Click rate estimation method, device, equipment and storage medium

Abstract

The application discloses a click rate prediction method and device, equipment and a storage medium, and belongs to the technical field of the Internet. The method comprises the following steps: obtaining to-be-detected data; performing feature extraction according to the to-be-detected data to obtain to-be-detected features; judging whether the to-be-detected features have a time window difference problem; if the to-be-detected features have the time window difference problem, performing multimedia information click rate prediction through a target click rate prediction model and the to-be-detected data to obtain a click rate prediction result. According to the scheme, if the to-be-detected features have the time window difference problem, it is determined that the model needs to be started, and the multimedia information click rate prediction is performed through the target click rate prediction model, so that the click rate prediction service can be optimized, and the multimedia information click rate prediction effect is improved.

Description

Technical Field

[0001] This invention relates to the field of Internet technology, and in particular to a method, apparatus, device, and storage medium for predicting click-through rates. Background Technology

[0002] With the rapid development of internet technology, numerous internet applications have emerged. Computational advertising, as a core product of internet applications, has driven innovation in business models and the monetization of commercial traffic. The overall process of computational advertising is divided into four important modules: recall, coarse ranking, fine ranking, and re-ranking. Fine ranking is the most critical module in the entire advertising business process, and ad click-through rate (CTR) prediction is the core technology within the fine ranking module. In other words, ad CTR prediction plays a crucial role in the monetization level of computational advertising traffic and the revenue of internet application products.

[0003] Due to the technical characteristics of click-through rate (CTR) prediction models, their high dependence on data and incremental updates over time windows have become bottlenecks hindering their efficient and rapid iteration, and a long-term business requirement for advertising algorithm engineers. Existing CTR prediction methods are prone to inaccuracies and generally ineffective.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a click-through rate (CTR) prediction method, apparatus, device, and storage medium, aiming to solve the technical problem of how to optimize CTR prediction services and improve the CTR prediction effect for multimedia information.

[0006] To achieve the above objectives, the present invention provides a click-through rate (CTR) prediction method, the CTR prediction method comprising:

[0007] Acquire the data to be tested;

[0008] Feature extraction is performed based on the data to be detected to obtain the features to be detected;

[0009] Determine whether the feature to be detected has a time window difference problem;

[0010] If the feature to be detected has a time window difference problem, then the click-through rate of multimedia information is predicted by using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

[0011] Optionally, the target click-through rate prediction model is obtained by transfer learning from a long-time window model and a short-time window model, wherein the accuracy of the long-time window model is higher than that of the short-time window model.

[0012] Optionally, before obtaining the click-through rate prediction result by predicting the click-through rate of multimedia information using the target click-through rate prediction model and the data to be detected, the method further includes:

[0013] Acquire data from a first time window and a second time window, wherein the data volume of the first time window is greater than that of the second time window.

[0014] The initial model is trained using the first time window data and the second time window data respectively to obtain a long time window model and a short time window model;

[0015] Based on the long-term window model and the short-term window model, a transfer learning method is used to obtain a target click-through rate prediction model.

[0016] Optionally, the step of performing transfer learning based on the long-time window model and the short-time window model to obtain the target click-through rate prediction model includes:

[0017] The sparse model corresponding to the long window model is selected as the candidate sparse model.

[0018] The dense model corresponding to the short time window model is selected as the candidate dense model;

[0019] A target click-through rate prediction model is generated based on the candidate sparse model and the candidate dense model.

[0020] Optionally, generating the target click-through rate prediction model based on the candidate sparse model and the candidate dense model includes:

[0021] The candidate sparse model is used as the hot start model for the short time window model;

[0022] The target click-through rate prediction model is obtained by jointly training the candidate sparse model and the candidate dense model after a warm start using the short time window model.

[0023] Optionally, the step of jointly training the short-time window model based on the candidate sparse model and the candidate dense model after a warm start to obtain the target click-through rate prediction model includes:

[0024] The candidate sparse model is replaced in the short time window model by the candidate sparse model after the warm start, so as to perform a warm start on the candidate dense model.

[0025] The short-time-window model is updated by jointly training the candidate sparse model and the candidate dense model.

[0026] The updated short time window model is used as the target click-through rate prediction model.

[0027] Optionally, if the feature to be detected has a time window difference problem, then the click-through rate (CTR) of multimedia information is predicted using a target CTR prediction model and the data to be detected, and the CTR prediction result is obtained, including:

[0028] If the features to be detected have time window differences, then determine whether there are differences in the sorting of multimedia information;

[0029] If there is no difference in the ranking of multimedia information, the click-through rate of multimedia information is predicted using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

[0030] Optionally, determining whether there is a difference in the sorting of multimedia information includes:

[0031] The test data is input into the long window model and the target click-through rate prediction model respectively to obtain the first click-through rate prediction result and the second click-through rate prediction result.

[0032] The first multimedia information ranking result is determined based on the first click-through rate prediction result;

[0033] The second multimedia information ranking result is determined based on the second click-through rate prediction result;

[0034] Compare the first multimedia information sorting result with the second multimedia information sorting result;

[0035] Determine whether there are differences in the sorting of multimedia information based on the comparison results.

[0036] Optionally, after determining whether there is a difference in the sorting of multimedia information, the method further includes:

[0037] If there are differences in the sorting of multimedia information, the click-through rate of the multimedia information is predicted based on the optimized target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

[0038] Optionally, before obtaining the click-through rate prediction result by predicting the click-through rate of multimedia information based on the optimized target click-through rate prediction model and the data to be detected, the method further includes:

[0039] The target click-through rate prediction model is optimized by knowledge distillation based on the long window model.

