further configured to: when the number of candidate recommendation information in the second candidate recommendation information set exceeds the threshold for initiating the deduplication function, the number of candidate recommendation information in the third candidate recommendation information set is less than the threshold for initiating the deduplication function, and the second candidate recommendation information set is not empty, perform the following processing on the candidate recommendation information with the highest click-through rate in the second candidate recommendation information set: determine the semantic distance between the information with the highest click-through rate and the first information in the second candidate recommendation information set; when the semantic distance is greater than the semantic distance threshold, transfer the candidate recommendation information with the highest click-through rate from the second candidate recommendation information set to the third candidate recommendation information set and update it to the first information in the third candidate recommendation information set; when the semantic distance is not greater than the semantic distance threshold, transfer the candidate recommendation information with the highest click-through rate from the second candidate recommendation information set to the substitute recommendation information set.

[0164] In the above scheme, the information transfer module 2554 is also used to: transfer all candidate recommendation information in the second candidate recommendation information set to the third candidate recommendation information set when the number of candidate recommendation information in the second candidate recommendation information set does not exceed the threshold for starting the deduplication function.

[0165] In the above scheme, the information transfer module 2554 is further configured to: when the number of information in the third candidate recommendation information set is less than the threshold for starting the deduplication function, and the second candidate recommendation information set is empty, perform the following processing on the substitute recommendation information set: assign the same sampling probability to each piece of information in the substitute recommendation information set; randomly extract information from the substitute recommendation information set according to the same sampling probability to transfer it to the third candidate recommendation information set, until the number of candidate recommendation information in the third candidate recommendation information set reaches the threshold for starting the deduplication function.

[0166] In the above scheme, the information ranking module 2555 is further configured to: remove the candidate recommendation information with the highest click-through rate from the third candidate recommendation information set and determine it as the candidate recommendation information ranked first; when the third candidate recommendation information set is not empty, perform the following processing for each piece of information in the third candidate recommendation information set: determine the ranking score of the candidate recommendation information based on the click-through rate of the candidate recommendation information in the third candidate recommendation information set and the semantic distance between it and the candidate recommendation information ranked first; determine the candidate recommendation information with the highest ranking score as the candidate recommendation information ranked next and remove it from the third candidate recommendation information set.

[0167] In the above scheme, the information ranking module 2555 is also used to: obtain the semantic feature distance between the information and the candidate recommendation information ranked first; and determine the ranking score of the candidate recommendation information based on the scoring function.

[0168] score=pctr A +w s *1 / ed AB ;

[0169] Among them, pctr A For the click-through rate of candidate recommendation information, w s For hyperparameters, d AB This represents the semantic distance between the information and the candidate recommendation information ranked first.

[0170] This invention provides a 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 invention, for example... Figures 4A-4E The AI-based recommendation method is shown.

[0171] In some embodiments, the storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEP ROM, 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.

[0172] 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.

[0173] 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 collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0174] 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.

[0175] In summary, by performing multiple information transfers on candidate recommendation information through the embodiments of the present invention, and ensuring that the semantic distance between the information transferred each time is greater than the semantic distance threshold, the duplication rate of candidate recommendation information is low and the coverage is wide. The candidate recommendation information after information transfer is then sorted to obtain fine-grained sorting results.

[0176] The above are merely embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based recommendation method, characterized in that, The method includes: Based on the click-through rate of each piece of information in the first candidate recommendation information set, non-linear probability sampling is performed on the first candidate recommendation information set, and The sampled information is used to form a second candidate recommendation information set; wherein, the nonlinear probability sampling is used to make the information with a higher click rate in the first candidate recommendation information set more likely to be sampled; The information with the highest click-through rate in the second candidate recommendation information set is transferred to the third candidate information set, and becomes the first piece of information in the third candidate information set. Continue to transfer some information from the second candidate recommendation information set to the third candidate information set, and the semantic distance between the information transferred in two adjacent transfers is greater than or equal to the semantic distance threshold; The information in the third candidate recommendation information set is sorted, and Perform recommendation operations based on multiple pieces of information included in the ranking results; The step of performing nonlinear probability sampling on the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set includes: When the number of information in the first candidate recommendation information set exceeds the threshold for the number of times the sampling function is activated, the nonlinear sampling probability of the information is determined based on the nonlinear sampling probability function, using the click rate of the information as a parameter in the nonlinear sampling probability function. Based on the nonlinear sampling probability of the information, the information is randomly sampled, and the multiple sampled information are used to form the second candidate recommendation information set. When the number of information in the first candidate recommendation information set does not exceed the threshold of the number of times the sampling function is activated, it is determined that all information in the first candidate recommendation information set belongs to the second candidate recommendation information set.

