Training method, device and storage medium for click-through rate prediction model of edge-cloud relay
By using a collaborative training mechanism between the cloud and the device, a lightweight model is trained using data prior to data regulation and updated in real time on the device. This solves the privacy and performance issues of click-through rate prediction models under data regulation, and achieves more accurate click-through rate prediction.
Patent Information
- Application Number
- CN202311618807.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-30
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-11-30
AI Technical Summary
Existing click-through rate prediction models face challenges in protecting user data privacy and updating models, especially under data regulation, where it is difficult to effectively utilize old and real-time data for model updates and optimization.
By using pre-data-monitoring data in the cloud to perform group meta-learning training on multiple lightweight and personalized device-side models, and then using real-time data after data-monitoring on the device side for collaborative training, combined with a model selector and collaborative learning mechanism, the system achieves automated model selection and personalized updates.
It achieves more accurate click-through rate prediction results while protecting user data privacy, and improves the model's predictive performance.
Smart Images

Figure CN117609788B_ABST
Abstract
Description
Technical Field
[0001] This disclosure belongs to the field of edge intelligent recommendation technology, and specifically relates to a method, device and storage medium for training a click-through rate prediction model that combines edge and cloud computing. Background Technology
[0002] Click-through rate (CTR) prediction aims to forecast the probability of a user clicking on a recommended item, playing a crucial role in recommendation systems, online advertising, and web search. To improve model accuracy, models have evolved from simple logistic regression (LR) and factorization machines (FM) to deep neural networks (DNN), Wide & Deep, and Deep Interest Networks (DIN). Furthermore, large amounts of user interaction data are collected by operators to train advanced CTR prediction models. Despite significant progress in CTR prediction, models often exhibit significant performance gaps between the training and inference phases. In practical applications, user click behavior typically changes over time due to variations in user needs or interests. Therefore, models deployed on end devices can easily become outdated, leading to decreased prediction performance. Thus, to maintain prediction performance, models need to be updated regularly with new data. However, CTR prediction involves privacy-sensitive data such as user profiles and behavioral information, including user ID, gender, age, and clicked items. With increasing user privacy awareness, users may refuse to share their private data. Moreover, many data protection laws and regulations both domestically and internationally impose strict regulations and restrictions on the collection and use of user data. Under the new form of data supervision, it is of great research significance to make full use of the old data collected before data supervision and the real-time data scattered on the terminal devices after data supervision to update the click-through rate prediction model. Summary of the Invention
[0003] This disclosure aims to at least partially address one of the technical problems in the related art.
[0004] To address this, this disclosure proposes a cloud-edge click-through rate (CTR) prediction model training method. This method involves first training multiple lightweight, highly capable, and personalized edge models in the cloud using pre-data monitoring data and models. Then, on each edge device, the edge models are collaboratively trained using real-time data after data monitoring. This approach protects user data privacy while enabling model updates, providing users with more accurate CTR prediction results.
[0005] Another objective of this disclosure is to propose a training device for a click-through rate prediction model for edge-cloud relay.
[0006] Another objective of this disclosure is to provide a computer storage medium.
[0007] To achieve the above objectives, the first aspect of this disclosure proposes a method for training a click-through rate prediction model for edge-cloud relay, comprising:
[0008] The cloud server uses old data collected before data supervision to perform group meta-learning training to obtain multiple initial device-side models;
[0009] The cloud server trains a model selector to automatically select the initial device-side model for the user from multiple initial device-side models based on user preferences;
[0010] The device downloads the initial device model and uses local data after data monitoring to relay the training of the device model. The devices achieve collaborative training through the sharing of device model parameters to obtain a personalized click-through rate prediction model.
[0011] In some embodiments, the grouped meta-learning training first trains a click-through rate prediction model on a cloud server using data collected in the cloud before data supervision, and then groups the cloud-collected data based on user embedding vectors used to characterize user features, and trains the click-through rate prediction model in groups to obtain multiple device-side models, which serve as the multiple initial device-side models.