[0040] Optionally, the step of performing knowledge distillation on the target click-through rate prediction model based on the long-term window model to obtain an optimized target click-through rate prediction model includes:

[0041] The soft tag loss function is determined based on the first click-through rate prediction result corresponding to the long window model and the second click-through rate prediction result corresponding to the target click-through rate prediction model.

[0042] Determine the hard tag loss function based on the second click-through rate prediction result;

[0043] The overall loss function is determined based on the hard label loss function and the soft label loss function;

[0044] Based on the overall loss function, knowledge distillation is performed on the target click-through rate prediction model to obtain an optimized target click-through rate prediction model.

[0045] Optionally, determining the soft tag loss function based on the first click-through rate (CTR) prediction result corresponding to the long-term window model and the second CTR prediction result corresponding to the target CTR prediction model includes:

[0046] The first click-through rate (CTR) estimate is determined based on the first CTR estimate result corresponding to the long window model.

[0047] The second click-through rate (CTR) estimate is determined based on the second CTR estimate result corresponding to the target CTR estimate model.

[0048] Calculate the mean squared error based on the first and second click-through rate estimates;

[0049] The soft label loss function is determined based on the mean square error.

[0050] Optionally, determining the overall loss function based on the hard-label loss function and the soft-label loss function includes:

[0051] Obtain the first weight value corresponding to the hard label loss function, and obtain the second weight value corresponding to the soft label loss function;

[0052] The weighted calculation is performed based on the hard label loss function, the soft label loss function, the first weight value, and the second weight value.

[0053] The overall loss function is determined based on the weighted calculation results.

[0054] Furthermore, to achieve the above objectives, the present invention also proposes a click-through rate (CTR) prediction device, the CTR prediction device comprising:

[0055] The data acquisition module is used to acquire the data to be detected.

[0056] The feature extraction module is used to extract features from the data to be detected to obtain the features to be detected.

[0057] The feature detection module is used to determine whether the feature to be detected has a time window difference problem;

[0058] The click-through rate (CTR) prediction module is used to predict the CTR of multimedia information by using the target CTR prediction model and the data to be detected if there is a time window difference problem in the feature to be detected, and to obtain the CTR prediction result.

[0059] Optionally, the target click-through rate prediction model is obtained by transfer learning from a long-time window model and a short-time window model, wherein the accuracy of the long-time window model is higher than that of the short-time window model.

[0060] Optionally, the click-through rate prediction device further includes: a model training module and a transfer learning module;

[0061] The data acquisition module is also used to acquire first time window data and second time window data, wherein the amount of data in the first time window data is greater than that in the second time window data.

[0062] The model training module is used to train the initial model based on the first time window data and the second time window data respectively, to obtain a long time window model and a short time window model.

[0063] The transfer learning module is used to perform transfer learning based on the long-time window model and the short-time window model to obtain a target click-through rate prediction model.

[0064] Optionally, the transfer learning module is further configured to use the sparse model corresponding to the long time window model as a candidate sparse model; use the dense model corresponding to the short time window model as a candidate dense model; and generate a target click-through rate prediction model based on the candidate sparse model and the candidate dense model.

[0065] Optionally, the transfer learning module is further configured to use the candidate sparse model as the hot-start model of the short-time window model; and to obtain the target click-through rate prediction model by jointly training the short-time window model based on the candidate sparse model and the candidate dense model after the hot start.

[0066] In addition, to achieve the above objectives, the present invention also proposes a click-through rate (CTR) prediction device, which includes: a memory, a processor, and a CTR prediction program stored in the memory and executable on the processor. When the CTR prediction program is executed by the processor, it implements the CTR prediction method as described above.

[0067] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a click-through rate (CTR) prediction program, which, when executed by a processor, implements the CTR prediction method as described above.

[0068] The click-through rate (CTR) prediction method proposed in this invention involves: acquiring data to be detected; extracting features from the data to be detected; determining whether the features to be detected have a time window difference problem; if the features to be detected have a time window difference problem, then using a target CTR prediction model and the data to be detected to predict the CTR of multimedia information, thereby obtaining the CTR prediction result. This solution can determine the need for a model hot-start when the features to be detected have a time window difference problem, and then use the target CTR prediction model to predict the CTR of multimedia information, thereby optimizing the CTR prediction business and improving the CTR prediction effect for multimedia information. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the click-through rate prediction device structure in the hardware operating environment involved in the embodiments of the present invention;

[0070] Figure 2 This is a flowchart illustrating the first embodiment of the click-through rate prediction method of the present invention;

[0071] Figure 3 This is a flowchart illustrating the second embodiment of the click-through rate prediction method of the present invention;

[0072] Figure 4 This is a schematic diagram of the WS2ND model structure of an embodiment of the click-through rate prediction method of the present invention;

[0073] Figure 5 This is a flowchart illustrating the third embodiment of the click-through rate prediction method of the present invention;

[0074] Figure 6 This is a flowchart of a hot-start combined with model knowledge distillation technique based on transfer learning, according to an embodiment of the click-through rate prediction method of the present invention.

[0075] Figure 7 This is a schematic diagram of the knowledge distillation structure based on the WS2ND model structure, representing an embodiment of the click-through rate prediction method of the present invention.

[0076] Figure 8 This is a schematic diagram of the functional modules of the first embodiment of the click-through rate prediction device of the present invention.

[0077] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0078] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0079] Reference Figure 1 , Figure 1 This is a schematic diagram of the click-through rate prediction device structure for the hardware operating environment involved in the embodiments of the present invention.

[0080] like Figure 1 As shown, the click-through rate prediction device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and input units such as buttons; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed random access memory (RAM) or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0081] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the click-through rate prediction device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0082] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a click rate prediction program.

