Information recommendation method, device, electronic device, storage medium and program product

Abstract

The present application provides an artificial intelligence-based information recommendation method, device, electronic device, computer-readable storage medium, and computer program product; the method comprises: performing a first recommendation operation on first information to an object within a first period, and obtaining a first operation label value corresponding to the first information; determining a first error in the first period based on the first operation label value and a first score, sending the first error to each feature device, performing an update process on a first parameter based on the first error in the first period to obtain a second parameter, performing privacy encoding process on the information in the second period based on the second parameter and second original data to obtain a second feature of each information in the second period; receiving the second feature of the information, and performing a fusion process on the second feature based on an exploration mechanism to obtain a second score for the information; and performing a recommendation operation on the second information to the object. Through the present application, recommendation accuracy can be improved while ensuring information privacy.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to an information recommendation method, device, electronic device, computer-readable storage medium, and computer program product based on artificial intelligence. Background Art

[0002] Artificial Intelligence (AI) is the theory, method, 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 knowledge to achieve the best results.

[0003] Information recommendation is a key application of artificial intelligence. Its accuracy relies on comprehensive learning of object features, which can help improve the accuracy of recommendation systems for users. In many real-world scenarios, object features are often distributed across different parties. Because privacy-preserving information isolation mechanisms exist between parties, it's inappropriate for each party to share object features with the party performing the recommendation task. Consequently, related technologies struggle to balance the need for user privacy protection with the need for comprehensive learning of object features. Summary of the Invention

[0004] The embodiments of the present application provide an artificial intelligence-based information recommendation method, device, electronic device, computer-readable storage medium, and computer program product, which can improve recommendation accuracy while ensuring information privacy.

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

[0006] An embodiment of the present application provides an artificial intelligence-based information recommendation method, which is applied to a recommendation system. The recommendation system includes a recommendation device and multiple feature devices corresponding to multiple business parties. The method includes:

[0007] The recommendation device performs a first recommendation operation on first information to the object in a first period, and obtains a first operation label value corresponding to the first information, wherein the first information is at least one information ranked first in descending order based on the first score among multiple information in the first period;

[0008] Determine a first error of the first period based on the first operation tag value and the first score of the first information, and send the first error to each of the feature devices, so that each of the feature devices performs the following processing:

[0009] Based on the first error of the first cycle, updating the first parameter to obtain the second parameter of the business party in the second cycle, and performing privacy encoding processing on multiple information in the second cycle based on the second parameter and second original data corresponding to the business party to obtain a second feature of each information in the second cycle;

[0010] The second cycle is a cycle adjacent to the first cycle, the first parameter is a decision parameter of the service party of the feature device in the first cycle, and the second parameter is a decision parameter of the service party of the feature device in the second cycle;

[0011] For each piece of information in the second period, the recommendation device receives a second feature of the information sent by each feature device, and performs a fusion process based on the exploration mechanism on the second feature to obtain a second score for the information;

[0012] In the second period, the recommendation device performs a recommendation operation on the object for second information, wherein the second information is at least one piece of information ranked higher in descending order based on the second score among the plurality of pieces of information in the second period.

[0013] An embodiment of the present application provides an artificial intelligence-based information recommendation device, which is applied to a recommendation system. The recommendation system includes a recommendation device and multiple feature devices corresponding to multiple business parties. The device includes:

[0014] a recommendation module, configured to, within a first period, execute, by the recommendation device, a first recommendation operation on first information to an object, and obtain a first operation label value corresponding to the first information, wherein the first information is at least one of the plurality of information in the first period that is ranked first in descending order based on the first score;

[0015] an interaction module, configured to determine a first error of the first period based on the first operation tag value and the first score of the first information, and send the first error to each of the feature devices, so that each of the feature devices performs the following processing:

[0016] Based on the first error of the first cycle, updating the first parameter to obtain the second parameter of the business party in the second cycle, and performing privacy encoding processing on multiple information in the second cycle based on the second parameter and second original data corresponding to the business party to obtain a second feature of each information in the second cycle;

[0017] The second cycle is a cycle adjacent to the first cycle, the first parameter is a decision parameter of the service party of the feature device in the first cycle, and the second parameter is a decision parameter of the service party of the feature device in the second cycle;

[0018] a fusion module, configured to, for each piece of information in the second period, receive, by the recommendation device, the second feature of the information sent by each feature device, and perform fusion processing on the second feature based on the exploration mechanism to obtain a second score for the information;

[0019] The recommendation module is further configured to, within the second period, enable the recommendation device to perform a recommendation operation on second information to the object, wherein the second information is at least one piece of information ranked higher in descending order based on the second score among multiple pieces of information in the second period.

[0020] In the above scheme, before the recommendation device performs the first recommendation operation of the first information to the object within the first cycle, the recommendation module is also used to: for each of the information in the first cycle, the recommendation device receives the first feature of the information sent by each of the feature devices, and performs fusion processing on the first feature based on the exploration mechanism to obtain the first score of the information; wherein, for each of the feature devices, the first feature sent by the feature device is obtained by privacy encoding based on the first parameter corresponding to the business party and the first original data corresponding to the business party.

[0021] In the above scheme, the recommendation module is also used to: when the interactive behavior of the object with respect to the first information is collected within the first period, determine the first operation label value as the full score value in the scoring standard of the first score; when the interactive behavior of the object with respect to the first information is not collected within the first period, determine the first operation label value as the zero score value in the scoring standard of the first score.

[0022] In the above scheme, the interaction module is also used to: perform any one of the following processes: determining the difference between the first operation label value and the first score, squaring the difference to obtain the first error; determining the absolute value between the first operation label value and the first score, and using the absolute value as the first error; performing cross-entropy calculation on the first operation label value and the first score, and using the cross-entropy calculation result as the first error.

[0023] In the above scheme, the interaction module is also used to: obtain the gradient of the first error of the first period corresponding to the first parameter; multiply the gradient by the learning rate of the first period to obtain a parameter change value; and subtract the first parameter from the parameter change value to obtain the second parameter.

[0024] In the above scheme, the interaction module is also used to: perform the following processing on each information in the second period: obtain the recommended reference features of the information from the second original data, wherein the recommended reference features include at least one of the following: information features of the information, object features of the object; transpose the original feature matrix composed of the recommended reference features to obtain the original transposed matrix; multiply the original transposed matrix with the second parameter to obtain the encoding features of the information; superimpose Gaussian noise that meets the differential privacy standard on the encoding features of the information to obtain the second feature of the information.

[0025] In the above solution, the fusion module is further used to: sum the second features sent by multiple feature devices to obtain the third score.

[0026] In the above solution, the fusion module is further used to: obtain the weight of each of the business parties; and based on the weight of each of the business parties, perform weighted sum processing on the second features sent by the feature devices of multiple business parties to obtain the third score.