[0012] In some embodiments, the cloud server uses old data collected before data supervision to perform group meta-learning training to obtain multiple initial device-side models, including:
[0013] S11, cloud servers utilize datasets Train a program with initial parameters w s Click-through rate prediction model f s (w s During training, parameter w s Update and loss function Set them to:
[0014]
[0015]
[0016] in, This represents a dataset constructed from data collected in the cloud prior to data regulation. Representatives from users before data regulation The dataset consists of the collected data. y represents the user set before data regulation. i A label indicating whether the user has interacted with item i. The click-through rate prediction model f represents the cloud server side. s (w sPredict the probability that a user will interact with item i, where η1 is the first learning rate. Represents a dataset;
[0017] S12, the cloud server uses the trained click-through rate prediction model f s (w s Get each user User Embedding Matrix According to the user embedding matrix Clustering users yields user grouping results. in, Representing the k-th group of users, Represents the number of groups;
[0018] S13, within each group k, train the device-side model using meta-learning to obtain an initial device-side model parameter for each group. Specifically, it includes:
[0019] S131, Randomly initialize device-side model parameters
[0020] S132, Repeat the following steps until the device-side model converges: 1) From the kth group of users A random sample of users 2) For each user in B First, perform a local update, which will randomly initialize the device-side model parameters. Assign a value to parameter θ u Then calculate the loss function. And update η2 represents the second learning rate. Representative task τ u The corresponding support set; 3) Perform a global update. η3 represents the third learning rate. Representative task τ u The corresponding query set;
[0021] S133, using the dataset Training parameters
[0022] In some embodiments, the cloud server trains a model selector to automatically select an initial device-side model for a user from multiple initial device-side models based on user preferences, including:
[0023] S21, using the click-through rate prediction model trained on the data previously collected in the cloud via the cloud server, the click-through rate prediction model for each item is obtained. Embedded vector e iBased on the embedding vectors of all items, the items are clustered into groups. Group correspond Each item category has several categories. Embedding vectors of items belonging to this category Taking the average yields a cluster center embedding vector. Represent the characteristics of each item category c;
[0024] S22, for users Each interactive item Representatives from users before data regulation The collected data forms a dataset, which is used to calculate interactive items. Embedded vector Embedded vector e with each cluster center c Similarity between users The category of each interactive item;
[0025] S23, based on user The categories of interactive items are used by users. preferences Modeling, the formula is as follows:
[0026]
[0027] In the formula, λ c On behalf of users The number of interacted items belonging to category c. This represents the total number of items that the user has interacted with;
[0028] S24, By modeling user preferences, construct the training dataset for the model selector according to the following formula.
[0029]
[0030] In the formula, The user obtained through step S21 The clustering results correspond to an initialization device-side model ID on the cloud server;
[0031] S25, using the training dataset The model selector is trained, and the training task of the model selector is regarded as a classification task, for the user. Newly generated interactive data on the device side Construct the input to the model selector, and initialize the device-side model ID as the output of the model selector. This represents the user set before data regulation.
[0032] In some embodiments, a multilayer perceptron network is used as the model selector s, and in the training dataset The model selector s is then trained using the cross-entropy loss function.
[0033] In some embodiments, the devices achieve collaborative training to obtain a personalized click-through rate prediction model through sharing of device-side model parameters, including:
[0034] S31, in each round of training, from the user set after data supervision Randomly select a portion of users Each participating user The initial device-side model is trained and updated several times using the local data generated after data supervision to obtain the updated device-side model parameters.
[0035] S32, each participating user Upload the updated device-side model parameters to the cloud server;
[0036] S33, for each group k, the cloud server will download the aggregated update parameter w of the model parameters of all participating users who initialized the device-side model for that group. k The updated formula is as follows:
[0037]
[0038] In the formula, k u This represents the clustering result of user u when the cloud server performs group meta-learning training. This represents the interactive data generated on the user's device after data supervision.
[0039] S34, using each set of parameters w k Update all User's device-side model;
[0040] S35. Repeat steps S31 to S34 several times until the model's on-device click-through rate prediction performance reaches the target level.