[0083] exist Figure 1 In the click-through rate (CTR) prediction device shown, the network interface 1004 is mainly used to connect to the external network and communicate with other network devices; the user interface 1003 is mainly used to connect to the user device and communicate with the user device; the device of the present invention calls the CTR prediction program stored in the memory 1005 through the processor 1001 and executes the CTR prediction method provided in the embodiment of the present invention.

[0084] Based on the above hardware structure, an embodiment of the click-through rate prediction method of the present invention is proposed.

[0085] Reference Figure 2 , Figure 2This is a flowchart illustrating the first embodiment of the click-through rate prediction method of the present invention.

[0086] In the first embodiment, the click-through rate prediction method includes:

[0087] Step S10: Obtain the data to be detected.

[0088] It should be noted that the execution subject of this embodiment can be a click-through rate prediction device, which can be a computer device with data processing function, or other devices that can achieve the same or similar functions. This embodiment does not limit this; in this embodiment, a computer device is used as an example for explanation.

[0089] It should be noted that the multimedia information in this embodiment may include, but is not limited to, advertisements, videos, images, and audio, as well as other types of multimedia information. This embodiment does not impose any limitations on this; in this embodiment, advertisements are used as an example of multimedia information for explanation. Accordingly, the multimedia information click-through rate prediction in this embodiment may specifically be an advertisement click-through rate prediction. The aforementioned advertisements may include, but are not limited to, various types of advertisements such as image advertisements, text advertisements, keyword advertisements, ranking advertisements, and video advertisements. This embodiment does not impose any limitations on this.

[0090] It should be understood that in recent years, driven by the rapid development of machine learning and deep learning, ad click-through rate (CTR) prediction technology has been continuously iterating and updating. Deep learning models, with their superior generalization performance, have gained recognition from major internet companies, making deep learning-based ad CTR prediction a widely accepted application. Ad CTR prediction models are based on daily-level feature data and are trained incrementally. Therefore, CTR prediction models trained with different time windows, especially those with longer time windows, have better prediction capabilities and stronger generalization abilities than those with shorter time windows. For the iteration and updating of existing models, it is necessary not only to optimize and update existing models of similar performance but also to further compensate for the model gains brought by the time window. Therefore, the high dependence of CTR prediction models on data features and the incremental updates accompanying time windows have become bottlenecks restricting their efficient and rapid iteration, and a long-term business requirement for advertising algorithm engineers.

[0091] To address the issue of compensating for the time window gains, some internet companies have proposed a "timed restart" solution. This involves retraining a model based on a manually set time window and directly replacing the online incremental model. This ensures that the data time window upon which the online incremental model relies is controllable, and that the gains of the new model during model iteration do not need to compensate for the time window gains of the original model. However, existing "timed restart" solutions are "lossy" in model prediction. Longer time window models see more data samples than shorter time window models. Therefore, for head features with high feature frequency, feature weights and representations learn to fit better due to gradient updates from multiple data samples. For long-tail features with low feature frequency, models that have seen more data are more likely to avoid the "cold start" problem. Thus, the gains are "lossy" for the model. The main work of this paper is to propose a better solution to address the time window model gain problem, and this solution has been tested and applied in real-world business scenarios.

[0092] It should be noted that this solution offers a superior approach to addressing the revenue issues associated with time-window models, and has been tested and applied in real-world business scenarios. This solution optimizes ad click-through rate prediction for search advertising scenarios, primarily to compensate for the additional model gains resulting from longer window times. To achieve this technical approach, this solution mainly utilizes transfer learning techniques.

[0093] It should be noted that the aforementioned transfer learning technique specifically refers to transfer learning warm-up techniques. Transfer learning warm-up, or transfer learning warm-up, involves using existing knowledge (source domain) and new knowledge (target domain) to transfer knowledge from the source domain to the target domain. In particular, in deep learning, transfer learning studies how to apply existing models to new, different, but related domains. Traditional machine learning is less flexible and produces less satisfactory results when dealing with tasks involving variations in data distribution, feature dimensionality, and model output. Transfer learning relaxes these assumptions. It organically utilizes knowledge from the source domain to better model the target domain, adapting to changes in data distribution, feature dimensionality, and model output. Furthermore, when labeled data is scarce, transfer learning can effectively utilize labeled data from related domains to perform data labeling. This approach employs the concept of transfer learning, using the click-through rate prediction of the long-term window model as the source domain and the click-through rate prediction of the short-term window model as the target domain. By warm-starting the short-term window model with the sparse model of the long-term window model, the approach further fits the data that the short-term window model has not seen.

[0094] It should be understood that when a multimedia information click-through rate prediction instruction is received, the data to be detected can be obtained according to the multimedia information click-through rate prediction instruction. Specifically, the data to be detected can be obtained from the advertising log system based on the instruction, which contains advertising data of previously recommended advertisements. Alternatively, the data to be detected can be obtained through other means based on the instruction. This embodiment does not limit this.

[0095] Step S20: Extract features based on the data to be detected to obtain the features to be detected.

[0096] It should be understood that after obtaining the data to be detected, in order to determine whether a model warm-start is needed, feature extraction can be performed on the data to obtain the features to be detected, and then the model warm-start can be used to determine whether a model warm-start is needed. If a model warm-start is needed, the target click-through rate (CTR) prediction model in this solution is used to predict the ad CTR. If a model warm-start is not needed, a conventional CTR prediction model is used. In this embodiment, the conventional CTR prediction model can include various models capable of CTR prediction, and this embodiment does not limit the specific model form of the conventional CTR prediction model.