[0027] In the above scheme, the fusion module is also used to: sum the second features sent by multiple feature devices to obtain a third score; splice the second features sent by multiple feature devices to obtain spliced ​​features corresponding to the second period; perform exploration processing based on the spliced ​​features corresponding to the second period to obtain a fourth score of the information; and use the sum of the third score and the fourth score as the second score.

[0028] In the above scheme, the fusion module is also used to: obtain a splicing feature matrix corresponding to the second period composed of splicing features corresponding to the second period; transpose the splicing feature matrix corresponding to the second period to obtain a splicing transposed matrix corresponding to the second period; multiply the splicing transposed matrix corresponding to the second period with the inverse matrix of the historical feature matrix corresponding to the second period to obtain a first multiplication matrix, and multiply the first multiplication matrix with the splicing feature matrix corresponding to the second period to obtain a second multiplication matrix; perform square root processing on the second multiplication matrix to obtain a square root matrix, and obtain the norm corresponding to the square root matrix; obtain a fourth score positively correlated with the norm.

[0029] In the above scheme, the fusion module is also used to: transpose the splicing feature matrix corresponding to the first period to obtain a splicing transposed matrix corresponding to the first period; multiply the splicing feature matrix corresponding to the first period with the splicing transposed matrix corresponding to the first period to obtain a third multiplied matrix; obtain the historical feature matrix corresponding to the first period; add the historical feature matrix corresponding to the first period and the third multiplied matrix to obtain a historical feature matrix corresponding to the second period.

[0030] An embodiment of the present application provides an electronic device, including:

[0031] a memory for storing computer-executable instructions;

[0032] The processor is used to implement the artificial intelligence-based information recommendation method provided in the embodiment of the present application when executing the computer-executable instructions stored in the memory.

[0033] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for causing a processor to execute instructions to implement the artificial intelligence-based information recommendation method provided in the embodiment of the present application.

[0034] An embodiment of the present application provides a computer program product, including a computer program or computer-executable instructions. When the computer program or computer-executable instructions are executed by a processor, the artificial intelligence-based information recommendation method provided in the embodiment of the present application is implemented.

[0035] The embodiments of the present application have the following beneficial effects:

[0036] Through the embodiments of the present application, the recommendation device completes the recommendation of information and obtains real feedback, and then returns the error to multiple business parties, so that the business parties can update the corresponding parameters based on the error. After each business party completes the privacy encoding based on the updated parameters, multiple business parties can send the features obtained by the privacy encoding to the recommendation device. The recommendation device fuses the features obtained by multiple privacy encodings, so that the availability of the features can still be guaranteed from the statistics obtained by the fusion while hiding the second original data. Recommendations are made based on the feature fusion results, so that the features can be comprehensively learned and the recommendation accuracy can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0039] Figure 3A-3CThis is a flowchart of an information recommendation method based on artificial intelligence provided by an embodiment of the present application;

[0040] Figure 4 It is a data relationship diagram of the recommendation system provided by the embodiment of the present application;

[0041] Figure 5 This is a schematic diagram of the coding process of the information recommendation method based on artificial intelligence provided in an embodiment of the present application;

[0042] Figure 6 This is a data exchange diagram of the artificial intelligence-based information recommendation method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0044] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be 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" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application 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 those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0047] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0048] 1) Federated Learning: This is a distributed machine learning technology. Its core idea is to conduct distributed model training among multiple data sources with local data. Without the need to exchange local individual or sample data, a global model is constructed based on virtual fused data by simply exchanging model parameters or intermediate results, thereby achieving a balance between data privacy protection and data sharing and computing.

[0049] 2) Feature encoding: The collected data is often messy and may contain various non-numeric symbols, such as Chinese characters. However, machine learning models require numeric data, as only numeric data can be used for calculations. Therefore, we need to encode each special feature value, which is also a quantization process.

[0050] In many real-world business scenarios, object features are often distributed across different parties. Due to information isolation mechanisms in place to protect privacy, these parties are not advised to share object features with the party performing the recommendation task. This makes it difficult for related technologies to simultaneously address both user privacy protection and the need to comprehensively learn object features. To address this privacy-preserving recommendation requirement, related technologies typically employ federated recommendation schemes based on deep neural networks. Specifically, in the distributed feature scenario, cryptographic privacy-enhancing schemes are used to train multiple neural networks while preserving privacy. These purely neural network-based schemes often struggle to adapt to cold-start recommendation scenarios, specifically when a new user or product is launched online. Because neural network performance often relies on extensive user data, achieving good performance in cold-start recommendation scenarios often leads to biased recommendations toward existing popular items and overlooking newly launched ones.

[0051] The embodiments of the present application provide an artificial intelligence-based information recommendation method, device, electronic device, computer-readable storage medium, and computer program product, which can improve recommendation accuracy while ensuring information privacy.

[0052] The following describes an exemplary application of an electronic device provided by an embodiment of the present application. The electronic device provided by an embodiment of the present application includes a recommendation device and a feature device, and both the recommendation device and the feature device can be servers. The following describes an exemplary application when the electronic device is implemented as a server.

[0053] See also Figure 1 , Figure 1This is a structural diagram of the information recommendation system based on artificial intelligence provided by an embodiment of the present application. In order to support a news application, the terminal 400 is connected to the server 200-2 (recommendation device) through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two. In the first cycle, the terminal 400 sends a recommendation request to the server 200-2. The server 200-2 sends a feature acquisition request to multiple servers 200-1 (feature devices). The multiple servers 200-1 perform differential privacy encoding on the feature data based on their respective first parameters to obtain the first feature. For each information in the first cycle, the server 200-2 receives the first feature of the information sent by each server 200-1, and performs fusion processing on the first feature based on the exploration mechanism to obtain the first score of the information in the first cycle. In the first cycle, the server 200-2 sends the first information with the highest first score to the terminal 400 for presentation, and obtains the first operation label value corresponding to the first information. The server 200-2 performs differential privacy encoding on the feature data based on the first operation label value and the first information. The first score determines the first error of the first period and sends the first error to each server 200-1. Each server 200-1 performs update processing on the first parameter based on the first error of the first period to obtain the second parameter of the business party in the second period. Based on the second parameter and the second original data corresponding to the business party, the multiple information of the second period are differentially privacy encoded to obtain the second feature of each information of the second period. For each information of the second period, the server 200-2 receives the second feature of the information sent by each server 200-1, and performs fusion processing on the second feature based on the exploration mechanism to obtain the second score of the information of the second period. The server 200-2 sends the second information with the highest second score to the terminal 400 for presentation.

[0054] In some embodiments, server 200-1 and server 200-2 can be independent physical servers, or a server cluster or distributed system composed of multiple physical servers. They can also be cloud servers that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Terminal 400 can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application.

[0055] In some embodiments, the terminal or server can implement the information recommendation method based on artificial intelligence provided by the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; it can be a native application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as a news APP or an e-commerce APP; it can also be a small program, that is, a program that can be run only by downloading it into a browser environment; it can also be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module or plug-in.