[0041] In some embodiments, in S31, the FedProx algorithm is used to perform several rounds of local training updates on the initial device-side model. The loss function used in the FedProx algorithm is... The update method for device-side model parameters is as follows:
[0042]
[0043]
[0044] In the formula, For users The interactive data generated by the device after data supervision For local training users The device-side model parameters, For users corresponding The device-side model parameters of the group are η4, which is the fourth learning rate, and μ is the hyperparameter.
[0045] In some embodiments, in S31, the FedAvg algorithm is used to perform several rounds of local training updates on the initial device-side model. The loss function used in the FedAvg algorithm is... The update method for device-side model parameters is as follows:
[0046]
[0047]
[0048] In the formula, For users The interactive data generated by the device after data supervision For local training users The device-side model parameters, For users corresponding The device-side model parameters of the group, where η5 is the fifth learning rate.
[0049] A second aspect of this disclosure provides a click-through rate (CTR) prediction model training apparatus for implementing the click-through rate prediction model training method according to any embodiment of the first aspect of this disclosure, comprising:
[0050] The initialization device-side model training module is used to perform group meta-learning training on the old data collected before data supervision through the cloud server to obtain multiple initialization device-side models;
[0051] The model selector training module is used to train a model selector via a cloud server to automatically select the initial device-side model for the user from multiple initial device-side models based on user preferences.
[0052] The device-side model collaborative training module is used to enable devices to download the initial device-side model and use local data after data supervision to relay the training of the device-side model. Devices achieve collaborative training through the sharing of device-side model parameters to obtain personalized click-through rate prediction models.
[0053] A computer-readable storage medium is provided in the third aspect of this disclosure, the computer-readable storage medium storing computer instructions for causing the computer to perform a click-through rate prediction model training method according to any embodiment of the first aspect of this disclosure.
[0054] This disclosure provides a click-through rate prediction model training method, apparatus, and storage medium for edge-cloud relay, which has the following characteristics and beneficial effects:
[0055] 1. By first training multiple lightweight, highly capable, and personalized edge models in the cloud using data and models from before data supervision, and then collaboratively training the edge models on each edge device using real-time data after data supervision, user data privacy is protected while model updates are achieved, providing users with more accurate click-through rate prediction results.
[0056] 2. It includes a group meta-learning mechanism that can fully utilize data collected in the cloud before data supervision to train a model with good initialization for end devices.
[0057] 3. Includes an automated model selection mechanism that can automatically select an initialization model for the end device based on user preferences.
[0058] 4. It includes a personalized collaborative learning mechanism that enables real-time updates of the user device-side model without sharing user data after data supervision. Attached Figure Description
[0059] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which:
[0060] Figure 1 This is a diagram illustrating the application scenarios for the click-through rate prediction model training method for edge-cloud relay provided in this embodiment of the disclosure.
[0061] Figure 2 This is a flowchart of the click-through rate prediction model training method for end-to-end cloud relay provided in this embodiment of the disclosure;
[0062] Figure 3 This is a logic diagram of the click-through rate prediction model training method for end-to-end cloud relay provided in this embodiment of the disclosure;
[0063] Figure 4 This is a performance verification diagram of the click-through rate prediction model training method for edge-cloud relay provided in this embodiment of the disclosure;
[0064] Figure 5 This is a schematic diagram of the structure of the click-through rate prediction model training device for edge-cloud relay provided in this embodiment of the disclosure;
[0065] Figure 6 This is a schematic diagram of the structure of an electronic device for implementing a computer-readable storage medium, provided in an embodiment of this disclosure. Detailed Implementation
[0066] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.
[0067] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.
[0068] The following describes, with reference to the accompanying drawings, a method, apparatus, and storage medium for training a click-through rate prediction model for edge-cloud relay based on embodiments of the present disclosure.
