A model training method and device based on edge computing

Abstract

The present disclosure relates to the field of computer technology, and in particular to a model training method and device based on edge computing, which are used to solve the problem of inaccurate recommendation solutions given by a recommendation model. The method is as follows: within the jurisdiction of the edge computing device, the sample user data of each user collected is input into a first recommendation model for federal training, and when a preset training end condition is met, the target first recommendation model obtained by training is output, wherein a piece of sample user data includes the user's human feature information and real behavior information, and the first recommendation model is used to predict the user's behavior and recommend related objects; in this way, the first recommendation model can be federally trained on the edge computing device without leaking user privacy and reducing network data transmission volume, thereby obtaining a more accurate and more widely applicable target first recommendation model, and then making more accurate behavior predictions for new users and formulating more accurate recommendation solutions.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a model training method and device based on edge computing. Background Art

[0002] In various industries, how to make appropriate and accurate information recommendations based on user preferences is becoming increasingly important. Therefore, the establishment of the first recommendation model has been widely used in various industries.

[0003] For example, in the retail industry, by capturing the user's preferences through the first recommendation model and recommending a type of food to the user, the user can have a satisfying food experience.

[0004] In the prior art, the training process of the first recommendation model is often performed on the network side, which inevitably requires the collection of a large amount of sample user data. However, the sample user data collected on the network side has the following defects:

[0005] 1) The sample user data obtained has a single dimension.

[0006] In order to obtain an accurate and effective first recommendation model, the network side needs to collect a large amount of sample user data. However, due to the restrictions of privacy protection clauses, the collection of a large amount of typical sample user data has encountered major obstacles.

[0007] For example, in the retail industry, sample user data is obtained mainly from user order information in a retail location and membership information pre-entered by users. Although this sample user data is typical, due to privacy protection regulations, the network side cannot obtain effective typical sample user data, making the accuracy of the first recommendation model obtained through training unable to meet usage requirements.

[0008] 2) The amount of sample user data obtained is huge. If all of it is concentrated on the network side for training, the excessive amount and concentration of sample user data will inevitably cause an operating burden on the network side equipment and even lead to insufficient computing power.

[0009] In view of this, it is necessary to redesign a model training system, application method and device based on edge computing to overcome the above-mentioned defects. Summary of the Invention

[0010] The present disclosure provides a model training method and device based on edge computing, which is used to solve the problems of the single dimension of sample user data information obtained in the prior art and the huge amount of information of the sample user data, which leads to insufficient computing power of network-side equipment, thereby causing the trained model to be inaccurate and having regional deviations.

[0011] The specific technical solutions provided by this disclosure are as follows:

[0012] First, a model training system based on edge computing is proposed, including several edge computing devices, wherein:

[0013] Any edge computing device that performs the following operations:

[0014] Collect sample user data of each user within the jurisdiction, where a piece of sample user data includes the user's physical characteristics information and real behavior information;

[0015] Based on the obtained sample user data, a first recommendation model is federated trained, wherein the first recommendation model is used to predict user behavior and recommend related objects;

[0016] When the preset training end condition is met, the target first recommendation model obtained through training is output.

[0017] Optionally, within the jurisdiction, sample user data of each user is collected, wherein a piece of sample user data includes the user's physical feature information and real behavior information, including:

[0018] Within the jurisdiction, the appearance attribute information of each user is collected and analyzed by smart cameras as human feature information;

[0019] Also, within the jurisdiction, the behavior-related object attribute information of each user is collected and analyzed through the point-of-sale information management system POS machine as real behavior information.

[0020] Optionally, before performing federated training on the first recommendation model based on the obtained sample user data, the method further includes:

[0021] The user identification ID of each sample user data is homomorphically encrypted and sent to a third-party device;

[0022] Based on the feedback result of the third-party device, filter out the user IDs that have appeared on each other specified edge computing device from each user ID;

[0023] Based on the filtered user ID, the corresponding sample user data is further filtered out.

[0024] Optionally, the performing federated training on the first recommendation model based on the obtained sample user data includes:

[0025] Each time sample user data is input into the first recommendation model, a corresponding first gradient is obtained based on an error between the output predicted behavior information and the actual behavior information;

[0026] Use federated learning to obtain the second gradient sent by each other edge computing device;

[0027] Parameters of the first recommendation model are adjusted based on the obtained first gradient and each second gradient.

[0028] Optionally, the method of adopting a federated learning approach to obtain a second gradient sent by each other edge computing device includes:

[0029] Send the specified user ID to the third-party device;

[0030] Based on the feedback results of the third-party device, the second gradients reported by the designated other edge computing devices are obtained, wherein a second gradient is obtained by training the corresponding second recommendation model by a designated other edge computing device based on the sample user data corresponding to the designated user ID.

[0031] Optionally, adjusting parameters of the first recommendation model based on the obtained first gradient and second gradient includes:

[0032] Linearly superimposing the obtained first gradient and each second gradient, and then adjusting the parameters of the first recommendation model based on the third gradient obtained after the superposition; or

[0033] The obtained second gradients are combined to obtain a fourth gradient, the fourth gradient is combined with the first gradient again to obtain a fifth gradient, and parameters of the first recommendation model are adjusted based on the fifth gradient.

[0034] Optionally, after outputting the target first recommendation model obtained through training, the method further includes:

[0035] Obtaining human feature information to be predicted, and inputting the human feature information to be predicted into the target first recommendation model to obtain corresponding predicted behavior information, and recommending corresponding associated objects based on the predicted behavior information; or

[0036] Obtain the human feature information to be predicted and the corresponding user ID to be predicted, and through a third-party device, determine whether the user ID to be predicted appears within the jurisdiction of other designated edge computing devices, input the human feature information to be predicted into the target first recommendation model, obtain corresponding predicted behavior information, and recommend corresponding associated objects based on the predicted behavior information.

[0037] Secondly, a model training method based on edge computing is proposed, including:

[0038] receiving an acquisition request from an edge computing device, the acquisition request being sent by the edge computing device during a process of federated training of a first recommendation model based on sample user data of each user, the first recommendation model being used to predict user behavior and recommend associated objects, wherein a piece of sample user data includes human feature information and actual behavior information of the user;

[0039] The information reported by each designated other edge computing device is fed back to the edge computing device, so that the edge computing device adjusts the parameters of the first recommendation model.