[0056] See also Figure 2 , Figure 2 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, taking the electronic device as server 200-2 as an example, Figure 2 The server 200-2 shown includes: at least one processor 210, a memory 250, at least one network interface 220 and a user interface 230. The various components in the terminal 400 are coupled together via a bus system 240. It is understood that the bus system 240 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 240 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 240 is not described in detail. Figure 2 Various buses are labeled as bus system 240 .

[0057] The 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., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0058] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 250 may optionally include one or more storage devices that are physically remote from the processor 210.

[0059] The memory 250 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 250 described in the embodiments of the present application is intended to include any suitable type of memory.

[0060] In some embodiments, the memory 250 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0061] Operating system 251, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0062] The network communication module 252 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 220. Exemplary network interfaces 220 include Bluetooth, Wireless LAN (WiFi), and Universal Serial Bus (USB).

[0063] In some embodiments, the artificial intelligence-based information recommendation device provided in the embodiments of the present application can be implemented in software. Figure 2 An artificial intelligence-based information recommendation device 255 stored in a memory 250 is shown, which can be software in the form of programs and plug-ins, including the following software modules: a recommendation module 2551, an interaction module 2552, and a fusion module 2553. These modules are logical, and therefore can be arbitrarily combined or further split according to the functions implemented. The functions of each module will be explained below.

[0064] The artificial intelligence-based information recommendation method provided in the embodiment of the present application will be explained in combination with the exemplary application and implementation of the server 200-2 provided in the embodiment of the present application.

[0065] See also Figure 3A , Figure 3A This is a flow chart of an information recommendation method based on artificial intelligence provided by an embodiment of the present application. The information recommendation method is applied to a recommendation system. The recommendation system includes a recommendation device and multiple feature devices corresponding to multiple business parties. Figure 3A Steps 101 to 109 are shown for explanation.

[0066] In step 101 , a recommendation device performs a first recommendation operation of first information on an object within a first period.

[0067] For example, the object is the service object of the recommendation system. For example, the object of the news recommendation system is the user who uses the news app, and the object of the video recommendation system is the user who uses the video app. The first information is the information pushed to the user, and the first recommendation operation is to send the first information to the terminal used by the user, so that it is presented on the terminal.

[0068] In some embodiments, before the recommendation device performs the first recommendation operation for the first information to the object in the first cycle in step 101, the following technical solution may also be executed: for each information in the first cycle, the recommendation device receives the first feature of the information sent by each feature device, and performs a fusion process on the first feature based on the exploration mechanism to obtain a first score for the information; wherein, for each feature device, the first feature sent by the feature device is obtained by privacy encoding based on the first parameter of the corresponding business party and the first original data corresponding to the business party. Through the embodiments of the present application, the recommendations in the first cycle can also be obtained based on comprehensive learning of the features, thereby ensuring a steady improvement in the accuracy of recommendations during the iterative process.

[0069] As an example, the first cycle is any recommendation cycle, which starts when a user initiates a recommendation request and ends when the user initiates another recommendation request. In each recommendation cycle, the recommendation device performs a recommendation to the user.

[0070] For example, a recommendation cycle has five pieces of information as recommendation candidates. The recommendation system includes three business parties. A privacy protection mechanism based on information isolation exists between the three business parties and the recommendation device. Specifically, the three business parties and the recommendation device do not directly share their stored raw data. Business party A stores the user's object raw data in dimension A, such as the user's purchase data. Business party B stores the user's object raw data in dimension B, such as the user's video viewing data. Because object raw data changes over time, the raw data in the first cycle is referred to as the first raw data, and the raw data in the second cycle is referred to as the second raw data for distinction. The following example uses information A, the object "Xiao Ming," and business party A as an example. Based on business party A's first parameter and the first raw data stored in business party A, information A in the first cycle is privacy-encoded. This yields the first feature of information A in the first cycle. The first raw data includes the object raw data of "Xiao Ming" in dimension A and the information raw data of information A. The recommendation device receives the first features of information A sent by the feature devices of the three business parties and fuses the three first features using an exploration mechanism to obtain a first score for the information. The plurality of information in the first period is sorted in descending order based on the first score, and the plurality of information with the highest sorting order is recommended as the first information.

[0071] In step 102 , the recommendation device obtains a first operation tag value corresponding to the first information.

[0072] As an example, the first information is at least one information that ranks first among the multiple information in the first period and is sorted in descending order based on the first score. Obtaining the first operation tag value corresponding to the first information in step 102 can be achieved through the following technical solution: when an interactive behavior of the subject with respect to the first information is collected within the first period, the first operation tag value is determined to be a full score within the scoring standard of the first score; when no interactive behavior of the subject with respect to the first information is collected within the first period, the first operation tag value is determined to be a zero score within the scoring standard of the first score.

[0073] As an example, if the first score is from zero to one, it is equivalent to 1 being the full score in the scoring standard and 0 being the zero score in the scoring standard. If the first score is from zero to ten, it is equivalent to 10 being the full score in the scoring standard and 0 being the zero score in the scoring standard. If the first score is from zero to one hundred, it is equivalent to 100 being the full score in the scoring standard and 0 being the zero score in the scoring standard. Interactive behaviors include click behaviors, like behaviors, play behaviors, etc. The first operation label value is used to indicate whether interactive behaviors for the first information are collected in the first period. When the interactive behaviors for the first information of the object are collected in the first period, it indicates that the recommendation of the first information is accurate, and thus the first operation label value of the first information is the full score. When the interactive behaviors for the first information of the object are not collected in the first period, it indicates that the recommendation of the first information to the object is inaccurate, and thus the first operation label value of the first information is zero.

[0074] In step 103 , the recommendation device determines a first error of the first period based on the first operation label value and the first score of the first information.

[0075] In some embodiments, determining the first error of the first cycle based on the first operation label value and the first score of the first information in step 103 can be achieved by the following technical solution: for each first information, performing any one of the following processes: determining the difference between the first operation label value and the first score, squaring the difference to obtain the first error of the first information; determining the absolute value between the first operation label value and the first score, and using the absolute value as the first error of the first information; performing cross entropy calculation on the first operation label value and the first score, and using the cross entropy calculation result as the first error of the first information; when there are multiple first information, summing the first errors of multiple first information to obtain the first error of the first cycle; when there are only one first information, determining the first error of the first information as the first error of the first cycle. Through the embodiments of the present application, a first error that characterizes the recommendation effect of the entire first cycle can be obtained, thereby improving the effectiveness of parameter updating.

[0076] As an example, take a certain first information as an example: when the first operation label value of the first information is 1, and the first score of the first information is 0.8, the first error of the first information is 0.2 (absolute value), or the first error of the first information can also be 0.04 (the square of the difference). The first errors of multiple first information are summed to obtain the first error of the first period. The first error of the first period is used to characterize the recommendation accuracy of the recommendation system in the first period, and the recommendation accuracy is negatively correlated with the first error.