[0069] Figure 1 This diagram illustrates the application scenarios for the click-through rate prediction model training method for edge-cloud relay provided in this embodiment. For example... Figure 1 As shown, before data regulation, the cloud server collected a large amount of user interaction data and used this data to train a click-through rate (CTR) prediction model. After data regulation, user interaction data generated by end devices can no longer be collected in the cloud, but is instead distributed across the end devices. The end-cloud relay CTR prediction model training method provided in this disclosure proposes to fully utilize the user interaction data collected by the cloud before data regulation, the CTR prediction model trained using the collected user interaction data, and the distributed user interaction data distributed across the end devices after data regulation, to train and update the CTR prediction model on the end devices, thereby improving the predictive performance of the user model.
[0070] Figure 2 This is a flowchart of the click-through rate prediction model training method for edge-cloud relay provided in this embodiment of the disclosure. Figure 2 As shown, the method includes, but is not limited to, the following steps:
[0071] S1, the cloud server uses old data collected before data supervision to perform group meta-learning training to obtain multiple initial device-side models;
[0072] S2, the cloud server trains a model selector to automatically select the initial device-side model for the user from multiple initial device-side models based on user preferences;
[0073] S3: The device downloads the initial device model and uses local data after data supervision to relay the training of the device model. The devices achieve collaborative training through the sharing of device model parameters to obtain a personalized click-through rate prediction model.
[0074] As an example, Figure 3 A logic diagram of the click-through rate prediction model training method for edge-cloud relay provided in this embodiment of the disclosure is shown below. Figure 3 As shown:
[0075] Specifically, to fully utilize the user interaction data collected by the cloud server before data supervision, the click-through rate prediction model training method of the end-to-cloud relay in this disclosure provides a grouped meta-learning training method. First, a server model is trained on the cloud server using data collected in the cloud before data supervision. Then, the cloud data is grouped based on user embedding vectors used to represent user characteristics. The grouped server model is then trained to obtain multiple device-side models, which serve as the initial models for the device-side. Let... This represents a dataset constructed from data collected in the cloud prior to data regulation, in which Representatives from users before data regulation The dataset consists of the collected data. This represents the user set before data regulation. The grouping meta-learning training method of this disclosure specifically includes the following steps:
[0076] S11, the cloud server first utilizes the dataset Train a more advanced initialization parameter w s Click-through rate prediction model f s (w s During training, parameter w s Update and loss function Set them to:
[0077]
[0078]
[0079] Among them, y i A label indicating whether the user has interacted with item i. The click-through rate prediction model f represents the cloud server side. s (w s Predict the probability that a user will interact with item i, where η1 is the first learning rate. It refers to a specific dataset in general.
[0080] S12, then the cloud server uses the trained click-through rate prediction model f s (w s Get each user User Embedding Matrix Then based on the user embedding matrix Perform k-means clustering on the users to obtain the user grouping results. in, Representing the k-th group of users, This represents the number of groups.
[0081] S13, then within each group k, the device-side model f is trained using meta-learning to obtain an initial device-side model parameter for each group. Meta-learning is employed for each user Click-through rate prediction can be considered a task. Each task Divided into a support set and a query set These are used to calculate the training loss and test loss for each task, respectively. Specifically, the support set is used in the local update phase, where the algorithm adjusts the initialization parameters of the device-side model based on each support set. This facilitates the learning process. Conversely, the query set is used in the global update phase. The algorithm calculates the loss function value of the query set using the model parameters adjusted in the local update phase, and trains and updates the initial device-side model parameters by minimizing the loss function. This involves learning how to learn. Through the aforementioned meta-learning process, the initial device-side models obtained by the cloud server can efficiently adapt to new tasks. The specific meta-learning method is as follows:
[0082] S131, First, randomly initialize the device-side model parameters.