[0040] Optionally, before receiving the acquisition request from the edge computing device, the method further includes:

[0041] Accepting a screening request from the edge computing device, wherein the screening request carries the user ID of each sample user data collected by the edge computing device, and the user ID is homomorphically encrypted;

[0042] Based on the reported information of each designated other edge computing device, the obtained user ID is filtered, and the filtered user ID is notified to the edge computing device, so that the edge computing device trains the first recommendation model based on the sample user data corresponding to the filtered user ID, and the filtered user ID has appeared in each designated other edge computing device.

[0043] Optionally, transmitting information reported by each designated other edge computing device to the edge computing device, so that the edge computing device adjusts parameters of the first recommendation model, includes:

[0044] The second gradient reported by each designated other edge computing device is fed back to the edge computing device, so that the edge computing device adjusts the parameters of the first recommendation model based on the obtained first gradient and each second gradient, wherein the first gradient is obtained by the edge computing device through federated training of the first recommendation model based on the sample user data of each user, and the each second gradient is obtained by each designated other edge computing device through training the corresponding second recommendation model based on its own sample user data.

[0045] Optionally, feeding back the second gradient reported by each designated other edge computing device to the edge computing device includes:

[0046] Based on the specified user ID carried in the acquisition request, the second gradient reported by each of the specified other edge computing devices corresponding to the specified user ID is fed back to the edge computing device; wherein, the second gradient is obtained by each of the specified other edge computing devices, respectively based on the sample user data corresponding to the specified user ID within their respective jurisdictions, by training the corresponding second recommendation model.

[0047] In a third aspect, a model training device based on edge computing includes:

[0048] Any edge computing device includes the following:

[0049] A collection unit is used to collect sample user data of each user within the jurisdiction, wherein a piece of sample user data includes the user's physical feature information and real behavior information;

[0050] a training unit, configured to perform federated training on a first recommendation model based on the obtained sample user data, wherein the first recommendation model is used to predict user behavior and recommend related objects;

[0051] The output unit is used to output the target first recommendation model obtained through training when a preset training end condition is met.

[0052] Optionally, within the jurisdiction, sample user data of each user is collected, wherein a piece of sample user data includes the user's body feature information and real behavior information, and the collection unit is used to:

[0053] Within the jurisdiction, the appearance attribute information of each user is collected and analyzed by smart cameras as human feature information;

[0054] Also, within the jurisdiction, the behavior-related object attribute information of each user is collected and analyzed through the point-of-sale information management system POS machine as real behavior information.

[0055] Optionally, before performing federated training on the first recommendation model based on the obtained sample user data, the collection unit is further configured to:

[0056] The user identification ID of each sample user data is homomorphically encrypted and sent to a third-party device;

[0057] Based on the feedback result of the third-party device, filter out the user IDs that have appeared on each other specified edge computing device from each user ID;

[0058] Based on the filtered user ID, the corresponding sample user data is further filtered out.

[0059] Optionally, the first recommendation model is federated trained based on the obtained sample user data, and the training unit is configured to:

[0060] Each time sample user data is input into the first recommendation model, a corresponding first gradient is obtained based on an error between the output predicted behavior information and the actual behavior information;

[0061] Use federated learning to obtain the second gradient sent by each other edge computing device;

[0062] Parameters of the first recommendation model are adjusted based on the obtained first gradient and each second gradient.

[0063] Optionally, the federated learning method is adopted to obtain the second gradient sent by each other edge computing device, and the training unit is used to:

[0064] Send the specified user ID to the third-party device;

[0065] Based on the feedback results of the third-party device, the second gradients reported by the designated other edge computing devices are obtained, wherein a second gradient is obtained by training the corresponding second recommendation model by a designated other edge computing device based on the sample user data corresponding to the designated user ID.

[0066] Optionally, the first recommendation model is adjusted in parameters based on the obtained first gradient and second gradient, and the training unit is configured to:

[0067] Linearly superimposing the obtained first gradient and each second gradient, and then adjusting the parameters of the first recommendation model based on the third gradient obtained after the superposition; or

[0068] The obtained second gradients are combined to obtain a fourth gradient, the fourth gradient is combined with the first gradient again to obtain a fifth gradient, and parameters of the first recommendation model are adjusted based on the fifth gradient.

[0069] Optionally, after outputting the target first recommendation model obtained through training, the training unit is further configured to:

[0070] Obtaining human feature information to be predicted, and inputting the human feature information to be predicted into the target first recommendation model to obtain corresponding predicted behavior information, and recommending corresponding associated objects based on the predicted behavior information; or

[0071] Obtain the human feature information to be predicted and the corresponding user ID to be predicted, and through a third-party device, determine whether the user ID to be predicted appears within the jurisdiction of other designated edge computing devices, input the human feature information to be predicted into the target first recommendation model, obtain corresponding predicted behavior information, and recommend corresponding associated objects based on the predicted behavior information.

[0072] In a fourth aspect, a device for model training based on edge computing includes:

[0073] a first communication unit, configured to receive an acquisition request from an edge computing device, the acquisition request being sent by the edge computing device during federated training of a first recommendation model based on sample user data of each user; the first recommendation model being configured to predict user behavior and recommend associated objects; wherein a piece of sample user data includes a user's body feature information and actual behavior information;

[0074] The second communication unit is used to feed back information reported by each designated other edge computing device to the edge computing device, so that the edge computing device adjusts parameters of the first recommendation model.

[0075] Optionally, before receiving the acquisition request from the edge computing device, the first communication unit is further configured to:

[0076] Accepting a screening request from the edge computing device, wherein the screening request carries the user ID of each sample user data collected by the edge computing device, and the user ID is homomorphically encrypted;

[0077] Based on the reported information of each designated other edge computing device, the obtained user ID is filtered, and the filtered user ID is notified to the edge computing device, so that the edge computing device trains the first recommendation model based on the sample user data corresponding to the filtered user ID, and the filtered user ID has appeared in each designated other edge computing device.

[0078] Optionally, the information reported by each designated other edge computing device is transmitted to the edge computing device, so that the edge computing device adjusts parameters of the first recommendation model, and the second communication unit is used to:

[0079] The second gradient reported by each designated other edge computing device is fed back to the edge computing device, so that the edge computing device adjusts the parameters of the first recommendation model based on the obtained first gradient and each second gradient, wherein the first gradient is obtained by the edge computing device through federated training of the first recommendation model based on the sample user data of each user, and the each second gradient is obtained by each designated other edge computing device through training the corresponding second recommendation model based on its own sample user data.