[0077] In step 104 , the recommender device sends the first error to each feature device.

[0078] In step 105, each feature device performs an update process on the first parameter based on the first error in the first cycle to obtain the second parameter of the business party in the second cycle.

[0079] In some embodiments, in step 105, based on the first error of the first cycle, the first parameter is updated to obtain the second parameter of the business party in the second cycle. This can be achieved through the following technical solutions: obtaining the gradient of the first parameter corresponding to the first error of the first cycle; multiplying the gradient by the learning rate of the first cycle to obtain the parameter change value; and subtracting the first parameter from the parameter change value to obtain the second parameter of the business party in the second cycle. By updating parameters in distributed feature devices through the embodiments of the present application, a distributed deployment of a multi-armed bandit is achieved, thereby improving recommendation accuracy while achieving privacy protection.

[0080] As an example, the i-th business party among the multiple business parties obtains the first error ∈ t Afterwards, the parameters are updated using the following formula (1):

[0081]

[0082] Among them, the left side of the equation is the second parameter encoded by the characteristic device of the i-th business party in the second cycle. is the first parameter encoded by the characteristic device of the i-th service party in the first cycle, l t is the learning rate of the first cycle, is the gradient of the first error of the first cycle corresponding to the first parameter.

[0083] In step 106, each feature device performs privacy encoding processing on multiple information in the second cycle based on the second parameter and the second original data corresponding to the business party to obtain the second feature of each information in the second cycle.

[0084] As an example, the privacy encoding process represents encoding the second original data based on a privacy protection technology, for example, injecting Gaussian noise when encoding, for example, performing identity mapping processing when encoding, and so on.

[0085] As an example, the second cycle is the next cycle adjacent to the first cycle, the first parameter is the decision parameter of the business party of the feature device in the first cycle, the second parameter is the decision parameter of the business party of the feature device in the second cycle, the second parameter is updated based on the first parameter and the first error of the first cycle, the first cycle is any recommendation cycle, and the second cycle is the next cycle adjacent to the first cycle. For example, the first cycle starts at the moment when the user initiates the recommendation request for the first time, and the first cycle ends at the moment when the user initiates the recommendation request for the second time. The second cycle starts at the moment when the user initiates the recommendation request for the second time, and the second cycle ends at the moment when the user initiates the recommendation request for the third time.

[0086] In some embodiments, see Figure 3B , Figure 3B This is a flow chart of the information recommendation method based on artificial intelligence provided by an embodiment of the present application. In step 106, privacy encoding processing is performed on multiple information in the second period based on the second parameter and the second original data corresponding to the business party to obtain the second feature of each information in the second period. Figure 3B Steps 1061 to 1064 are implemented.

[0087] In step 1061 , recommended reference features of information are obtained from the second original data.

[0088] The recommended reference feature includes at least one of the following: an information feature of information, and an object feature of an object.

[0089] For example, in the second cycle, five pieces of information are recommended as candidates. The recommendation system includes three business parties. A privacy protection mechanism based on information isolation exists between these three business parties and the recommendation device. This means that the three business parties and the recommendation device do not directly share their stored raw data. Business party A stores the user's object raw data in dimension A, such as the user's purchase data, and business party B stores the user's object raw data in dimension B, such as the user's video viewing data. Because object raw data changes over time, the raw data in the first cycle is referred to as the first raw data, and the raw data in the second cycle is referred to as the second raw data for distinction.

[0090] The following is an example of information A, object "Xiao Ming" and business party A. The second original data includes the object original data of object "Xiao Ming" in dimension A (for example, the user's purchase record) and the information original data of information A (for example, the release record of information A). The original features of the object are extracted from the object original data included in the second original data, and the original features of the information are extracted from the information original data included in the second original data. The original data are all actually collected data and may contain various non-numeric special symbols, such as Chinese. In fact, the data required by the machine learning model is digital, because only digital types can be calculated. Therefore, for various special feature values, we need to encode them accordingly, which is also a quantization process. The original features of the object and the original features of the information are obtained through the quantization process.

[0091] In step 1062 , the original feature matrix composed of the recommended reference features is transposed to obtain an original transposed matrix.

[0092] In step 1063, the original transposed matrix is ​​multiplied by the second parameter to obtain the encoding feature of the information.

[0093] In step 1064, Gaussian noise that meets the differential privacy standard is superimposed on the encoded feature of the information to obtain a second feature of the information.

[0094] As an example, steps 1062 to 1064 are performed according to the following formula (2):

[0095]

[0096] in, represents Gaussian noise that meets the differential privacy criteria, is the original feature matrix of business party i, is the second parameter of business party i in the second cycle, It is information a provided by business party i t The second feature is to further enhance the strength of privacy protection, Gaussian noise Gaussian noise of a magnitude that meets the differential privacy criterion is selected. Two variance boundaries can usually be used to limit the magnitude of the differential privacy criterion, that is, the variance of the Gaussian noise is greater than the first threshold and less than the second threshold, and the first threshold is less than the second threshold. The larger the variance, the better the privacy, and the smaller the variance, the higher the data availability.

[0097] In step 107 , for each message in the second cycle, the recommendation device receives the second feature of the message sent by each feature device.

[0098] As an example, for example, the second cycle includes 5 information as recommendation candidates, and the recommendation system includes 3 business parties. Taking information A as an example, the recommendation device receives the second feature of the information sent by the feature devices corresponding to the 3 business parties.

[0099] In step 108 , the recommendation device performs a fusion process on the second feature based on the exploration mechanism to obtain a second score of the information.

[0100] As an example, the fusion processing representation based on the exploration mechanism takes uncertainty information into account when fusing the second features sent by multiple feature devices, so that the degree of interest of the object in the information in the cold start scenario can be explored based on the uncertainty information.

[0101] In some embodiments, see Figure 3C , Figure 3C This is a flow chart of the information recommendation method based on artificial intelligence provided by the embodiment of the present application. In step 108, the second feature is fused based on the exploration mechanism to obtain the second score of the information. Figure 3C Steps 1081 to 1084 are shown to be implemented.

[0102] In step 1081, the second features sent by multiple feature devices are fused to obtain a third score.

[0103] In some embodiments, in step 1081, the second features sent by multiple feature devices are fused to obtain a third score, which can be achieved through the following technical solution: the second features sent by multiple feature devices are summed to obtain the third score.

[0104] As an example, see formula (3):

[0105]

[0106] Among them, p t, is the third fraction, is the second feature sent by the feature device of a business party i, and M is the number of business parties.

[0107] In some embodiments, in step 1081, the second features sent by multiple feature devices are fused to obtain a third score, which can be achieved through the following technical solution: obtaining the weight of each business party; based on the weight of each business party, performing weighted summation on the second features sent by the feature devices of multiple business parties to obtain the third score.

[0108] As an example, see formula (4):

[0109]

[0110] Among them, p t, is the third fraction, is the second feature sent by the feature device of a business party i, W i is the weight of a business party i, and M is the number of business parties.