[0083] S132, then repeat the following steps until the device-side model converges: 1) From A random sample of users 2) For each user u∈B in B, first perform a local update, which will randomly initialize the device-side model parameters. Assign a value to parameter θ u Then calculate the loss function. And update η2 represents the second learning rate. Representative task τ u The corresponding support set; 3) Perform a global update. η3 represents the third learning rate. Representative task τ u The corresponding query set;
[0084] S133, finally utilizing the dataset collected in the cloud before data supervision. train
[0085] The click-through rate prediction model training method for edge-cloud relay in this embodiment of the present disclosure provides a group-based learning method that ultimately yields... Initialize device-side model parameters and user grouping results
[0086] Through the aforementioned grouping meta-learning training method, the cloud server trains multiple initial device-side models. Devices can then download one of these initial device-side models and train it locally to adapt to new user-generated data. To enable the selection of a suitable initial device-side model for the user, the end-to-cloud relay click-through rate prediction model training method of this disclosure provides an automated model selection mechanism. The cloud server trains a model selector based on the grouping results and item embedding vectors to automatically select an initial device-side model for the user. Specifically, it includes the following steps:
[0087] S21, by utilizing the pre-trained click-through rate prediction model f obtained in step S11 s (w s You can get each item. Embedded vector e i (This vector is the click-through rate prediction model f) s (w s The intermediate layer outputs features. Then, based on the embedding vectors of all items, the items are clustered to classify them into groups. Group correspond There are several categories. For each item category... Embedding vectors of items belonging to this category Taking the average yields a cluster center embedding vector. This represents the characteristics of each item category c.
[0088] S22, then, for the user Each interactive item By calculating its embedding vector Embedded vector e with each cluster center c Similarity between users The category of each interactive item.
[0089] S23, based on user The categories of interactive items are used by users. preferences Modeling, the formula is as follows:
[0090]
[0091] In the formula, λc On behalf of the user The number of interacted items belonging to category c. This represents the total number of items that the user has interacted with.
[0092] S24, By modeling user preferences, construct the training dataset for the model selector according to the following formula.
[0093]
[0094] In the formula, The user obtained through step S21 The clustering results correspond to an initialization device-side model ID on the cloud server.
[0095] S25, using the training dataset The model selector is trained, and this training task can be viewed as a classification task. Specifically, a multilayer perceptron (MLP) network is used as the model selector, and the training dataset... The model selector s is trained using the cross-entropy loss function. The model selector is trained on a cloud server and downloaded to each device for local inference. (The last sentence appears to be incomplete and possibly refers to a different topic.) On the device side, firstly based on the user Newly generated interactive data on the device side The input to the model selector is the user. preferences This aligns with the process of building the training dataset for the model selector. The model selector then outputs an ID, and the device downloads the corresponding initial device-side model from the cloud server.
[0096] After devices acquire their initial device-side models, they will relay-train them locally using newly generated data after data supervision to adapt them to the user's real-time preferences. To promote efficient collaborative training among users, the edge-cloud relay click-through rate prediction model training method of this disclosure provides a personalized collaborative training mechanism that performs personalized device-side model aggregation based on the user's initial device-side model selection. Each user first downloads the trained model selector s from the cloud server, and then utilizes locally generated interaction data... The input p to construct the model selector network u This is consistent with the process of constructing training data for the model selector. Then, the model selector outputs k. u For each group, the device will download the initial device model for that group from the cloud server. Next, the devices will begin an iterative collaborative learning process, which includes:
[0097] S31, in each round, from the user set after data supervision Randomly select a portion of users Each participating user The initial device-side model is trained and updated locally for τ rounds using the FedProx algorithm based on its local data.
[0098]
[0099]
[0100] In the formula, The loss function of the FedProx algorithm is... For users Newly generated interactive data on the device side For local training users The device-side model parameters, For users corresponding The device-side model parameters of the group are η4, which is the fourth learning rate, and μ is the hyperparameter.
[0101] Alternatively, the FedAvg algorithm can be used to perform several rounds of local training and updates on the initial device-side model. The loss function used in the FedAvg algorithm is... The update method for device-side model parameters is as follows:
[0102]
[0103]
[0104] In the formula, η5 is the fifth learning rate.
[0105] S32, then, each participating user Upload updated device model parameters To the cloud server.
[0106] S33, for each group k, the cloud server will download the aggregated update parameter w of the model parameters of all participating users who initialized the device-side model for that group. k The updated formula is as follows:
[0107]
[0108] In the formula, This represents the user set of the kth group that was selected to participate in the initialization of the model. u This represents the clustering result of user u when the cloud server performs group meta-learning training. This represents the interactive data generated on the user's device after data supervision.