[0080] Optionally, the second gradient reported by each designated other edge computing device is fed back to the edge computing device, and the second communication unit is used to:

[0081] Based on the specified user ID carried in the acquisition request, the second gradient reported by each of the specified other edge computing devices corresponding to the specified user ID is fed back to the edge computing device; wherein, the second gradient is obtained by each of the specified other edge computing devices, respectively based on the sample user data corresponding to the specified user ID within their respective jurisdictions, by training the corresponding second recommendation model.

[0082] In a fifth aspect, an edge computing device is proposed, including:

[0083] a memory for storing executable instructions;

[0084] A processor is used to read and execute the executable instructions stored in the memory to implement the method described in any one of the first aspects above.

[0085] In a sixth aspect, a network device is provided, comprising:

[0086] a memory for storing executable instructions;

[0087] A processor is used to read and execute the executable instructions stored in the memory to implement the method described in any one of the second aspects above.

[0088] In a seventh aspect, a computer-readable storage medium is proposed. When the instructions in the computer-readable storage medium are executed by a processor, the processor is enabled to execute any method described in the first aspect.

[0089] In an eighth aspect, a computer-readable storage medium is proposed. When the instructions in the computer-readable storage medium are executed by a processor, the processor is enabled to execute any method described in the second aspect.

[0090] In an embodiment of the present disclosure, the edge computing device inputs the collected sample user data of each user into the first recommendation model within its jurisdiction for federated training, and outputs the target first recommendation model obtained through training when the preset training end conditions are met, wherein a piece of sample user data includes the user's human feature information and real behavior information, and the first recommendation model is used to predict the user's behavior and recommend related objects; in this way, the first recommendation model can be federated trained on the edge computing device without leaking user privacy and reducing the amount of network data transmission, thereby obtaining a more accurate and more widely applicable target first recommendation model, and then based on the target first recommendation model, more accurate behavior predictions can be made for new users, and more accurate recommendation plans can be formulated. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0092] Figure 1A A schematic diagram of a hardware architecture provided in an embodiment of the present disclosure;

[0093] Figure 1B A schematic diagram of the deployment of various edge computing devices in an embodiment of the present disclosure;

[0094] Figure 1C A schematic diagram of the interaction between an edge computing device and other edge computing devices in an embodiment of the present disclosure;

[0095] Figure 2 This is a schematic diagram of the process of training the first recommendation model based on edge computing in an embodiment of the present disclosure;

[0096] Figure 3 This is a schematic diagram of the process of training the first recommendation model based on edge computing assistance in an embodiment of the present disclosure;

[0097] Figure 4 This is a schematic diagram of the logical architecture of the edge computing device in the embodiment of the present disclosure;

[0098] Figure 5 This is a schematic diagram of the physical architecture of the edge computing device in an embodiment of the present disclosure;

[0099] Figure 6 This is a schematic diagram of the logical architecture of a network device in an embodiment of the present disclosure;

[0100] Figure 7 FIG. 1 is a schematic diagram of the physical architecture of a network device in an embodiment of the present disclosure. FIG. DETAILED DESCRIPTION

[0101] In order to solve the problem of single dimension of sample user data obtained on the network side and insufficient computing power caused by obtaining a huge amount of sample user data, in the embodiment of the present disclosure, an edge computing device is set based on artificial intelligence (AI). The edge computing device refers to an open platform that is physically close to the edge of the network and is used to integrate network, computing, storage, and application core capabilities. The edge computing device can be used to collect sample user data on the user side and complete the training of the first recommendation model. Most importantly, the trained first recommendation model can be used to predict user behavior and recommend related objects.

[0102] Figure 1A Schematic diagram of a hardware architecture provided in an embodiment of the present disclosure. Figure 1A As shown, the hardware architecture 100 may include an edge computing device 101 and a network device 102;

[0103] The edge computing device 101 is connected to each device 103 and is used to obtain corresponding sample user data based on each device 103;

[0104] The network device 102 is used for data transmission between the edge computing devices 101 .

[0105] Optionally, in the embodiment of the present disclosure, the edge computing devices 101 may be interconnected through the network device 102, or may be directly interconnected with each other without the help of the network device 102.

[0106] The preferred embodiments of the present disclosure are further described in detail below with reference to the accompanying drawings.

[0107] For example, see Figure 1B As shown, taking traditional retail as an example, it is assumed that a computer deployed in a retail store can be used as an edge computing device.

[0108] And assume that in a retail store, smart cameras can be used as collection devices to collect human characteristics information of users visiting the store;

[0109] And it is assumed that in a retail store, a point of sales (POS) machine can be used as a collection device for collecting real behavior information of users who come to the store.

[0110] Edge computing devices can be deployed in each area, so that a large amount of sample user data can be obtained by expanding the jurisdiction of the edge computing devices. Then, the first recommendation model obtained based on the sample user data has a wider applicability. In actual applications, each edge computing device can be interconnected through a third-party device, or can be directly interconnected without the help of a third-party device. Figure 1C As shown, in the embodiment of the present disclosure, the first case is taken as an example for description.

[0111] Based on the above distributed architecture, see Figure 2 As shown, in the subsequent embodiments, taking an edge computing device as an example (hereinafter referred to as edge computing device X), the detailed training process of the first recommendation model is introduced as follows:

[0112] Step 200: The edge computing device X collects sample user data of each user within its jurisdiction, where each piece of sample user data includes the user's body feature information and real behavior information.

[0113] In a specific implementation, the edge computing device X can collect sample user data of each user through the connected collection device within its jurisdiction.

[0114] For example, see Figure 1B As shown in the figure, taking the traditional retail industry as an example, assuming that a computer deployed in a retail store can serve as an edge computing device X, then edge computing device X can collect data through smart cameras and obtain human feature information after analysis; and collect data through POS terminals and obtain real behavior information after analysis.