2. The method according to claim 1, characterized in that, Before performing nonlinear probability sampling on the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, the method further includes: Obtain the user characteristics of the objects to be recommended, and obtain other objects to be recommended whose similarity to the user characteristics exceeds the user similarity threshold; Among them, the other objects to be recommended are users in the recommendation log other than the objects to be recommended; Obtain information from the candidate recommendation information database corresponding to other objects to be recommended, and determine the click-through rate of the objects to be recommended in relation to the information; Multiple pieces of information with click-through rates exceeding a click-through rate threshold are selected from the candidate recommendation information database, and these multiple pieces of information with click-through rates exceeding the click-through rate threshold are used as information corresponding to the object to be recommended, so as to form a first candidate recommendation information set.

3. The method according to claim 1, characterized in that, Before performing nonlinear probability sampling on the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, the method further includes: Obtain the candidate recommendation information features of the information; Predict the click-through rate of the information based on the candidate recommendation information features and user features; The click-through rate prediction model used in the prediction is trained using user features, candidate recommendation information features, and the actual click-through rate in the recommendation log as training samples, with the goal of minimizing the error between the actual click-through rate and the predicted click-through rate.

4. The method according to claim 1, characterized in that, The semantic distance between two adjacent transfers is obtained based on the semantic features of the information in the two adjacent transfers, where the semantic features are the latent semantic vectors of the information.

5. The method according to claim 1, characterized in that, The step of determining the nonlinear sampling probability of the information based on the nonlinear sampling probability function, using the click rate of the information as a parameter, includes: The nonlinear sampling probability of the information is determined based on the following nonlinear sampling probability function: pro_sampling=1 / (1+e -prob_coef*(pctr-prob_bias) ) The pro_sampling is the non-linear sampling probability of the information, the prob_coef and the prob_bias are hyperparameters, e is the natural base, and pctr is the predicted click-through rate for the information.

6. The method according to claim 1, characterized in that, The nonlinear sampling probability based on the information is used to randomly sample the information, and the sampled information is used to form the second candidate recommendation information set, including: For each piece of information in the first set of candidate recommendation information, perform the following processing: Random numbers are generated using a random function, and the nonlinear sampling probability of the information is proportionally converted to the same numerical range as the random numbers. Information corresponding to nonlinear sampling probabilities greater than the converted random number is identified as randomly sampled information and transferred from the first candidate recommendation information set to the second candidate recommendation information set.

7. The method according to claim 6, characterized in that, The method further includes: When the number of pieces of information in the second candidate recommendation information set is less than the sampling output number threshold, the following processing is performed on each piece of information in the first candidate recommendation information set: Assign the same sampling probability to each piece of information in the first candidate recommendation information set; Information is randomly extracted from the first candidate recommendation information set with the same sampling probability and transferred to the second candidate recommendation information set until the number of information in the second candidate recommendation information set reaches the sampling output number threshold.

8. The method according to claim 1, characterized in that, The step of transferring a portion of the information from the second candidate recommendation information set to the third candidate information set, wherein the semantic distance between two adjacent transfers is greater than or equal to a semantic distance threshold, includes: When the number of information in the second candidate recommendation information set exceeds the threshold for triggering the deduplication function, the number of information in the third candidate recommendation information set is less than the threshold for triggering the deduplication function, and the second candidate recommendation information set is not empty, the following processing is performed on the information with the highest click-through rate in the second candidate recommendation information set: Determine the semantic distance between the information with the highest click-through rate in the second candidate recommendation information set and the first information; When the semantic distance is greater than the semantic distance threshold, the information with the highest click-through rate is transferred from the second candidate recommendation information set to the third candidate recommendation information set and updated to the first information in the third candidate recommendation information set; When the semantic distance is not greater than the semantic distance threshold, the information with the highest click-through rate is transferred from the second candidate recommendation information set to the alternative recommendation information set.