[0109] S34, using each set of parameters w k Update all User's device-side model.
[0110] S35, repeat steps S31 to S34 several times until the on-device prediction performance of the model reaches the target level.
[0111] Verification of the effectiveness of the embodiments disclosed herein:
[0112] The effectiveness of the edge-cloud relay click-through rate (CTR) prediction model training method provided in this disclosure embodiment is verified using three public datasets (MovieLens, Frappé, and KKBox) and one industrial dataset (Meituan). For the three public datasets, the training data is randomly divided into two parts: 5% of the data represents user interaction data collected by the cloud server before data privacy regulations were implemented, and the remainder represents privacy-preserving data distributed across multiple devices. For the Meituan dataset, data collected from December 17, 2022 to December 20, 2022 is labeled as cloud data collected and aggregated before the implementation of data privacy regulations, while data generated from February 17, 2023 to February 20, 2023 is considered data generated after the implementation of data privacy regulations. This data is distributed and stored on user devices. The device-side CTR prediction model f employs two widely used logistic regression (LR) models or deep neural network (DNN) models, while the server-side CTR prediction model f is more advanced. s The xDeepFM model is adopted. The click-through rate prediction model training method for edge-cloud relay provided in this disclosure is named RelayRec. Its performance is compared with eight methods, including a combination of two initialization model training methods and four collaborative training methods. The two initialization model training methods include training the initial device-side model directly using old data collected in the cloud before data supervision, and training the initial meta-model using meta-learning methods. The four collaborative training methods include edge fine-tuning (training only using edge data), federated averaging (FedAvg), FedProx, and group collaboration. Two performance evaluation metrics are included: AUC (Area Under the Curve), which quantifies the probability that a randomly selected positive sample ranks higher than a randomly selected negative sample; a higher AUC value indicates better click-through rate prediction performance. The other is LogLoss (Logarithmic Loss), also known as binary cross-entropy loss, used as a metric to measure loss in binary classification scenarios; a lower Logloss value indicates better click-through rate prediction performance.
[0113] Figure 4The performance comparison between RelayRec and other methods shows that, in most cases, RelayRec outperforms the comparison schemes. In particular, it demonstrates the most significant performance improvement on the Frappé dataset, with an average AUC increase of 7.2% and a LogLoss reduction of 14.9%. Furthermore, RelayRec also performs significantly on the Meituan Industrial Dataset, with an AUC increase of approximately 5.9% and a LogLoss reduction of approximately 5.4%. Overall, RelayRec improves AUC by an average of 5% and reduces LogLoss by 8.1%. These performance improvements demonstrate the superiority of RelayRec's edge-cloud relay training method in terms of click-through rate prediction model training performance.
[0114] It is understood that the click-through rate (CTR) prediction model training method provided in this disclosure, which trains multiple lightweight, highly capable, and personalized end-device models in the cloud using data and models collected before data supervision, and then collaboratively trains the end-device models on each device using real-time data collected after data supervision, protects user data privacy while enabling model updates and providing users with more accurate CTR prediction results. Specifically, the proposed group meta-learning mechanism fully utilizes data collected in the cloud before data supervision to train well-initialized models for end devices; the proposed automated model selection mechanism automatically selects initialization models for end devices based on user preferences; and the proposed personalized collaborative learning mechanism enables real-time updates of user device end-device models after data supervision without sharing user data.
[0115] To achieve the above embodiments, such as Figure 5 As shown, this embodiment also provides a click-through rate prediction model training device 100 for end-to-end cloud relay. The device 100 includes an initialization device-side model training module 1, a model selector training module 2, and a device-side model collaborative training module 3.
[0116] Initialize device-side model training module 1, which is used to perform group meta-learning training on the old data collected before data supervision through the cloud server to obtain multiple initial device-side models;
[0117] Model selector training module 2 is used to train a model selector via a cloud server to automatically select an initial device-side model for the user from multiple initial device-side models based on user preferences;
[0118] The device-side model collaborative training module 3 is used to enable the device to download the initial device-side model and use the local data after data supervision to relay the training of the device-side model. The devices achieve collaborative training through the sharing of device-side model parameters to obtain a personalized click-through rate prediction model.