[0111] In step 1082, the second features sent by multiple feature devices are spliced ​​to obtain spliced ​​features corresponding to the second period.

[0112] In step 1083, an exploration process is performed based on the splicing features corresponding to the second period to obtain a fourth score of the information.

[0113] In some embodiments, the exploration processing based on the splicing features in step 1083 to obtain the fourth score of the information can be achieved by the following technical solutions: obtaining a splicing feature matrix corresponding to the second period composed of the splicing features of the corresponding second period; transposing the splicing feature matrix corresponding to the second period to obtain a splicing transposed matrix corresponding to the second period; multiplying the splicing transposed matrix corresponding to the second period with the inverse matrix of the historical feature matrix corresponding to the second period to obtain a first multiplication matrix, and multiplying the first multiplication matrix with the splicing feature matrix corresponding to the second period to obtain a second multiplication matrix; performing square root processing on the second multiplication matrix to obtain a square root matrix, and obtaining the norm of the corresponding square root matrix; obtaining a fourth score positively correlated with the norm.

[0114] As an example, see formula (5):

[0115]

[0116] Among them, p t, is the second fraction, is the third score, α is a hyperparameter, is the square root matrix, is the concatenated feature matrix, is the concatenated transposed matrix, It is the inverse matrix of the historical characteristic matrix corresponding to the second period.

[0117] In some embodiments, the splicing feature matrix corresponding to the first period is transposed to obtain a splicing transposed matrix corresponding to the first period; the splicing feature matrix corresponding to the first period and the splicing transposed matrix corresponding to the first period are multiplied to obtain a third multiplication matrix; the historical feature matrix corresponding to the first period is obtained; the historical feature matrix corresponding to the first period and the third multiplication matrix are added to obtain a historical feature matrix corresponding to the second period.

[0118] As an example, see formula (6):

[0119]

[0120] in, is the historical characteristic matrix corresponding to the second period, is the historical characteristic matrix corresponding to the first period, is the third multiplication matrix, is the concatenated feature matrix corresponding to the first cycle, is the concatenated transposed matrix corresponding to the first period.

[0121] In step 1084 , the sum of the third score and the fourth score is used as the second score.

[0122] In the above embodiment, if the second score only includes the third score, our recommendation iteration process can easily fall into a local extreme value. The significance of the fourth score is that if the recommendation system knows too little about a certain information (cold start), then its average return (the third score) at this time has a very low confidence level, a high uncertainty level, and a large confidence interval. This is equivalent to not believing that the average return at this time is its true average return. Therefore, the fourth score is needed to increase the probability of selecting this information, so as to learn richer object features and explore deeper interests by selecting this information.

[0123] In step 109 , within the second period, the recommendation device performs a recommendation operation on the object with respect to the second information.

[0124] As an example, the second information is at least one information among the plurality of information in the second period that is ranked higher in descending order based on the second score.

[0125] Through the embodiments of the present application, the recommendation device completes the recommendation of information and obtains real feedback, and then returns the error to multiple business parties, so that the business parties can update the corresponding parameters based on the error. After each business party completes the differential privacy encoding based on the updated parameters, multiple business parties can send the features obtained by the differential privacy encoding to the recommendation device. The recommendation device fuses the features obtained by multiple differential privacy encodings, so that the availability of the features can still be guaranteed from the statistics obtained by fusion while hiding the second original data. Recommendations are made based on the feature fusion results, so that the features can be comprehensively learned and the recommendation accuracy can be improved.

[0126] The following describes an exemplary application of the embodiments of the present application in a practical application scenario.

[0127] During the first cycle, the terminal 400 sends a recommendation request to the server 200-2, and the server 200-2 sends a feature acquisition request to multiple servers 200-1 (feature devices). The multiple servers 200-1 perform differential privacy encoding on the feature data based on their respective first parameters to obtain the first feature. For each information in the first cycle, the server 200-2 receives the first feature of the information sent by each server 200-1, and performs fusion processing on the first feature based on the exploration mechanism to obtain the first score of the information in the first cycle; during the first cycle, the server 200-2 sends the first information with the highest first score to the terminal 400 for presentation, and obtains the first operation label value corresponding to the first information. The server 200-2 determines the first error of the first cycle based on the first operation label value and the first score of the first information, and sends the first error to each server 200-1. Each server 200-1 performs update processing on the first parameter based on the first error of the first cycle to obtain the second parameter of the business party in the second cycle.

[0128] During the second period, the terminal 400 sends a recommendation request to the server 200-2, and the server 200-2 sends a feature acquisition request to multiple servers 200-1. The multiple servers 200-1 perform differential privacy encoding processing on multiple information of the second period based on their respective second parameters to obtain the second feature of each information of the second period. For each information of the second period, the server 200-2 receives the second feature of the information sent by each server 200-1, and performs fusion processing on the second feature based on the exploration mechanism to obtain the second score of the information of the second period; the server 200-2 sends the second information with the highest second score to the terminal 400 for presentation.

[0129] In some embodiments, recommendation systems in related technologies are typically built under the assumption of centralized training, which assumes that the business side has access to all user data. However, with the increasing emphasis on privacy protection, it has become difficult for different departments to share user object data. However, the breadth of object data can significantly impact the accuracy of the recommendation system. Therefore, to achieve a highly accurate recommendation system, it is necessary to obtain different types of object data for a particular user from multiple business departments that are privacy-isolated. For example, the object data for a user's book purchases and the user's gaming data must come from different business departments, but these object data can all contribute to a comprehensive representation of the user.

[0130] See also Figure 4 , Figure 4This is a data relationship diagram of the recommendation system provided by the embodiment of the present application. For example, the public welfare project recommendation system is used to perform recommendation tasks for public welfare projects, that is, to recommend public welfare projects to users. For a single user, the object data required for recommendation not only comes from the department corresponding to the public welfare project recommendation system, but is also distributed in different other departments, such as public welfare organization departments. There is privacy isolation between the two departments, that is, data will not be shared. In response to the above contradictions, the embodiment of the present application provides a recommendation system suitable for distributed feature scenarios. Based on local feature encoding, differential privacy technology, and project cold start exploration mechanism, a recommendation system suitable for multi-party joint modeling and privacy protection is constructed, so that multiple business departments can complete the construction of the recommendation system without exchanging the user's original object data.

[0131] The embodiment of the present application first introduces a centralized recommendation scenario (i.e., all object data of all users belong to the same department). At the beginning of the recommendation task, there is a lack of samples and the acquisition of samples is streaming. The multi-armed bandit model can learn from each sample and make the best possible decision for each sample. Specifically, when the decision maker is given T experimental opportunities, for each experimental opportunity, the decision maker can observe a sample and make a decision based on this sample and historical experience information. Then the real business scenario will give the decision maker a feedback, which depends on the nature of the sample and the action of the decision maker. The decision maker tries to use this sample and the observed benefits to infer unknown rules, and use it to improve the decision when the next experimental opportunity comes. The ultimate goal of the decision maker is to maximize the cumulative benefits of these T experimental opportunities.