[0115] When edge computing device X collects data through a smart camera and obtains human feature information after analysis, it includes: within its jurisdiction, edge computing device X can use smart cameras in retail stores to collect the appearance characteristics of users visiting the store, such as facial expressions, clothing, etc.; further, because the smart camera is equipped with a chip, the smart camera can further analyze the user's age and mood when visiting the store based on the user's appearance characteristics. Therefore, edge computing device X can use the appearance attribute information collected by the smart camera for each user as human feature information, where the appearance attribute information includes, but is not limited to, any one or any combination of the following information:

[0116] gender;

[0117] Dress;

[0118] age;

[0119] mood;

[0120] etc.

[0121] When edge computing device X collects data through a POS machine and obtains real behavior information after analysis, it includes: within the jurisdiction, edge computing device X can collect orders generated by users who visit the store through a POS machine in a retail store, and obtain behavior-related object attribute information of each user as real behavior information by analyzing the orders. The behavior-related object attribute information includes, but is not limited to, any one or any combination of the following information:

[0122] The name of the associated object;

[0123] The price of the associated object;

[0124] Attributes of the associated object;

[0125] etc.

[0126] For example, assume that the jurisdiction of edge computing device X is a milk tea shop. Then, assume that the order information of the user who comes to the store and can be obtained by edge computing device X through the POS machine in the retail store is: milk tea, 5 yuan, slightly sweet; then, the corresponding behavior-related object attribute information can be any one or any combination of milk tea, 5 yuan, and slightly sweet. Accordingly, the above behavior-related object attribute information can be used as the real behavior information.

[0127] For example, assume that the jurisdiction of edge computing device X is a coffee shop. Suppose that the order information of the user who came to the store and can be obtained by edge computing device X through the retail store's POS machine is: Blue Mountain coffee, 20 yuan, no sugar, no milk; then the corresponding behavior-related object attribute information can be any one or any combination of Blue Mountain coffee, 20 yuan, no sugar, no milk, and accordingly, the above behavior-related object attribute information can be used as the real behavior information.

[0128] For example, assume that the jurisdiction of edge computing device X is a gym. Suppose that the order information of the user who came to the store and obtained by edge computing device X through the POS machine of the retail store is: sports drink, 10 yuan, sugar-free; then, the corresponding behavior-related object attribute information can be any one or any combination of sports drink, 10 yuan, and sugar-free. Accordingly, the above behavior-related object attribute information can be used as the real behavior information.

[0129] Furthermore, after executing step 200 and before executing step 210 , that is, after obtaining each sample user data and before performing federated training on the first recommendation model, the sample user data needs to be screened.

[0130] In the embodiment of the present disclosure, optionally, a vertical federated learning method is adopted to perform joint training of the first recommendation model between different edge computing devices. Therefore, it is necessary to perform sample alignment first, that is, to filter out sample user data generated by users who have appeared within the jurisdiction of different edge computing devices from each sample user data.

[0131] Specifically include:

[0132] a) Edge computing device X homomorphically encrypts the user ID of each sample user data obtained and sends it to a third-party device.

[0133] In practice, to prevent user privacy leaks, edge computing device X obtains the user IDs of each sample user within its jurisdiction and then homomorphically encrypts them before sending them to a third-party device. Because homomorphic encryption ensures data consistency before and after encryption, the third-party device can accurately match the user IDs provided by edge computing device X with those provided by other edge computing devices based on the homomorphically encrypted user IDs, thereby filtering out identical user IDs.

[0134] For example, assuming that the jurisdiction of edge computing device X is a milk tea shop, then edge computing device X will homomorphically encrypt the collected user ID "Xiao Wang" and send it to a third-party device, thereby avoiding the leakage of personal privacy information.

[0135] b) Based on the feedback result of the third-party device, the edge computing device X filters out the user IDs that have appeared on other designated edge computing devices from the user IDs.

[0136] In a specific implementation, the third-party device will notify edge computing device X of which user IDs have appeared within the jurisdiction of other designated edge computing devices based on the user IDs sent by edge computing device X. Then, edge computing device X will filter out these user IDs.

[0137] Optionally, the number of designated other edge computing devices may be two or more. For ease of description, the embodiment of the present disclosure is described by taking the example of the presence of two designated other edge computing devices.

[0138] For example, the jurisdiction of edge computing device X is a milk tea shop, the jurisdiction of the first designated other edge device is a coffee shop, and the jurisdiction of the second designated other edge device is a gym.

[0139] Suppose that a third-party device compares the user IDs sent by edge computing device X with the user IDs of other designated edge computing devices and notifies edge computing device X that user IDs such as "Xiao Wang" and "Xiao Zhang" that appear in the milk tea shop also appear in the designated coffee shop and gym. In this case, edge computing device X filters out "Xiao Wang" and "Xiao Zhang."

[0140] c) The edge computing device X further filters out the corresponding sample user data based on the filtered user ID.

[0141] In a specific implementation, edge computing device X filters out sample user data corresponding to the filtered user IDs, thereby obtaining sample user data for subsequent training of the first recommendation model. Optionally, each of the other designated edge computing devices filters out sample user data corresponding to the filtered user IDs, thereby obtaining sample user data for subsequent training of their respective second recommendation models.

[0142] For example, the jurisdiction of edge computing device X is still the milk tea shop, the jurisdiction of the first other edge device is the coffee shop, and the jurisdiction of the second other edge device is the gym.

[0143] Assume that edge computing device X, based on notification from a third-party device, filters out the user IDs "Xiao Wang" and "Xiao Zhang." Then, further, the sample user data corresponding to "Xiao Wang" and "Xiao Zhang" are filtered out from the sample user data of each milk tea shop.

[0144] Optionally, based on notifications from third-party devices, sample user data corresponding to "Xiao Wang" and sample user data corresponding to "Xiao Zhang" are filtered out from various sample user data in the coffee shop; based on notifications from third-party devices, sample user data corresponding to "Xiao Wang" and sample user data corresponding to "Xiao Zhang" are also filtered out from various sample user data in the gym.

[0145] Step 210: The edge computing device X performs federated training on a first recommendation model based on the obtained sample user data, wherein the first recommendation model is used to predict user behavior and recommend related objects.

[0146] In the specific implementation, taking the training of the first recommendation model as an example, the training process of the first recommendation model is introduced as follows:

[0147] Edge computing device X inputs the selected sample user data into the first recommendation model for federated training. During the training process, each time the sample user data is input into the first recommendation model, a first gradient is obtained based on the error between the output predicted behavior information and the actual behavior information. Then, federated learning is used to obtain the respective second gradients reported by other designated edge computing devices through a third-party device. The second gradients are obtained by training the corresponding second recommendation models based on the sample user data corresponding to the same user ID within the jurisdiction of the other designated edge computing devices. Parameters of the first recommendation model are then adjusted based on the obtained first gradient and each second gradient.