[0097] It is understandable that in the scenario of predicting ad click-through rate, some ad features related to the ad can be extracted from the data to be detected. For example, ad features may include, but are not limited to, ad duration, ad type and ad playback time. This embodiment does not limit this.

[0098] Step S30: Determine whether there is a time window difference problem in the feature to be detected.

[0099] It should be understood that the presence of time window differences in the data to be detected can be determined based on the characteristics to be detected. Specifically, if the time span between the data to be detected is large, it can be determined that there is a time window difference problem; if the time span between the data to be detected is small, it can be determined that there is no time window difference problem.

[0100] Step S40: If the feature to be detected has a time window difference problem, then the click-through rate of multimedia information is predicted by the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

[0101] It should be understood that if a time window difference is detected in the feature to be detected, it can be determined that a model warm-start is required. Therefore, in this case, the click-through rate of multimedia information can be predicted by using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result can be obtained.

[0102] It should be noted that the target click-through rate (CTR) prediction model in this solution can be obtained through transfer learning from the long-term window model and the short-term window model. Both the long-term window model and the short-term window model are CTR prediction models. However, since the long-term window model has more training data than the short-term window model, the accuracy of the long-term window model is higher than that of the short-term window model.

[0103] Understandably, this solution employs a target click-through rate (CTR) prediction model obtained through transfer learning from both long-window and short-window models to predict CTR for multimedia information when a warm-start model is required. This ensures that the benefits of the long-window model are not lost. Simultaneously, for algorithm engineers who offline mine and optimize the CTR prediction model, there's no need to expend effort compensating for the benefits gained from the short-window model; they can focus on business improvement, further enhancing model iteration efficiency and traffic monetization capabilities. This has profound and significant implications for optimizing CTR prediction-related business processes.

[0104] In this embodiment, data to be detected is acquired; features are extracted from the data to be detected to obtain the features to be detected; it is determined whether the features to be detected have a time window difference problem; if the features to be detected have a time window difference problem, the click-through rate (CTR) of multimedia information is predicted using a target CTR prediction model and the data to be detected to obtain the CTR prediction result. This solution can determine the need for a model hot start when the features to be detected have a time window difference problem, and then perform CTR prediction of multimedia information using a target CTR prediction model, thereby optimizing the CTR prediction business and improving the CTR prediction effect of multimedia information.

[0105] In one embodiment, such as Figure 3 As shown, based on the first embodiment, a second embodiment of the click-through rate prediction method of the present invention is proposed. Before step S10, the method further includes:

[0106] Step S01: Obtain first time window data and second time window data, wherein the amount of data in the first time window data is greater than that in the second time window data.

[0107] It should be understood that in order to train long-time window models and short-time window models, first-time window data and second-time window data can be obtained from the advertising log system. The time window of the first-time window data is longer than that of the second-time window data, so the amount of data in the first-time window data is greater than that in the second-time window data.

[0108] Step S02: Train the initial model based on the first time window data and the second time window data respectively to obtain a long time window model and a short time window model.

[0109] It should be understood that after obtaining the first time window data and the second time window data, the initial model can be trained based on the first time window data to obtain a long time window model, and the initial model can be trained based on the second time window data to obtain a short time window model. Since the amount of data in the first time window data is greater than that in the second time window data, the accuracy of the trained long time window model is higher than that of the short time window model.

[0110] It should be noted that the initial model in this embodiment can be a neural network model, including but not limited to deep neural network models, convolutional neural network models, and other types of neural network models. This embodiment does not limit this.

[0111] Step S03: Perform transfer learning based on the long-time window model and the short-time window model to obtain the target click-through rate prediction model.

[0112] It should be understood that, in order to achieve a hot start for the model, transfer learning can be performed based on the long-time window model and the short-time window model to obtain the target click-through rate prediction model.

[0113] It should be noted that the ad click-through rate prediction model uses an embedding + MLP structure, also known as a sparse + dense structure. The sparse part is the vectorized representation of the feature values, and the dense part is a DNN network. In this embodiment, the sparse part of the model is referred to as the sparse model, and the dense part is referred to as the dense model.

[0114] Understandably, it can be referenced. Figure 4 , Figure 4 This is a schematic diagram of the WS2ND model structure. Figure 4 The dashed line model represents a short-term window model, the solid line model represents a long-term window model, and the model obtained on the right (dash line on top, solid line on the bottom) is the WS2ND model, which is the target click-through rate prediction model in this scheme. The upper part of each model is a dense model, and the lower part is a sparse model.

[0115] Understandably, the sparse model corresponding to the long time window model can be used as the candidate sparse model, and the dense model corresponding to the short time window model can be used as the candidate dense model. Then, the target click-through rate prediction model can be generated based on the candidate sparse model and the candidate dense model.

[0116] Furthermore, to achieve better model training results, the candidate sparse model can be used as the hot start model for the short time window model. The short time window model is then jointly trained based on the hot start candidate sparse model and candidate dense model to obtain the target click-through rate prediction model.

[0117] It should be understood that the sparse model in the short time window model can be replaced by the candidate sparse model after the warm start, so as to perform a warm start on the candidate dense model. The candidate sparse model and the candidate dense model are jointly trained to update the short time window model, and the updated short time window model is used as the target click-through rate prediction model.

[0118] In its implementation, this scheme employs deep learning transfer learning techniques, using a sparse model of a long-time window model as a "warm-start" model for a short-time window model. The short-time window model is then jointly trained based on the warm-started long-time window sparse model and the new dense model, i.e., WS2ND (Warm-up Sparse to NewDense). This approach ensures that the "contribution" of the long-time window model to the updating and iteration of the sparse feature parts is not lost, while also guaranteeing the "pluggable" convenience for adding new features and changing network structures.