[0132] Multi-armed bandit algorithms are widely used in cold-start algorithm design, such as recommendation systems. Recommendation systems face a cold-start problem when they need to perform recommendations without having a comprehensive understanding of users' current needs. Therefore, they must experimentally perform recommendations for each new user to gain comprehensive empirical information for recommendations, while also ensuring that recommendations are as effective as possible for each user. For example, in online shopping recommendation systems, the system, as the decision-maker, observes a user's basic information, such as age and browsing history, during each decision-making process and recommends one of multiple candidate products to the user. After viewing the recommended product, the customer decides whether to click or purchase it, while being observed by the recommendation system. Based on user feedback, the recommendation system attempts to update its recommendation mechanism to improve the effectiveness of recommendations for subsequent users.

[0133] In some embodiments, the application process of the multi-armed bandit in a centralized recommendation scenario is as follows:

[0134] For t in[0:T]: / / For the nth recommendation among T recommendations

[0135] For the tth recommendation, perform the following steps:

[0136] Calculate the mean reward: is the predicted feedback value, is the information a used for the tth recommendation t The local object characteristics of yes The transposed matrix of is the parameter used for the tth recommendation;

[0137] Calculate the reward cap: is the predicted recommendation score, is the predicted feedback value, α is a hyperparameter, yes The transposed matrix of is the inverse matrix of the cumulative characteristic matrix;

[0138] Select the information a that maximizes the reward cap t And get the reward observation value r t :a t =argmax a p t, ;

[0139] Parameter update:

[0140] From the above process, we can see that for each piece of information, the multi-armed bandit passes the parameter Predict a given user's response to information t By optimizing the parameter θ a , thereby modeling the user's preference for the information and making accurate recommendations to the user.

[0141] The embodiment of the present application extends the multi-armed bandit algorithm in a centralized scenario to a distributed feature scenario. The core of the embodiment of the present application lies in how to jointly build a multi-armed bandit with all parties while protecting the privacy of object features in a distributed feature scenario.

[0142] The above section has provided a detailed introduction to the multi-armed bandit in centralized scenarios. However, in most real-world scenarios, the local object features used to recommend information are often distributed across M departments, each of which holds local object features: However, due to compliance requirements, different departments cannot directly share the original object features of users. Therefore, it is necessary to design various privacy protection mechanisms to complete the construction of the recommendation system. To this end, the embodiment of the present application proposes a privacy-protected multi-armed bandit construction scheme, which combines local encoding technology with differential privacy technology to complete the protection of local object features. Department A is responsible for the construction of the recommendation system, that is, Department A will interact with other departments B in information. The artificial intelligence-based information recommendation method provided in the embodiment of the present application includes two stages: other departments B are responsible for local feature encoding, and department A completes the learning of the multi-armed bandit.

[0143] The local feature encoding process of other departments B is described in detail below. The privacy protection mechanism provided by the embodiment of the present application is to convert the parameter θ in the centralized scenario into a Split into M different departments B, there is information isolation between these departments B, that is, the original object characteristics of each department need to be protected. The i-th department in multiple departments B has local parameters Because different departments cannot know each other's local parameters, based on this, the embodiment of the present application uses local parameters to encode the local object features of the department, see formula (7):

[0144]

[0145] in, is the local object feature of department i, is the local parameter of department i, is the feature coding result of department i.

[0146] After completing the local encoding, it can be found that the exploration mechanism of the multi-armed bandit is still difficult to directly complete the calculation under the premise of rejecting shared features. Therefore, in order to complete the exploration in the cold start scenario, the embodiment of the present application injects additional Gaussian noise into the above-mentioned local encoding and completes the update of the local encoding according to the following formula (8):

[0147]

[0148] in, represents Gaussian noise, is the local object feature of department i, is the local parameter of department i, is the feature encoding result of department i. In order to further enhance the strength of privacy protection, Gaussian noise is added The magnitude is chosen to satisfy the differential privacy criteria.

[0149] See also Figure 5 , Figure 5This is a schematic diagram of the encoding process of the artificial intelligence-based information recommendation method provided in an embodiment of the present application. First, the local parameters are initialized, then the local features are encoded through the local parameters, and then Gaussian noise that meets the differential privacy standard is selected. Then, the Gaussian noise is injected into the local feature code, and the updated local feature code is sent to department A responsible for executing the recommendation task.

[0150] The following is a detailed introduction to the recommended solution based on the multi-armed bandit completed by Department A. Figure 6 , Figure 6 This is a data exchange diagram of the artificial intelligence-based information recommendation method provided by an embodiment of the present application. After completing local encoding, other departments B can send the encoded features to department A. After receiving the local encoding features of different departments B, department A completes the information recommendation and obtains the user's real feedback, and then returns the feedback error to other departments B. Other departments B then update the corresponding parameters based on the feedback error. The specific process is as follows:

[0151] For t in[0:T]: / / For the nth recommendation among T recommendations

[0152] For the tth recommendation, perform the following steps:

[0153] Received encoding characteristics from other department B: / / Feature encoding processing based on formula (8)

[0154] Combined calculation of rewards: is the predicted recommendation score, is the predicted feedback value, α is a hyperparameter, yes The transposed matrix of is the inverse matrix of the cumulative feature matrix, It's multiple departments The splicing result of

[0155] Select the information a that maximizes the reward cap t and get user feedback t :a t =argmax a p t, ;

[0156] Calculate the feedback local error according to the feedback: t =(r t -p t,a ) 2 ; / / ∈ t Is the feedback local error will feedback local error ∈ t Send to other department B.

[0157] The i-th department in multiple other departments B obtains feedback local error ∈ t Afterwards, the local parameters are updated using the following formula (9):

[0158]

[0159] in, is the parameter of the i-th department, l t is the learning rate for parameter update after the tth recommendation, It is the gradient of the parameter corresponding to the feedback local error.

[0160] In actual business scenarios, the recommendation system uses distributed collection. The embodiments of this application allow for collaborative deployment between devices in different departments to achieve higher training efficiency and feature utilization. At the same time, due to the use of a federated learning framework, only statistics are required rather than the second raw data, and data privacy protection between different servers can be guaranteed. In addition, the technical solution provided by the embodiments of this application is combined with technologies such as differential privacy to provide more comprehensive protection for data privacy.