[0148] Optionally, a third-party device can set a sample training frequency between each edge computing device so that the generation frequency of the first gradient and each second gradient is consistent, which facilitates subsequent model parameter adjustment.

[0149] For example, a third-party device sets up edge computing device X and other edge computing devices. During the iteration process, N sample user data are input each time to perform federated training on the first recommendation model and the second recommendation model used by each of them, thereby generating the first gradient and each second gradient at the same frequency.

[0150] In a specific implementation, during one iteration, edge computing device X sends a designated user ID to a third-party device, where there can be multiple designated user IDs. Based on the notification from the third-party device, each of the other designated edge computing devices then inputs the sample user data corresponding to the designated user ID into their respective second recommendation models for federated training, thereby obtaining respective second gradients. The second gradients are then sent to the third-party device. Based on the feedback from the third-party device, edge computing device X obtains the respective second gradients reported by the other designated edge devices.

[0151] For example, let's assume that the jurisdiction of edge computing device X is the milk tea shop, and the jurisdiction of other designated edge devices is the coffee shop.

[0152] Assume that the user ID specified by edge computing device X is "Xiao Wang". Edge computing device X trains the first recommendation model based on the sample user data corresponding to "Xiao Wang", thereby obtaining a first gradient. Edge computing device X then sends the specified user ID "Xiao Wang" to a third-party device. Based on the feedback from the third-party device, the coffee shop inputs the sample user data corresponding to "Xiao Wang" into the corresponding second recommendation model, thereby obtaining a corresponding second gradient. The second gradient is then sent to the third-party device. Edge computing device X then obtains the second gradient based on the feedback from the third-party device.

[0153] Furthermore, the edge computing device X can adjust the parameters of the first recommendation model based on the first gradient obtained by training the first recommendation model, combined with the second gradients obtained by training the second recommendation models reported by other designated edge computing devices provided by the third-party device. During the adjustment process, the first recommendation model includes model features preset for the jurisdiction of the corresponding edge computing device X. Therefore, during the adjustment process, whether the first gradient or the second gradients are used, the parameters are adjusted for the above-mentioned model features.

[0154] On the other hand, the combination of the first gradient and each second gradient can be flexibly adjusted according to the algorithm used to train the first recommendation model.

[0155] For example, if the logistic regression (LR) algorithm is used to perform federated training on the first recommendation model, the obtained first gradient and each second gradient can be directly linearly superimposed, and then the parameters of the first recommendation model are adjusted based on the third gradient obtained after superposition.

[0156] For another example, if the gradient descent tree (Gradient Boosting Decision Tree, GBDT) algorithm is used to perform federated training on the first recommendation model, the second gradients obtained are merged to obtain a fourth gradient, and then the fourth gradient is merged with the first gradient again to obtain a fifth gradient. Finally, based on the fifth gradient, the parameters of the first recommendation model are adjusted.

[0157] Step 220: When the preset training end condition is met, the edge computing device X outputs the target first recommendation model obtained through training.

[0158] In a specific implementation, after multiple rounds of iterative training, the edge computing device X can obtain a trained first recommendation model, and can recommend related objects to new users based on the obtained trained first recommendation model.

[0159] Specifically, edge computing device X obtains the human feature information to be predicted, inputs the human feature information to the target first recommendation model, obtains corresponding predicted behavior information, and recommends corresponding associated objects based on the predicted behavior information. Optionally, edge computing device X can obtain the user ID to be predicted from the human feature information to be predicted, and then record and identify the user ID to be predicted, so that other edge computing devices can feedback sample user data corresponding to the user ID to be predicted to the third-party device when querying the third-party device.

[0160] For example, the jurisdiction of edge computing device X is still the milk tea shop, the jurisdiction of the first other edge device is the coffee shop, and the jurisdiction of the second other edge device is the gym.

[0161] Suppose that the smart camera in a milk tea shop collects human feature information to be predicted, and inputs the human feature information (such as female, 20 years old, student clothing, lost) into the target first recommendation model, and obtains the corresponding predicted behavior information as "milk tea, 10 yuan, slightly sweet", and then based on the predicted behavior information, the corresponding related object is recommended as "original milk tea".

[0162] Optionally, the edge computing device X obtains the human feature information to be predicted and the corresponding user ID to be predicted, and determines through a third-party device whether the user ID to be predicted appears within the jurisdiction of other designated edge computing devices, inputs the human feature information to be predicted into the target first recommendation model, obtains corresponding predicted behavior information, and recommends corresponding associated objects based on the predicted behavior information.

[0163] For example, the jurisdiction of edge computing device X is still the milk tea shop, the jurisdiction of the first other edge device is the coffee shop, and the jurisdiction of the second other edge device is the gym.

[0164] Assume that the smart camera in a milk tea shop collects human feature information to be predicted, and the corresponding user ID to be predicted is "Xiao Li". Through a third-party device, it is determined that the user ID to be predicted has appeared in a coffee shop and a gym. The human feature information of "Xiao Li" (such as female, 25 years old, office worker attire, lost) is input into the target first recommendation model, and the predicted behavior information of "Xiao Li" can be obtained as "milk tea, 20 yuan, sugar-free", and then the corresponding related object is recommended as "refreshing milk green" based on the predicted behavior information.

[0165] On the other hand, during the above-mentioned iterative training process, edge computing device X also needs to send its own first gradient to other edge computing devices so that other edge computing devices can train their respective second recommendation models. The specific process is similar to the process of training the first recommendation model. The only difference is that the second recommendation model is used to predict the behavior of users within the jurisdiction of the other edge computing devices and recommend related objects. It will not be repeated here.

[0166] Corresponding to the above embodiment, in the above process, refer to Figure 3 As shown in the figure, the specific process executed by the third-party device is as follows:

[0167] Step 300: A third-party device receives an acquisition request from edge computing device X, where the acquisition request is sent by edge computing device X during federated training of a first recommendation model based on sample user data of each user. The first recommendation model is used to predict user behavior and recommend related objects. A piece of sample user data includes the user's physical characteristics and actual behavior information.