[0119] It should be noted that the aforementioned explanation of an embodiment of a click-through rate prediction model training method for end-to-end cloud relay also applies to an end-to-end cloud relay click-through rate prediction model training device in this embodiment, and will not be repeated here.
[0120] To implement the above embodiments, this disclosure also proposes a computer-readable storage medium storing a computer program that is executed by a processor to perform the click-through rate prediction model training method for end-to-cloud relay described above.
[0121] The following is for reference. Figure 6 The diagram illustrates a structural schematic of an electronic device suitable for implementing embodiments of the present disclosure. It should be noted that the electronic devices in the embodiments of the present disclosure may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs, desktop computers, and servers. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.
[0122] like Figure 6 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 101, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 102 or a program loaded from a storage device 108 into a random access memory (RAM) 103. The RAM 103 also stores various programs and data required for the operation of the electronic device. The processing unit 101, ROM 102, and RAM 103 are interconnected via a bus 104. An input / output (I / O) interface 105 is also connected to the bus 104.
[0123] Typically, the following devices can be connected to I / O interface 105: input devices 106 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, etc.; output devices 107 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 108 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.
[0124] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, this embodiment includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network via communication device 109, or installed from storage device 108, or installed from ROM 102. When the computer program is executed by processing device 101, it performs the functions defined above in the methods of embodiments of this disclosure.
[0125] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.
[0126] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.
[0127] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the aforementioned click-through rate prediction model training method for end-to-end cloud relay.
[0128] Computer program code for performing the operations of this disclosure can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and Python, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0129] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0130] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0131] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.
[0132] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0133] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0134] Those skilled in the art will understand that implementing all or part of the steps of the methods in the above embodiments can be accomplished by instructing related hardware through a program. The developed program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0135] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0136] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.
Claims
1. A method for training a click-through rate prediction model for edge-cloud relay, characterized in that, include: The cloud server uses old data collected before data supervision to perform group meta-learning training to obtain multiple initial device-side models; The cloud server trains a model selector to automatically select the initial device-side model for the user from multiple initial device-side models based on user preferences; The device downloads the initial device model and uses local data after data supervision to relay the training of the device model. The devices achieve collaborative training through the sharing of device model parameters to obtain a personalized click-through rate prediction model. The devices achieve collaborative training to obtain a personalized click-through rate prediction model through sharing of model parameters on the device side, including: S31, in each round of training, from the user set after data supervision Randomly select a portion of users Each participating user The initial device-side model is trained and updated several times using the local data generated after data supervision to obtain the updated device-side model parameters. S32, each participating user Upload the updated device-side model parameters to the cloud server; S33, for each group k, the cloud server will download the aggregated and updated parameters of all participating users' model parameters for the initialization of the device-side model in that group. The updated formula is as follows: In the formula, = , When performing group learning training on behalf of the cloud server, the user The clustering results Representing users after data regulation Interactive data generated on the device side; S34, using each set of parameters Update all User's device-side model; S35, repeat steps S31 to S34 several times until the model's on-device click-through rate prediction performance reaches the target level.
2. The click-through rate prediction model training method according to claim 1, characterized in that, The grouped meta-learning training first trains a click-through rate prediction model on a cloud server using data collected in the cloud before data supervision. Then, the data collected in the cloud is grouped based on user embedding vectors used to represent user characteristics. The click-through rate prediction model is trained in groups to obtain multiple device-side models, which serve as the multiple initial device-side models.