[0161] It is understandable that in the embodiments of the present application, when data related to user information, object characteristics, etc. is involved, when the embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0162] The following continues to describe the exemplary structure of the artificial intelligence-based information recommendation device 255 provided in the embodiment of the present application as a software module. In some embodiments, such as Figure 2As shown, the software modules in the artificial intelligence-based information recommendation device 255 stored in the memory 250 may include: a recommendation module 2551, which is used to recommend that the device perform a first recommendation operation of the first information to the object within a first cycle, and obtain a first operation label value corresponding to the first information, wherein the first information is at least one information that is ranked first in descending order based on the first score among multiple information in the first cycle; an interaction module 2552, which is used to determine a first error of the first cycle based on the first operation label value and the first score of the first information, and send the first error to each feature device, so that each feature device performs the following processing: based on the first error of the first cycle, update the first parameter to obtain the second parameter of the business party in the second cycle, and based on the second parameter and the second error corresponding to the business party The original data performs privacy encoding processing on multiple information in the second period respectively to obtain the second feature of each information in the second period; wherein the second period is the next period adjacent to the first period, the first parameter is the decision parameter of the business party of the feature device in the first period, and the second parameter is the decision parameter of the business party of the feature device in the second period; the fusion module 2553 is used for, for each information in the second period, the recommendation device receives the second feature of the information sent by each feature device, and performs fusion processing on the second feature based on the exploration mechanism to obtain the second score of the information; the recommendation module 2551 is also used for, within the second period, the recommendation device to perform a recommendation operation on the second information to the object, wherein the second information is at least one information in the multiple information in the second period that is ranked higher in descending order based on the second score.

[0163] In some embodiments, before the recommendation device performs the first recommendation operation of the first information to the object within the first cycle, the recommendation module 2551 is also used to: for each information in the first cycle, the recommendation device receives the first feature of the information sent by each feature device, and performs fusion processing on the first feature based on the exploration mechanism to obtain the first score of the information; wherein, for each feature device, the first feature sent by the feature device is obtained by privacy encoding based on the first parameter of the corresponding business party and the first original data corresponding to the business party.

[0164] In some embodiments, the recommendation module 2551 is also used to: when the object's interactive behavior towards the first information is collected within the first period, determine the first operation label value as the full score value in the scoring standard of the first score; when the object's interactive behavior towards the first information is not collected within the first period, determine the first operation label value as the zero score value in the scoring standard of the first score.

[0165] In some embodiments, the interaction module 2552 is also used to: perform any one of the following processes: determining the difference between the first operation label value and the first score, squaring the difference to obtain a first error; determining the absolute value between the first operation label value and the first score, and using the absolute value as the first error; performing cross-entropy calculation on the first operation label value and the first score, and using the cross-entropy calculation result as the first error.

[0166] In some embodiments, the interaction module 2552 is further used to: obtain the gradient of the first parameter corresponding to the first error in the first period; multiply the gradient by the learning rate of the first period to obtain the parameter change value; subtract the first parameter from the parameter change value to obtain the second parameter of the business party in the second period.

[0167] In some embodiments, the interaction module 2552 is also used to: perform the following processing on each information of the second period: obtain recommended reference features of the information from the second original data, wherein the recommended reference features include at least one of the following: information features of the information, object features of the object; transpose the original feature matrix composed of the recommended reference features to obtain the original transposed matrix; multiply the original transposed matrix with the second parameter to obtain the coding features of the information; superimpose Gaussian noise that meets the differential privacy standard on the coding features of the information to obtain the second feature of the information.

[0168] In some embodiments, the fusion module 2553 is further used to: sum the second features sent by multiple feature devices to obtain a third score.

[0169] In some embodiments, the fusion module 2553 is further used to: obtain the weight of each business party; and based on the weight of each business party, perform weighted sum processing on the second features sent by the feature devices of multiple business parties to obtain a third score.

[0170] In some embodiments, the fusion module 2553 is further used to: sum the second features sent by multiple feature devices to obtain a third score; splice the second features sent by multiple feature devices to obtain spliced ​​features corresponding to the second period; perform exploration processing based on the spliced ​​features corresponding to the second period to obtain a fourth score of the information; and use the sum of the third score and the fourth score as the second score.

[0171] In some embodiments, the fusion module 2553 is further used to: obtain a splicing feature matrix corresponding to the second period composed of splicing features corresponding to the second period; transpose the splicing feature matrix corresponding to the second period to obtain a splicing transposed matrix corresponding to the second period; multiply the splicing transposed matrix corresponding to the second period with the inverse matrix of the historical feature matrix corresponding to the second period to obtain a first multiplication matrix, and multiply the first multiplication matrix with the splicing feature matrix corresponding to the second period to obtain a second multiplication matrix; perform square root processing on the second multiplication matrix to obtain a square root matrix, and obtain the norm of the corresponding square root matrix; obtain a fourth score that is positively correlated with the norm.

[0172] In some embodiments, the fusion module 2553 is further used to: transpose the splicing feature matrix corresponding to the first period to obtain a splicing transposed matrix corresponding to the first period; multiply the splicing feature matrix corresponding to the first period with the splicing transposed matrix corresponding to the first period to obtain a third multiplication matrix; obtain the historical feature matrix corresponding to the first period; and add the historical feature matrix corresponding to the first period with the third multiplication matrix to obtain a historical feature matrix corresponding to the second period.

[0173] The present invention provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the instructions, causing the electronic device to perform the artificial intelligence-based information recommendation method described above in the present invention.

[0174] The embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, the processor will execute the information recommendation method based on artificial intelligence provided by the embodiment of the present application, for example, Figure 3A-3C The information recommendation method based on artificial intelligence is shown.

[0175] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0176] In some embodiments, computer-executable instructions may be in 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 stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0177] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0178] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0179] To sum up, through the embodiments of the present application, the recommendation device completes the recommendation of information and obtains real feedback, and then returns the error to multiple business parties, so that the business parties can update the corresponding parameters based on the error. After each business party completes the differential privacy encoding based on the updated parameters, multiple business parties can send the features obtained by the differential privacy encoding to the recommendation device. The recommendation device fuses the features obtained by multiple differential privacy encodings, so that the availability of the features can still be guaranteed from the statistics obtained by fusion while hiding the second original data. Recommendations are made based on the feature fusion results, so that the features can be comprehensively learned and the recommendation accuracy can be improved.

[0180] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. An information recommendation method based on artificial intelligence, characterized in that: The information recommendation method is applied to a recommendation system, wherein the recommendation system includes a recommendation device and a plurality of feature devices corresponding one-to-one to a plurality of business parties. The method includes: The recommendation device performs a first recommendation operation on first information to the object in a first period, and obtains a first operation label value corresponding to the first information, wherein the first information is at least one information ranked first in descending order based on the first score among multiple information in the first period; Determine a first error of the first period based on the first operation tag value and the first score of the first information, and send the first error to each of the feature devices, so that each of the feature devices performs the following processing: Based on the first error of the first cycle, updating the first parameter to obtain the second parameter of the business party in the second cycle; and based on the second parameter and second original data corresponding to the business party, performing privacy encoding processing on multiple information in the second cycle to obtain a second feature of each information in the second cycle; The second cycle is a next cycle adjacent to the first cycle; For each piece of information in the second period, the recommendation device receives a second feature of the information sent by each feature device, and performs a fusion process based on the exploration mechanism on the second feature to obtain a second score for the information; In the second period, the recommendation device performs a recommendation operation on the object for second information, wherein the second information is at least one piece of information ranked higher in descending order based on the second score among the plurality of pieces of information in the second period.