[0168] Specifically, before receiving the acquisition request from edge computing device X, the third-party device may also accept a screening request from edge computing device X. The screening request carries the user IDs of each sample user data collected by edge computing device X, and the user IDs are homomorphically encrypted. Then, the third-party device screens the obtained user IDs based on the reported information from each designated other edge computing device, and notifies edge computing device X of the screened user IDs, instructing edge computing device X to perform federated training on the first recommendation model based on the sample user data corresponding to the screened user IDs, where the screened user IDs have appeared on each designated other edge computing device.

[0169] Step 310: The third-party device feeds back the information reported by each designated other edge computing device to the edge computing device X, instructing the edge computing device to adjust the parameters of the first recommendation model.

[0170] In a specific implementation, the third-party device, based on receiving the acquisition request from the edge computing device X, feeds back the second gradient reported by each designated other edge computing device to the edge computing device, so that the edge computing device adjusts the parameters of the first recommendation model based on the obtained first gradient and each second gradient, wherein the first gradient is obtained by the edge computing device through federated training of the first recommendation model based on the sample user data of each user, and the each second gradient is obtained by the each designated other edge computing device through training the corresponding second recommendation model based on its own sample user data.

[0171] Optionally, the third-party device feeds back the second gradient reported by each of the designated other edge computing devices corresponding to the designated user ID to the edge computing device X based on the designated user ID carried in the acquisition request; wherein the second gradient is obtained by each of the designated other edge computing devices, based on the sample user data corresponding to the designated user ID within their respective jurisdictions, by federally training the corresponding second recommendation model.

[0172] Based on the same inventive concept, see Figure 4As shown, in an embodiment of the present disclosure, an edge computing device is provided, which includes at least: a collection unit 401, a training unit 402, and an output unit 403, wherein:

[0173] The collection unit 401 is used to collect sample user data of each user within the jurisdiction, wherein a piece of sample user data includes the user's body feature information and real behavior information;

[0174] A training unit 402 is configured to perform federated training on a first recommendation model based on the obtained sample user data, wherein the first recommendation model is configured to predict user behavior and recommend related objects;

[0175] The output unit 403 is configured to output the target first recommendation model obtained through training when a preset training end condition is met.

[0176] Optionally, within the jurisdiction, sample user data of each user is collected, wherein a piece of sample user data includes the user's body feature information and real behavior information, and the collection unit 401 is used to:

[0177] Within the jurisdiction, the appearance attribute information of each user is collected and analyzed by smart cameras as human feature information;

[0178] Also, within the jurisdiction, the behavior-related object attribute information of each user is collected and analyzed through the point-of-sale information management system POS machine as real behavior information.

[0179] Optionally, before performing federated training on the first recommendation model based on the obtained sample user data, the collection unit 401 is further configured to:

[0180] The user identification ID of each sample user data is homomorphically encrypted and sent to a third-party device;

[0181] Based on the feedback result of the third-party device, filter out the user IDs that have appeared on each other specified edge computing device from each user ID;

[0182] Based on the filtered user ID, the corresponding sample user data is further filtered out.

[0183] Optionally, the first recommendation model is federated trained based on the obtained sample user data, and the training unit 402 is configured to:

[0184] Each time sample user data is input into the first recommendation model, a corresponding first gradient is obtained based on an error between the output predicted behavior information and the actual behavior information;

[0185] Use federated learning to obtain the second gradient sent by each other edge computing device;

[0186] Parameters of the first recommendation model are adjusted based on the obtained first gradient and each second gradient.

[0187] Optionally, the federated learning method is adopted to obtain the second gradient sent by each other edge computing device, and the training unit 402 is used to:

[0188] Send the specified user ID to the third-party device;

[0189] Based on the feedback results of the third-party device, the second gradients reported by the designated other edge computing devices are obtained, wherein a second gradient is obtained by training the corresponding second recommendation model by a designated other edge computing device based on the sample user data corresponding to the designated user ID.

[0190] Optionally, the first recommendation model is adjusted in parameters based on the obtained first gradient and second gradient, and the training unit 402 is configured to:

[0191] Linearly superimposing the obtained first gradient and each second gradient, and then adjusting the parameters of the first recommendation model based on the third gradient obtained after the superposition; or

[0192] The obtained second gradients are combined to obtain a fourth gradient, the fourth gradient is combined with the first gradient again to obtain a fifth gradient, and parameters of the first recommendation model are adjusted based on the fifth gradient.

[0193] Optionally, after outputting the target first recommendation model obtained through training, the training unit 402 is further configured to:

[0194] Obtaining human feature information to be predicted, and inputting the human feature information to be predicted into the target first recommendation model to obtain corresponding predicted behavior information, and recommending corresponding associated objects based on the predicted behavior information; or

[0195] Obtain the human feature information to be predicted and the corresponding user ID to be predicted, and through a third-party device, determine whether the user ID to be predicted appears within the jurisdiction of other designated edge computing devices, input the human feature information to be predicted into the target first recommendation model, obtain corresponding predicted behavior information, and recommend corresponding associated objects based on the predicted behavior information.

[0196] Based on the same inventive concept, see Figure 5 As shown, the disclosed embodiment provides an edge computing device, comprising at least:

[0197] Memory 501, used to store executable instructions;

[0198] The processor 502 is used to read and execute executable instructions stored in the memory to execute any one of the methods executed by the edge computing device in the above embodiments.

[0199] Based on the same inventive concept, see Figure 6 As shown, in an embodiment of the present disclosure, a network device (eg, a third-party device) is provided, comprising at least: a first communication unit 601, a second communication unit 602, wherein:

[0200] A first communication unit 601 is configured to receive an acquisition request from an edge computing device, the acquisition request being sent by the edge computing device during federated training of a first recommendation model based on sample user data of each user; the first recommendation model is configured to predict user behavior and recommend related objects; wherein a piece of sample user data includes a user's body feature information and actual behavior information;

[0201] The second communication unit 602 is configured to feed back information reported by each designated other edge computing device to the edge computing device, so as to enable the edge computing device to adjust parameters of the first recommendation model.