9. The method according to claim 8, characterized in that, The method further includes: When the number of information in the second candidate recommendation information set does not exceed the threshold for activating the deduplication function, all information in the second candidate recommendation information set is transferred to the third candidate recommendation information set.

10. The method according to claim 8, characterized in that, The method further includes: When the number of information in the third candidate recommendation information set is less than the threshold for initiating the deduplication function, and the second candidate recommendation information set is empty, the following processing is performed on the substitute recommendation information set: Assign the same sampling probability to each piece of information in the set of alternative recommendation information; Information from the substitute recommendation information set is randomly sampled with the same sampling probability and transferred to the third candidate recommendation information set until the number of information in the third candidate recommendation information set reaches the threshold for activating the deduplication function.

11. The method according to claim 1, characterized in that, The sorting process for the information in the third candidate recommendation information set includes: The information with the highest click rate in the third candidate recommendation information set is removed from the third candidate recommendation information set and determined as the information ranked first in the ranking result; When the third candidate recommendation information set is not empty, the following processing is performed for each piece of information in the third candidate recommendation information set: The ranking score of the information is determined based on the click-through rate of the information in the third candidate recommendation information set and the semantic distance between the information and the information ranked first. The information with the highest ranking score is determined as the next highest-ranked information in the ranking results and removed from the third candidate recommendation information set.

12. The method according to claim 11, characterized in that, The process of determining the ranking score of the information based on the click-through rate of the information in the third candidate recommendation information set and the semantic distance between the information and the information ranked first includes: Obtain the semantic feature distance between the information and the information ranked first; The ranking score of the information is determined based on the following scoring function: score=pctr A +w s *1 / ed AB ; Wherein, the pctr A The click-through rate of the information, w s As a hyperparameter, the d AB The semantic distance between the information and the information ranked first.

13. A recommendation device based on artificial intelligence, characterized in that, The device includes: The information sampling module is used to perform nonlinear probability sampling on the first candidate recommendation information set based on the click rate of each piece of information in the first candidate recommendation information set, and to form a second candidate recommendation information set from the sampled information; wherein, the nonlinear probability sampling is used to make the information with a higher click rate in the first candidate recommendation information set more likely to be sampled; The information transfer module is used to transfer the information with the highest click rate in the second candidate recommendation information set to the third candidate information set as the first information in the third candidate information set, and to continue to transfer some information in the second candidate recommendation information set to the third candidate information set, and the semantic distance between the information transferred in two adjacent transfers is greater than or equal to the semantic distance threshold. The information sorting module is used to sort the information in the third candidate recommendation information set and perform recommendation operations based on the multiple pieces of information included in the sorting result. The step of performing nonlinear probability sampling on the first candidate recommendation information set based on the click-through rate of each piece of information in the first candidate recommendation information set includes: when the number of pieces of information in the first candidate recommendation information set exceeds the threshold for activating the sampling function, determining the nonlinear sampling probability of the information based on the nonlinear sampling probability function, using the click-through rate of the information as a parameter in the nonlinear sampling probability function; performing random sampling processing on the information based on the nonlinear sampling probability of the information, and forming a second candidate recommendation information set with the multiple pieces of information obtained from the sampling; and determining that all information in the first candidate recommendation information set belongs to the second candidate recommendation information set when the number of pieces of information in the first candidate recommendation information set does not exceed the threshold for activating the sampling function.

14. The apparatus according to claim 13, characterized in that, The AI-based recommendation device also includes: The information generation module is configured to: acquire user characteristics of the object to be recommended, and acquire other objects to be recommended whose similarity to the user characteristics exceeds a user similarity threshold; wherein, the other objects to be recommended are users other than the object to be recommended in the recommendation log; acquire information corresponding to the other objects to be recommended in the candidate recommendation information database, and determine the click-through rate of the object to be recommended for the information; filter out multiple pieces of information whose click-through rate exceeds a click-through rate threshold from the candidate recommendation information database, and use the multiple pieces of information as candidate recommendation information corresponding to the object to be recommended, so as to form a first candidate recommendation information set.

15. The apparatus according to claim 13, characterized in that, The AI-based recommendation device also includes: The click-through rate (CTR) prediction module is further used to obtain candidate recommendation information features of the information; predict the CTR of the information based on the candidate recommendation information features and user features; wherein the CTR prediction model used for the prediction is trained using user features, candidate recommendation information features and actual CTR in the recommendation log as training samples, with the goal of minimizing the error between the actual CTR and the predicted CTR.