2. The method according to claim 1, characterized in that Before the recommendation device performs the first recommendation operation of the first information on the object within the first period, the method further includes: For each piece of information in the first period, the recommendation device receives a first feature of the information sent by each feature device, and performs a fusion process based on the exploration mechanism on the first feature to obtain a first score for the information; Among them, for each of the feature devices, the first feature sent by the feature device is obtained by privacy encoding based on the first parameter corresponding to the business party and the first original data corresponding to the business party.

3. The method according to claim 1, characterized in that The acquiring of a first operation tag value corresponding to the first information includes: When an interactive behavior of the subject with respect to the first information is collected within the first period, determining the first operation tag value as a full score in the first score grading standard; When no interactive behavior of the object with respect to the first information is collected within the first period, the first operation tag value is determined to be a zero score in the first score grading standard.

4. The method according to claim 1, wherein The determining, based on the first operation tag value and the first score of the first information, a first error of the first cycle includes: For each piece of the first information, perform any one of the following processes: determining a difference between the first operation label value and the first score, and squaring the difference to obtain a first error of the first information; determining an absolute value between the first operation label value and the first score, and using the absolute value as a first error of the first information; performing a cross entropy calculation on the first operation label value and the first score, and using a cross entropy calculation result as a first error of the first information; When there are multiple first information, summing up the first errors of the multiple first information to obtain the first error of the first period; When the number of the first information is one, the first error of the first information is determined as the first error of the first period.

5. The method according to claim 1, wherein The updating process of the first parameter based on the first error of the first cycle to obtain the second parameter of the business party in the second cycle includes: Obtaining a gradient of the first error of the first period corresponding to the first parameter; Multiplying the gradient by the learning rate of the first cycle to obtain a parameter change value; Subtract the first parameter from the parameter change value to obtain a second parameter of the business party in the second period.

6. The method according to claim 1, characterized in that The step of performing privacy encoding processing on each of the plurality of information in the second period based on the second parameter and the second original data corresponding to the business party to obtain a second feature of each of the information in the second period includes: The following processing is performed for each information of the second cycle: Acquire a recommended reference feature of the information from the second original data, wherein the recommended reference feature includes at least one of the following: an information feature of the information, an object feature of the object; Transposing an original feature matrix composed of the recommended reference features to obtain an original transposed matrix; Multiplying the original transposed matrix by the second parameter to obtain a coding feature of the information; Gaussian noise that meets the differential privacy standard is superimposed on the encoded feature of the information to obtain a second feature of the information.

7. The method according to claim 1, characterized in that The performing a fusion process based on an exploration mechanism on the second feature to obtain a second score of the information includes: fusing the second features sent by the plurality of feature devices to obtain a third score; performing splicing processing on the second features sent by the plurality of feature devices to obtain a spliced ​​feature corresponding to the second period; performing exploration processing based on the splicing features corresponding to the second period to obtain a fourth score for the information; The sum of the third score and the fourth score is used as the second score.

8. The method according to claim 7, characterized in that The fusing the second features sent by the plurality of feature devices to obtain a third score includes: The second features sent by the plurality of feature devices are summed to obtain the third score.

9. The method according to claim 7, characterized in that The fusing the second features sent by the plurality of feature devices to obtain a third score includes: Obtaining the weight of each of the business parties; Based on the weight of each of the business parties, weighted summation is performed on the second features sent by the feature devices of multiple business parties to obtain the third score.

10. The method according to claim 7, characterized in that The performing exploration processing based on the splicing feature to obtain a fourth score of the information includes: Obtaining a splicing feature matrix corresponding to the second period composed of splicing features corresponding to the second period; performing a transposition process on the splicing feature matrix corresponding to the second period to obtain a splicing transposed matrix corresponding to the second period; Multiplying the splicing transposed matrix corresponding to the second period by the inverse matrix of the historical feature matrix corresponding to the second period to obtain a first multiplication matrix, and multiplying the first multiplication matrix by the splicing feature matrix corresponding to the second period to obtain a second multiplication matrix; Performing a square root process on the second multiplication matrix to obtain a square root matrix, and acquiring a norm corresponding to the square root matrix; A fourth score positively correlated with the norm is obtained.

11. The method according to claim 10, characterized in that The method further comprises: Transposing the splicing feature matrix corresponding to the first period to obtain a splicing transposed matrix corresponding to the first period; Multiplying the splicing feature matrix corresponding to the first period and the splicing transposed matrix corresponding to the first period to obtain a third multiplication matrix; Obtaining a historical feature matrix corresponding to the first period; The historical feature matrix corresponding to the first period is added to the third multiplication matrix to obtain a historical feature matrix corresponding to the second period.

12. An information recommendation device based on artificial intelligence, characterized in that: The information recommendation device is applied to a recommendation system, which includes a recommendation device and a plurality of feature devices corresponding to a plurality of business parties. The device includes: a recommendation module, configured to, within a first period, execute, by the recommendation device, a first recommendation operation on first information to an object, and obtain a first operation label value corresponding to the first information, wherein the first information is at least one of the plurality of information in the first period that is ranked first in descending order based on the first score; an interaction module, configured to determine a first error of the first period based on the first operation tag value and the first score of the first information, and send the first error to each of the feature devices, so that each of the feature devices performs the following processing: Based on the first error of the first cycle, updating the first parameter to obtain the second parameter of the business party in the second cycle, and performing privacy encoding processing on multiple information in the second cycle based on the second parameter and second original data corresponding to the business party to obtain a second feature of each information in the second cycle; The second cycle is a cycle adjacent to the first cycle, the first parameter is a decision parameter of the service party of the feature device in the first cycle, and the second parameter is a decision parameter of the service party of the feature device in the second cycle; a fusion module, configured to, for each piece of information in the second period, receive, by the recommendation device, the second feature of the information sent by each feature device, and perform fusion processing on the second feature based on the exploration mechanism to obtain a second score for the information; The recommendation module is further configured to, within the second period, enable the recommendation device to perform a recommendation operation on second information to the object, wherein the second information is at least one piece of information ranked higher in descending order based on the second score among multiple pieces of information in the second period.

13. An electronic device, characterized in that: The electronic device comprises: a memory for storing computer-executable instructions; The processor is configured to implement the artificial intelligence-based information recommendation method according to any one of claims 1 to 11 when executing the computer-executable instructions stored in the memory.

14. A computer-readable storage medium storing computer-executable instructions, characterized in that: When the computer executable instructions are executed by a processor, the artificial intelligence-based information recommendation method according to any one of claims 1 to 11 is implemented.

15. A computer program product comprising a computer program or computer executable instructions, characterized in that When the computer program or computer executable instructions are executed by a processor, the artificial intelligence-based information recommendation method according to any one of claims 1 to 11 is implemented.

Images