[0119] In this embodiment, first time window data and second time window data are acquired, with the first time window data having a larger data volume than the second time window data. An initial model is trained using both the first and second time window data to obtain a long-time window model and a short-time window model. Transfer learning is then performed on the long-time window model and the short-time window model to obtain a target click-through rate (CTR) prediction model. By training the long-time window model and the short-time window model, and then performing transfer learning on these two models to obtain the target CTR prediction model, better CTR prediction results can be achieved when using the target CTR prediction model to predict the CTR of multimedia information.

[0120] In one embodiment, such as Figure 5 As shown, a third embodiment of the click-through rate prediction method of the present invention is proposed based on the first embodiment or the second embodiment. In this embodiment, the description is based on the first embodiment. Step S40 includes:

[0121] Step S401: If the feature to be detected has a time window difference problem, then determine whether there is a difference in the sorting of multimedia information.

[0122] Understandably, to achieve better click-through rate prediction results, one can refer to... Figure 6 , Figure 6 This is a flowchart of a hot-start combined with model knowledge distillation technique based on transfer learning. If there are time window differences in the features to be detected, it can be further determined whether there are differences in the ranking of multimedia information.

[0123] Furthermore, in order to accurately determine whether there are differences in the ranking of multimedia information, test data can be input into the long-term window model and the target click-through rate prediction model respectively to obtain the first click-through rate prediction result and the second click-through rate prediction result. Then, the first multimedia information ranking result is determined based on the first click-through rate prediction result, and the second multimedia information ranking result is determined based on the second click-through rate prediction result. The first multimedia information ranking result and the second multimedia information ranking result are compared, and the comparison result is used to determine whether there are differences in the ranking of multimedia information.

[0124] Understandably, if there are no differences in the sorting of multimedia information, it indicates that the target CTR prediction model and the long-term window model produce the same CTR prediction results, meaning the target CTR prediction model has high accuracy and can be directly used for CTR prediction. However, if there are differences in the sorting of multimedia information, it indicates that the target CTR prediction model and the long-term window model produce different CTR prediction results, meaning the target CTR prediction model is not accurate enough and needs optimization.

[0125] Step S402: If there is no difference in the sorting of multimedia information, the click-through rate of multimedia information is predicted using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

[0126] It should be understood that if there is no difference in the ranking of multimedia information, the click-through rate (CTR) of the multimedia information is predicted using the target CTR prediction model and the data to be tested, resulting in a CTR prediction result. If there is a difference in the ranking of multimedia information, the CTR of the multimedia information is predicted using the optimized target CTR prediction model and the data to be tested, resulting in a CTR prediction result.

[0127] It should be noted that this can be referred to Figure 7 , Figure 7 This diagram illustrates the knowledge distillation structure based on the WS2ND model. This solution can use knowledge distillation technology to optimize the target click-through rate (CTR) prediction model. It can perform knowledge distillation on the target CTR prediction model based on a long-term window model to obtain an optimized target CTR prediction model.

[0128] It should be understood that knowledge distillation follows a "teacher-student" training model. In knowledge distillation, the smaller model (student model) is typically supervised by a larger model (teacher model). The key challenge of the algorithm is how to transfer the knowledge transformed from the teacher model to the student model. Based on this model distillation technique, the teacher model is a long-window model, and the student model is a target click-through rate prediction model. While ensuring strong supervision of the existing learning task, the loss function of the student model is adjusted by the teacher model, thereby helping the student model learn more efficiently and effectively.

[0129] In practical implementation, the prediction direction of the WS2ND model structure combining the warm-start sparse model with the new dense model inevitably differs from that of the long-term window model. This is because the warm-start sparse model is a relatively mature model trained over a long time window. Binding it to the new dense model for incremental updates causes changes in the model's internal parameters due to the new gradient update direction, resulting in differences compared to the previous sparse parameters. Therefore, this solution introduces a "knowledge distillation" approach, which involves distilling the target click-through rate prediction model using the long-term window model.

[0130] Furthermore, in order to achieve better knowledge distillation results and improve the accuracy of the target click-through rate (CTR) prediction model, the step of performing knowledge distillation on the target CTR prediction model based on a long-term window model to obtain an optimized target CTR prediction model includes:

[0131] The soft tag loss function is determined based on the first click-through rate (CTR) prediction result corresponding to the long-term window model and the second CTR prediction result corresponding to the target CTR prediction model; the hard tag loss function is determined based on the second CTR prediction result; the overall loss function is determined based on the hard tag loss function and the soft tag loss function; and knowledge distillation is performed on the target CTR prediction model based on the overall loss function to obtain the optimized target CTR prediction model.

[0132] It should be noted that the overall loss function formula for knowledge distillation based on the WS2ND model structure is as follows:

[0133] L loss =αL hard label +βL soft label ;

[0134] Among them, L hard label This represents the loss function for click-through rate (CTR) prediction, specifically the cross-entropy loss function. The hard label is the label indicating whether a click has occurred in the CTR prediction scenario. L soft labelThe mean squared error function represents the error between the teacher model and the student model; the soft label is the mean squared error between the predicted values ​​of the teacher model and the predicted values ​​of the student model. α and β represent the loss weights of the hard-label loss function and the soft-label loss function, respectively, used to control the magnitude of the updated gradient during backpropagation.