[0202] Optionally, before receiving the acquisition request from the edge computing device, the first communication unit 601 is further configured to:

[0203] Accepting a screening request from the edge computing device, wherein the screening request carries the user ID of each sample user data collected by the edge computing device, and the user ID is homomorphically encrypted;

[0204] Based on the reported information of each designated other edge computing device, the obtained user ID is filtered, and the filtered user ID is notified to the edge computing device, so that the edge computing device trains the first recommendation model based on the sample user data corresponding to the filtered user ID, and the filtered user ID has appeared in each designated other edge computing device.

[0205] Optionally, the information reported by each designated other edge computing device is transmitted to the edge computing device, so that the edge computing device adjusts parameters of the first recommendation model, and the second communication unit 602 is used to:

[0206] The second gradient reported by each designated other edge computing device is fed back to the edge computing device, so that the edge computing device adjusts the parameters of the first recommendation model based on the obtained first gradient and each second gradient, wherein the first gradient is obtained by the edge computing device through federated training of the first recommendation model based on the sample user data of each user, and the each second gradient is obtained by each designated other edge computing device through training the corresponding second recommendation model based on its own sample user data.

[0207] Optionally, the second gradient reported by each designated other edge computing device is fed back to the edge computing device, and the second communication unit 602 is used to:

[0208] Based on the specified user ID carried in the acquisition request, the second gradient reported by each of the specified other edge computing devices corresponding to the specified user ID is fed back to the edge computing device; wherein, the second gradient is obtained by each of the specified other edge computing devices, respectively based on the sample user data corresponding to the specified user ID within their respective jurisdictions, by training the corresponding second recommendation model.

[0209] Based on the same inventive concept, see Figure 7 As shown, the disclosed embodiment provides a network device (e.g., a third-party device, comprising at least:

[0210] Memory 701, used to store executable instructions;

[0211] The processor 702 is configured to read and execute executable instructions stored in the memory, and execute any one of the methods executed by the third-party device in the above-mentioned embodiments.

[0212] Based on the same inventive concept, an embodiment of the present disclosure provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor, the processor is enabled to execute any one of the methods executed by the edge computing device in the above embodiments.

[0213] Based on the same inventive concept, an embodiment of the present disclosure provides a computer-readable storage medium. When instructions in the computer-readable storage medium are executed by a processor, the processor is enabled to execute any one of the methods executed by the third-party device in the above embodiments.

[0214] To sum up, in the embodiment of the present disclosure, the edge computing device inputs the collected sample user data of each user into the first recommendation model for federal training within its jurisdiction, and outputs the target first recommendation model obtained through training when the preset training end conditions are met, wherein a piece of sample user data includes the user's human feature information and real behavior information, and the first recommendation model is used to predict the user's behavior and recommend related objects; in this way, the first recommendation model can be federally trained on the edge computing device without leaking user privacy and reducing network data transmission volume, thereby obtaining a more accurate and more widely applicable target first recommendation model, and then based on the target first recommendation model, more accurate behavior predictions can be made for new users, and more accurate recommendation plans can be formulated.

[0215] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0216] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0217] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0218] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0219] Although the preferred embodiments of the present disclosure have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present disclosure.

[0220] Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present disclosure without departing from the spirit and scope of the embodiments of the present disclosure. Thus, if such changes and modifications of the embodiments of the present disclosure fall within the scope of the claims of the present disclosure and their equivalents, the present disclosure is intended to include such changes and modifications.

Claims

1. A model training system based on edge computing, characterized in that: It includes several edge computing devices, each of which corresponds to different retail businesses, wherein: Any edge computing device that performs the following operations: Collect sample user data from each user within the jurisdiction, where one piece of sample user data includes the user's physical characteristics and real behavior information, and the real behavior information includes attribute information of the user's behavior-related objects obtained by analyzing orders, where the orders are collected through a point-of-sale (POS) information management system. From each sample user data, filter out the sample user data generated by users that appear within the jurisdiction of different edge computing devices; Based on the screened sample user data, a first recommendation model is federated trained, wherein the first recommendation model is used to predict user behavior and recommend related objects; When the preset training end condition is met, the target first recommendation model obtained by training is output; Also includes: Obtain the human feature information to be predicted and the corresponding user ID to be predicted, and through a third-party device, determine whether the user ID to be predicted appears within the jurisdiction of other designated edge computing devices, input the human feature information to be predicted into the target first recommendation model, obtain corresponding predicted behavior information, and recommend corresponding associated objects based on the predicted behavior information.

2. The system according to claim 1, wherein The human body feature information includes the user's appearance attribute information obtained through a smart camera.

3. The system according to claim 1 or 2, characterized in that From each sample user data, filter out sample user data generated by users that appear within the jurisdiction of different edge computing devices, including: The user identification ID of each sample user data is homomorphically encrypted and sent to a third-party device; Based on the feedback result of the third-party device, filter out the user IDs that have appeared on each other specified edge computing device from each user ID; Based on the filtered user ID, the corresponding sample user data is further filtered out.

4. The system according to claim 1, wherein: The step of performing federated training on the first recommendation model based on the screened sample user data includes: Each time sample user data is input into the first recommendation model, a corresponding first gradient is obtained based on an error between the output predicted behavior information and the actual behavior information; Use federated learning to obtain the second gradient sent by each other edge computing device; Parameters of the first recommendation model are adjusted based on the obtained first gradient and each second gradient.

5. The system according to claim 4, wherein: The method of adopting the federated learning method to obtain the second gradient sent by each other edge computing device includes: Send the specified user ID to the third-party device; Based on the feedback results of the third-party device, the second gradients reported by the designated other edge computing devices are obtained, wherein a second gradient is obtained by training the corresponding second recommendation model by a designated other edge computing device based on the sample user data corresponding to the designated user ID.

6. The system according to claim 4, wherein: The step of adjusting parameters of the first recommendation model based on the obtained first gradient and second gradient includes: Linearly superimposing the obtained first gradient and each second gradient, and then adjusting the parameters of the first recommendation model based on the third gradient obtained after the superposition; or The obtained second gradients are combined to obtain a fourth gradient, the fourth gradient is combined with the first gradient again to obtain a fifth gradient, and parameters of the first recommendation model are adjusted based on the fifth gradient.