3. The click-through rate prediction model training method according to claim 1, characterized in that, The cloud server uses old data collected before data supervision to perform group meta-learning training to obtain multiple initialized device-side models, including: S11, cloud servers utilize datasets Train an initialization parameter of Click-through rate prediction model During training, parameters Update and loss function Set them to: in, This represents a dataset constructed from data collected in the cloud prior to data regulation. Representatives from users before data regulation The dataset consists of the collected data. Represents the user set before data regulation. Represents whether the user is associated with the item Interactive tags Click-through rate prediction model representing the cloud server side Predicting users and items Is there a probability of interaction? As the first learning rate, Represents a dataset; S12, the cloud server uses the trained click-through rate prediction model Get each user User Embedding Matrix : According to the user embedding matrix Clustering users yields user grouping results. ,in, Representing the k-th group of users, Represents the number of groups; S13, within each group k, train the device-side model using meta-learning to obtain an initial device-side model parameter for each group. Specifically, it includes: S131, Randomly initialize device-side model parameters ; S132, Repeat the following steps until the device-side model converges: 1) From the kth group of users A random sample of users ;2) For Every user in First, perform a local update, which will randomly initialize the device-side model parameters. Assigning a value to a parameter Then calculate the loss function. And update , Represents the second learning rate. Representative task The corresponding support set; 3) Perform a global update , Represents the third learning rate. Representative task The corresponding query set; S133, using the dataset Training parameters .
4. The click-through rate prediction model training method according to claim 2, characterized in that, The cloud server trains a model selector to automatically select an initial device-side model for the user from multiple initial device-side models based on user preferences, including: S21, using the click-through rate prediction model trained on the data previously collected in the cloud via the cloud server, the click-through rate prediction model for each item is obtained. Embedded vector Based on the embedding vectors of all items, the items are clustered into groups. Group ,correspond Each item category has several categories. Embedding vectors of items belonging to this category Taking the average yields a cluster center embedding vector. , representing each item category Features; S22, for users Each interactive item , Representatives from users before data regulation The collected data forms a dataset, which is used to calculate interactive items. Embedded vector Embedded vectors with each cluster center Similarity between users The category of each interactive item; S23, based on user The categories of interactive items are used by users. preferences Modeling, the formula is as follows: In the formula, On behalf of users The number of interacted items belonging to category c. This represents the total number of items that the user has interacted with; S24, By modeling user preferences, construct the training dataset for the model selector according to the following formula. : In the formula, The user obtained through step S21 The clustering results correspond to an initialization device-side model ID on the cloud server; S25, using the training dataset The model selector is trained, and the training task of the model selector is regarded as a classification task, for the user. Newly generated interactive data on the device side Construct the input to the model selector, and initialize the device-side model ID as the output of the model selector. This represents the user set before data regulation.
5. The click-through rate prediction model training method according to claim 4, characterized in that, A multilayer perceptron network is used as the model selector. and in the training dataset The model selector is applied using the cross-entropy loss function. Conduct training.
6. The click-through rate prediction model training method according to claim 1, characterized in that, In S31, the FedProx algorithm is used to perform several rounds of local training and updates on the initial device-side model. The loss function used in the FedProx algorithm is... The update method for device-side model parameters is as follows: In the formula, For users The interactive data generated by the device after data supervision For local training users The device-side model parameters, For users corresponding The group's device-side model parameters, The fourth learning rate, This is a hyperparameter.
7. The click-through rate prediction model training method according to claim 1, characterized in that, In S31, the FedAvg algorithm is used to perform several rounds of local training and updates on the initial device-side model. The loss function used in the FedAvg algorithm is... The update method for device-side model parameters is as follows: In the formula, For users The interactive data generated by the device after data supervision For local training users The device-side model parameters, This is the fifth learning rate.
8. A click-through rate prediction model training apparatus for implementing the click-through rate prediction model training method according to any one of claims 1 to 7, characterized in that, include: The initialization device-side model training module is used to perform group meta-learning training on the cloud server using old data collected before data supervision to obtain multiple initialization device-side models; The model selector training module is used to train a model selector via a cloud server to automatically select the initial device-side model for the user from multiple initial device-side models based on user preferences. The device-side model collaborative training module is used to enable devices to download the initial device-side model and use local data after data supervision to relay the training of the device-side model. Devices achieve collaborative training through the sharing of device-side model parameters to obtain personalized click-through rate prediction models.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the click-through rate prediction model training method according to any one of claims 1 to 7.