16. The apparatus according to claim 13, characterized in that, The semantic distance between two adjacent transfers is obtained based on the semantic features of the information in the two adjacent transfers, where the semantic features are the latent semantic vectors of the information.

17. The apparatus according to claim 13, characterized in that, The information sampling module is further configured to: determine the nonlinear sampling probability of the information according to the following nonlinear sampling probability function: pro_sampling=1 / (1+e -prob_coef*(pctr-prob_bias) ) The pro_sampling is the non-linear sampling probability of the information, the prob_coef and the prob_bias are hyperparameters, e is the natural base, and pctr is the predicted click-through rate for the information.

18. The apparatus according to claim 13, characterized in that, The information sampling module is further configured to: perform the following processing on each piece of information in the first candidate recommendation information set: generate a random number through a random function, and proportionally convert the nonlinear sampling probability of the information to the same numerical range as the random number; determine the information corresponding to the nonlinear sampling probability greater than the converted random number as randomly sampled information, and transfer it from the first candidate recommendation information set to the second candidate recommendation information set.

19. The apparatus according to claim 18, characterized in that, The information sampling module is further configured to: when the number of information in the second candidate recommendation information set is less than the sampling output number threshold, perform the following processing for each piece of information in the first candidate recommendation information set: assign the same sampling probability to each piece of information in the first candidate recommendation information set; randomly extract information from the first candidate recommendation information set according to the same sampling probability to transfer it to the second candidate recommendation information set until the number of information in the second candidate recommendation information set reaches the sampling output number threshold.

20. The apparatus according to claim 13, characterized in that, The information transfer module is further configured to: when the number of information in the second candidate recommendation information set exceeds the threshold for activating the deduplication function, the number of information in the third candidate recommendation information set is less than the threshold for activating the deduplication function, and the second candidate recommendation information set is not empty, perform the following processing on the information with the highest click-through rate in the second candidate recommendation information set: determine the semantic distance between the information with the highest click-through rate in the second candidate recommendation information set and the first information; when the semantic distance is greater than the semantic distance threshold, transfer the information with the highest click-through rate from the second candidate recommendation information set to the third candidate recommendation information set, and update it to the first information in the third candidate recommendation information set; When the semantic distance is not greater than the semantic distance threshold, the information with the highest click-through rate is transferred from the second candidate recommendation information set to the alternative recommendation information set.

21. The apparatus according to claim 20, characterized in that, The information transfer module is further configured to: transfer all information in the second candidate recommendation information set to the third candidate recommendation information set when the number of information in the second candidate recommendation information set does not exceed the threshold for activating the deduplication function.

22. The apparatus according to claim 20, characterized in that, The information transfer module is further configured to: when the number of information in the third candidate recommendation information set is less than the threshold for initiating the deduplication function, and the second candidate recommendation information set is empty, perform the following processing on the substitute recommendation information set: assign the same sampling probability to each piece of information in the substitute recommendation information set; randomly extract information from the substitute recommendation information set according to the same sampling probability to transfer it to the third candidate recommendation information set, until the number of information in the third candidate recommendation information set reaches the threshold for initiating the deduplication function.

23. The apparatus according to claim 13, characterized in that, The information sorting module is further configured to: remove the information with the highest click rate from the third candidate recommendation information set and determine it as the information ranked first; When the third candidate recommendation information set is not empty, the following processing is performed for each piece of information in the third candidate recommendation information set: based on the click-through rate of the information in the third candidate recommendation information set and the semantic distance between it and the information ranked first, the ranking score of the information is determined; the information with the highest ranking score is determined as the next ranked information and removed from the third candidate recommendation information set.

24. The apparatus according to claim 23, characterized in that, The information sorting module is further configured to: obtain the semantic feature distance between the information and the information ranked first; and determine the sorting score of the information based on a scoring function. score=pctr A +w s *1 / ed AB ; Wherein, the pctr A The click-through rate of the information, w s As a hyperparameter, the d AB The semantic distance between the information and the information ranked first.

25. 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 recommendation method according to any one of claims 1 to 12.

26. A computer-readable storage medium, characterized in that, It stores executable instructions for causing a processor to execute, thereby implementing the AI-based recommendation method as described in any one of claims 1 to 12.

Images