7. A model training method based on edge computing, characterized in that: include: Receive an acquisition request from an edge computing device, the acquisition request being sent by the edge computing device during a process of federated training of a first recommendation model based on sample user data of each user, the first recommendation model being used to predict user behavior and recommend associated objects, wherein a piece of sample user data includes human feature information and real behavior information of the user, the human feature information including appearance attribute information of the user obtained by a smart camera, the real behavior information including attribute information of behavior-associated objects of the user obtained by analyzing orders, the orders being collected by a point-of-sale (POS) management system, and the sample user data of each user being sample user data generated by users who have appeared within the jurisdiction of different edge computing devices, screened by the edge computing device; Feedback information reported by each designated other edge computing device to the edge computing device, so that the edge computing device adjusts parameters of the first recommendation model; Among them, after the edge computing device completes training of the first recommendation model and obtains the target first recommendation model, it obtains the human feature information to be predicted and the corresponding user ID to be predicted, and determines through a third-party device that the user ID to be predicted has appeared within the jurisdiction of other designated edge computing devices, and inputs the human feature information to be predicted into the target first recommendation model to obtain corresponding predicted behavior information, and recommends corresponding associated objects based on the predicted behavior information.

8. The method according to claim 7, wherein Before receiving the acquisition request from the edge computing device, the method further includes: Accepting a screening request from the edge computing device, wherein the screening request carries the user ID of each sample user data collected by the edge computing device, and the user ID is homomorphically encrypted; Based on the reported information of each designated other edge computing device, the obtained user ID is filtered, and the filtered user ID is notified to the edge computing device, so that the edge computing device trains the first recommendation model based on the sample user data corresponding to the filtered user ID, and the filtered user ID has appeared in each designated other edge computing device.

9. The method according to claim 7, wherein Feeding back information reported by each designated other edge computing device to the edge computing device, so that the edge computing device adjusts parameters of the first recommendation model, includes: The second gradient reported by each designated other edge computing device is fed back to the edge computing device, so that the edge computing device adjusts the parameters of the first recommendation model based on the obtained first gradient and each second gradient, wherein the first gradient is obtained by the edge computing device through federated training of the first recommendation model based on the sample user data of each user, and the each second gradient is obtained by each designated other edge computing device through training the corresponding second recommendation model based on its own sample user data.

10. The method according to claim 9, wherein Feeding back the second gradient reported by each designated other edge computing device to the edge computing device includes: Based on the specified user ID carried in the acquisition request, the second gradient reported by each of the specified other edge computing devices corresponding to the specified user ID is fed back to the edge computing device; wherein, the second gradient is obtained by each of the specified other edge computing devices, respectively based on the sample user data corresponding to the specified user ID within their respective jurisdictions, by training the corresponding second recommendation model.

11. A model training device based on edge computing, characterized in that: Applied to a model training system based on edge computing, the model training system includes several edge computing devices corresponding to different retail businesses, and the apparatus is applied to any edge computing device, including: A collection unit is configured to collect sample user data from each user within its jurisdiction, wherein one piece of sample user data includes the user's physical characteristics and real behavior information, wherein the real behavior information includes attribute information of the user's behavior-related objects obtained by analyzing orders, wherein the orders are collected via a point-of-sale (POS) intelligence management system; and filter out sample user data generated by users who have appeared within the jurisdiction of different edge computing devices from each piece of sample user data; a training unit, configured to perform federated training on a first recommendation model based on the obtained sample user data, wherein the first recommendation model is used to predict user behavior and recommend related objects; An output unit, configured to output the target first recommendation model obtained through training when a preset training end condition is met; The training unit is also used to obtain the human feature information to be predicted and the corresponding user ID to be predicted, and through a third-party device, determine whether the user ID to be predicted appears within the jurisdiction of other designated edge computing devices, input the human feature information to be predicted into the target first recommendation model, obtain corresponding predicted behavior information, and recommend corresponding associated objects based on the predicted behavior information.

12. The device according to claim 11, wherein The human body feature information includes the user's appearance attribute information obtained through a smart camera.

13. The device according to claim 11 or 12, characterized in that The acquisition unit is further configured to: The user identification ID of each sample user data is homomorphically encrypted and sent to a third-party device; Based on the feedback result of the third-party device, filter out the user IDs that have appeared on each other specified edge computing device from each user ID; Based on the filtered user ID, the corresponding sample user data is further filtered out.

14. The device according to claim 11, wherein The first recommendation model is federated trained based on the obtained sample user data, and the training unit is used to: Each time sample user data is input into the first recommendation model, a corresponding first gradient is obtained based on an error between the output predicted behavior information and the actual behavior information; Use federated learning to obtain the second gradient sent by each other edge computing device; Parameters of the first recommendation model are adjusted based on the obtained first gradient and each second gradient.

15. A device for model training based on edge computing, characterized in that: include: A first communication unit is configured to receive an acquisition request from an edge computing device, the acquisition request being sent by the edge computing device during federated training of a first recommendation model based on sample user data of each user; the first recommendation model being configured to predict user behavior and recommend associated objects; wherein a piece of sample user data includes a user's human feature information and real behavior information, the human feature information including appearance attribute information of the user obtained through a smart camera, the real behavior information including attribute information of behavior-associated objects of the user obtained through order analysis, the orders being collected through a point-of-sale (POS) management system; the sample user data of each user being sample user data generated by users who have appeared within the jurisdiction of different edge computing devices, screened by the edge computing device; a second communication unit, configured to feed back information reported by each designated other edge computing device to the edge computing device, so as to enable the edge computing device to adjust parameters of the first recommendation model; Among them, after the edge computing device completes training of the first recommendation model and obtains the target first recommendation model, it obtains the human feature information to be predicted and the corresponding user ID to be predicted, and determines through a third-party device that the user ID to be predicted has appeared within the jurisdiction of other designated edge computing devices, and inputs the human feature information to be predicted into the target first recommendation model to obtain corresponding predicted behavior information, and recommends corresponding associated objects based on the predicted behavior information.

16. An edge computing device, characterized in that: include: a memory for storing executable instructions; A processor, configured to read and execute the executable instructions stored in the memory to implement the method according to any one of claims 7 to 10.

17. A network device, characterized in that: include: a memory for storing executable instructions; A processor, configured to read and execute the executable instructions stored in the memory to implement the method according to any one of claims 7 to 10.

18. A computer-readable storage medium, characterized in that When the instructions in the computer-readable storage medium are executed by a processor, the processor is enabled to perform the method according to any one of claims 7 to 10.

Images