[0135] It should be noted that the formula for the hard-label loss function is as follows:

[0136]

[0137] It should be noted that the formula for the soft-label loss function is as follows:

[0138] L soft label =MSE(pred teacher ,pred WS2ND );

[0139] Understandably, a first click-through rate (CTR) estimate can be determined based on the long-term window model, and a second CTR estimate can be determined based on the target CTR estimate model. Then, the software tag loss function can be determined based on both the first and second CTR estimates. Specifically, this can be done as follows: determine the first CTR estimate based on the first CTR estimate corresponding to the long-term window model; determine the second CTR estimate based on the second CTR estimate corresponding to the target CTR estimate model; calculate the mean squared error (MSE) using the aforementioned soft tag loss function formula based on the first and second CTR estimates; and determine the soft tag loss function based on the MSE.

[0140] Understandably, the hard label loss function can be determined based on the second click-through rate (CTR) prediction result. Specifically, this can be done by: determining the first CTR prediction value based on the first CTR prediction result corresponding to the long-term window model; and then determining the hard label loss function based on the first CTR prediction value using the aforementioned hard label loss function formula.

[0141] Understandably, after determining the soft tag loss function and the hard tag loss function, the overall loss function can be determined based on these two functions. Then, knowledge distillation is performed on the target click-through rate (CTR) prediction model using this overall loss function to obtain an optimized model. Specifically, the first weight value (α) corresponding to the hard tag loss function and the second weight value (β) corresponding to the soft tag loss function are obtained. The overall loss function formula is then calculated using the hard tag loss function, the soft tag loss function, the first weight value, and the second weight value, and the overall loss function is determined based on the weighted calculation result.

[0142] In this embodiment, if the features to be detected exhibit time window differences, it is determined whether there are differences in the ranking of multimedia information. If there are no differences in the ranking of multimedia information, the click-through rate (CTR) of the multimedia information is predicted using the target CTR prediction model and the data to be detected, yielding the CTR prediction result. The accuracy of the target CTR prediction model can be checked by examining whether there are differences in the ranking of multimedia information. If the accuracy is insufficient, knowledge distillation is used to improve the accuracy of the target CTR prediction model. This solution combines transfer learning and knowledge distillation to train the target CTR prediction model, ensuring both the CTR prediction effect and accuracy of the target CTR prediction model.

[0143] Furthermore, this embodiment of the invention also proposes a storage medium storing a click-through rate (CTR) prediction program, which, when executed by a processor, implements the steps of the CTR prediction method described above.

[0144] Since this storage medium adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be repeated here.

[0145] In addition, refer to Figure 8 This invention also proposes a click-through rate (CTR) prediction device, which includes:

[0146] The data acquisition module 10 is used to acquire the data to be detected.

[0147] It should be noted that the multimedia information in this embodiment may include, but is not limited to, advertisements, videos, images, and audio, as well as other types of multimedia information. This embodiment does not impose any limitations on this; in this embodiment, advertisements are used as an example of multimedia information for explanation. Accordingly, the multimedia information click-through rate prediction in this embodiment may specifically be an advertisement click-through rate prediction. The aforementioned advertisements may include, but are not limited to, various types of advertisements such as image advertisements, text advertisements, keyword advertisements, ranking advertisements, and video advertisements. This embodiment does not impose any limitations on this.

[0148] It should be understood that in recent years, driven by the rapid development of machine learning and deep learning, ad click-through rate (CTR) prediction technology has been continuously iterating and updating. Deep learning models, with their superior generalization performance, have gained recognition from major internet companies, making deep learning-based ad CTR prediction a widely accepted application. Ad CTR prediction models are based on daily-level feature data and are trained incrementally. Therefore, CTR prediction models trained with different time windows, especially those with longer time windows, have better prediction capabilities and stronger generalization abilities than those with shorter time windows. For the iteration and updating of existing models, it is necessary not only to optimize and update existing models of similar performance but also to further compensate for the model gains brought by the time window. Therefore, the high dependence of CTR prediction models on data features and the incremental updates accompanying time windows have become bottlenecks restricting their efficient and rapid iteration, and a long-term business requirement for advertising algorithm engineers.

[0149] To address the issue of compensating for the time window gains, some internet companies have proposed a "timed restart" solution. This involves retraining a model based on a manually set time window and directly replacing the online incremental model. This ensures that the data time window upon which the online incremental model relies is controllable, and that the gains of the new model during model iteration do not need to compensate for the time window gains of the original model. However, existing "timed restart" solutions are "lossy" in model prediction. Longer time window models see more data samples than shorter time window models. Therefore, for head features with high feature frequency, feature weights and representations learn to fit better due to gradient updates from multiple data samples. For long-tail features with low feature frequency, models that have seen more data are more likely to avoid the "cold start" problem. Thus, the gains are "lossy" for the model. The main work of this paper is to propose a better solution to address the time window model gain problem, and this solution has been tested and applied in real-world business scenarios.

[0150] It should be noted that this solution offers a superior approach to addressing the revenue issues associated with time-window models, and has been tested and applied in real-world business scenarios. This solution optimizes ad click-through rate prediction for search advertising scenarios, primarily to compensate for the additional model gains resulting from longer window times. To achieve this technical approach, this solution mainly utilizes transfer learning techniques.

[0151] It should be noted that the aforementioned transfer learning technique specifically refers to transfer learning warm-up techniques. Transfer learning warm-up, or transfer learning warm-up, involves using existing knowledge (source domain) and new knowledge (target domain) to transfer knowledge from the source domain to the target domain. In particular, in deep learning, transfer learning studies how to apply existing models to new, different, but related domains. Traditional machine learning is less flexible and produces less satisfactory results when dealing with tasks involving variations in data distribution, feature dimensionality, and model output. Transfer learning relaxes these assumptions. It organically utilizes knowledge from the source domain to better model the target domain, adapting to changes in data distribution, feature dimensionality, and model output. Furthermore, when labeled data is scarce, transfer learning can effectively utilize labeled data from related domains to perform data labeling. This approach employs the concept of transfer learning, using the click-through rate prediction of the long-term window model as the source domain and the click-through rate prediction of the short-term window model as the target domain. By warm-starting the short-term window model with the sparse model of the long-term window model, the approach further fits the data that the short-term window model has not seen.

[0152] It should be understood that when a multimedia information click-through rate prediction instruction is received, the data to be detected can be obtained according to the multimedia information click-through rate prediction instruction. Specifically, the data to be detected can be obtained from the advertising log system based on the instruction, which contains advertising data of previously recommended advertisements. Alternatively, the data to be detected can be obtained through other means based on the instruction. This embodiment does not limit this.

[0153] The feature extraction module 20 is used to extract features based on the data to be detected, and obtain the features to be detected.

[0154] It should be understood that after obtaining the data to be detected, in order to determine whether a model warm-start is needed, feature extraction can be performed on the data to obtain the features to be detected, and then the model warm-start can be used to determine whether a model warm-start is needed. If a model warm-start is needed, the target click-through rate (CTR) prediction model in this solution is used to predict the ad CTR. If a model warm-start is not needed, a conventional CTR prediction model is used. In this embodiment, the conventional CTR prediction model can include various models capable of CTR prediction, and this embodiment does not limit the specific model form of the conventional CTR prediction model.

[0155] It is understandable that in the scenario of predicting ad click-through rate, some ad features related to the ad can be extracted from the data to be detected. For example, ad features may include, but are not limited to, ad duration, ad type and ad playback time. This embodiment does not limit this.

[0156] The feature detection module 30 is used to determine whether there is a time window difference problem in the feature to be detected.

[0157] It should be understood that the presence of time window differences in the data to be detected can be determined based on the characteristics to be detected. Specifically, if the time span between the data to be detected is large, it can be determined that there is a time window difference problem; if the time span between the data to be detected is small, it can be determined that there is no time window difference problem.

[0158] Click-through rate (CTR) prediction module 40 is used to predict the CTR of multimedia information by using the target CTR prediction model and the data to be detected if there is a time window difference problem in the feature to be detected, and to obtain the CTR prediction result.

[0159] It should be understood that if a time window difference is detected in the feature to be detected, it can be determined that a model warm-start is required. Therefore, in this case, the click-through rate of multimedia information can be predicted by using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result can be obtained.

[0160] It should be noted that the target click-through rate (CTR) prediction model in this solution can be obtained through transfer learning from the long-term window model and the short-term window model. Both the long-term window model and the short-term window model are CTR prediction models. However, since the long-term window model has more training data than the short-term window model, the accuracy of the long-term window model is higher than that of the short-term window model.

[0161] Understandably, this solution employs a target click-through rate (CTR) prediction model obtained through transfer learning from both long-window and short-window models to predict CTR for multimedia information when a warm-start model is required. This ensures that the benefits of the long-window model are not lost. Simultaneously, for algorithm engineers who offline mine and optimize the CTR prediction model, there's no need to expend effort compensating for the benefits gained from the short-window model; they can focus on business improvement, further enhancing model iteration efficiency and traffic monetization capabilities. This has profound and significant implications for optimizing CTR prediction-related business processes.

[0162] In this embodiment, data to be detected is acquired; features are extracted from the data to be detected to obtain the features to be detected; it is determined whether the features to be detected have a time window difference problem; if the features to be detected have a time window difference problem, the click-through rate (CTR) of multimedia information is predicted using a target CTR prediction model and the data to be detected to obtain the CTR prediction result. This solution can determine the need for a model hot start when the features to be detected have a time window difference problem, and then perform CTR prediction of multimedia information using a target CTR prediction model, thereby optimizing the CTR prediction business and improving the CTR prediction effect of multimedia information.

[0163] In one embodiment, the click-through rate prediction device further includes: a model training module and a transfer learning module; the data acquisition module is further configured to acquire first time window data and second time window data, wherein the amount of data in the first time window data is greater than that in the second time window data; the model training module is configured to train an initial model based on the first time window data and the second time window data respectively to obtain a long time window model and a short time window model; the transfer learning module is configured to perform transfer learning based on the long time window model and the short time window model to obtain a target click-through rate prediction model.

[0164] In one embodiment, the transfer learning module is further configured to use the sparse model corresponding to the long time window model as a candidate sparse model; use the dense model corresponding to the short time window model as a candidate dense model; and generate a target click-through rate prediction model based on the candidate sparse model and the candidate dense model.

[0165] In one embodiment, the transfer learning module is further configured to use the candidate sparse model as a hot-start model for the short-time window model; and to obtain a target click-through rate prediction model by jointly training the short-time window model based on the candidate sparse model and the candidate dense model after the hot start.

[0166] In one embodiment, the transfer learning module is further configured to replace the sparse model in the short time window model with the candidate sparse model after a warm start, so as to perform a warm start on the candidate dense model; perform joint training on the candidate sparse model and the candidate dense model to update the short time window model; and use the updated short time window model as the target click-through rate prediction model.

[0167] In one embodiment, the click-through rate prediction module 40 is further configured to determine whether there is a difference in the sorting of multimedia information if there is a time window difference problem in the feature to be detected; if there is no difference in the sorting of multimedia information, the click-through rate of multimedia information is predicted by the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

[0168] In one embodiment, the click-through rate (CTR) prediction module 40 is further configured to input test data into the long-term window model and the target CTR prediction model respectively to obtain a first CTR prediction result and a second CTR prediction result; determine a first multimedia information ranking result based on the first CTR prediction result; determine a second multimedia information ranking result based on the second CTR prediction result; compare the first multimedia information ranking result with the second multimedia information ranking result; and determine whether there is a difference in the multimedia information ranking based on the comparison result.

[0169] In one embodiment, the click-through rate (CTR) prediction device further includes a model optimization module. If there are differences in the sorting of multimedia information, the CTR of the multimedia information is predicted based on the optimized target CTR prediction model and the data to be detected, and the CTR prediction result is obtained.

[0170] In one embodiment, the model optimization module is further configured to perform knowledge distillation on the target click-through rate prediction model based on the long-term window model to obtain an optimized target click-through rate prediction model.

[0171] In one embodiment, the model optimization module is further configured to: determine a soft tag loss function based on the first click-through rate (CTR) prediction result corresponding to the long-term window model and the second CTR prediction result corresponding to the target CTR prediction model; determine a hard tag loss function based on the second CTR prediction result; determine an overall loss function based on the hard tag loss function and the soft tag loss function; and perform knowledge distillation on the target CTR prediction model based on the overall loss function to obtain an optimized target CTR prediction model.

[0172] In one embodiment, the model optimization module is further configured to: determine a first click-through rate (CTR) estimate based on the first CTR estimate result corresponding to the long-term window model; determine a second CTR estimate based on the second CTR estimate result corresponding to the target CTR estimate model; calculate a mean squared error based on the first CTR estimate and the second CTR estimate; and determine a soft tag loss function based on the mean squared error.

[0173] In one embodiment, the model optimization module is further configured to obtain a first weight value corresponding to the hard-label loss function and a second weight value corresponding to the soft-label loss function; perform a weighted calculation based on the hard-label loss function, the soft-label loss function, the first weight value, and the second weight value; and determine the overall loss function based on the weighted calculation result.

[0174] Other embodiments or specific implementation methods of the click-through rate prediction device described in this invention can be referred to the above-described method embodiments, and will not be repeated here.

[0175] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0176] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0177] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This estimation machine software product is stored in an estimation machine readable storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a smart device (which may be a mobile phone, estimation machine, click-through rate prediction device, or web click-through rate prediction device, etc.) to execute the methods described in the various embodiments of the present invention.

[0178] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A click-through rate (CTR) prediction method, characterized in that, The click-through rate prediction method includes: Acquire the data to be tested; Feature extraction is performed based on the data to be detected to obtain the features to be detected; Determine whether the feature to be detected has a time window difference problem; If the feature to be detected has a time window difference problem, then the click-through rate of multimedia information is predicted by using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained. Before obtaining the click-through rate prediction result by using the target click-through rate prediction model and the data to be detected to predict the click-through rate of multimedia information, the process also includes: Acquire data from a first time window and a second time window, wherein the data volume of the first time window is greater than that of the second time window. The initial model is trained using the first time window data and the second time window data respectively to obtain a long time window model and a short time window model; A target click-through rate prediction model is obtained by performing transfer learning based on the long-term window model and the short-term window model.

2. The click-through rate prediction method as described in claim 1, characterized in that, The target click-through rate prediction model is obtained by transfer learning from a long-time window model and a short-time window model. The accuracy of the long-time window model is higher than that of the short-time window model.

3. The click-through rate prediction method as described in claim 2, characterized in that, The step of performing transfer learning based on the long-time window model and the short-time window model to obtain the target click-through rate prediction model includes: The sparse model corresponding to the long window model is selected as the candidate sparse model. The dense model corresponding to the short time window model is selected as the candidate dense model; A target click-through rate prediction model is generated based on the candidate sparse model and the candidate dense model.

4. The click-through rate prediction method as described in claim 3, characterized in that, The step of generating a target click-through rate prediction model based on the candidate sparse model and the candidate dense model includes: The candidate sparse model is used as the hot start model for the short time window model; The target click-through rate prediction model is obtained by jointly training the candidate sparse model and the candidate dense model after a warm start using the short time window model.

5. The click-through rate prediction method as described in claim 4, characterized in that, The step of jointly training the short-time window model based on the candidate sparse model and the candidate dense model after a warm start to obtain the target click-through rate prediction model includes: The candidate sparse model is replaced in the short time window model by the candidate sparse model after the warm start, so as to perform a warm start on the candidate dense model. The short-time-window model is updated by jointly training the candidate sparse model and the candidate dense model. The updated short time window model is used as the target click-through rate prediction model.

6. The click-through rate prediction method as described in any one of claims 1 to 5, characterized in that, If the feature to be detected has a time window difference problem, then the click-through rate (CTR) of multimedia information is predicted using the target CTR prediction model and the data to be detected, and the CTR prediction result is obtained, including: If the features to be detected have time window differences, then determine whether there are differences in the sorting of multimedia information; If there is no difference in the ranking of multimedia information, the click-through rate of multimedia information is predicted using the target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

7. The click-through rate prediction method as described in claim 6, characterized in that, The determination of whether there are differences in the sorting of multimedia information includes: The test data is input into the long window model and the target click-through rate prediction model respectively to obtain the first click-through rate prediction result and the second click-through rate prediction result. The first multimedia information ranking result is determined based on the first click-through rate prediction result; The second multimedia information ranking result is determined based on the second click-through rate prediction result; Compare the first multimedia information sorting result with the second multimedia information sorting result; Determine whether there are differences in the sorting of multimedia information based on the comparison results.

8. The click-through rate prediction method as described in claim 6, characterized in that, After determining whether there are differences in the sorting of multimedia information, the method further includes: If there are differences in the sorting of multimedia information, the click-through rate of the multimedia information is predicted based on the optimized target click-through rate prediction model and the data to be detected, and the click-through rate prediction result is obtained.

9. The click-through rate prediction method as described in claim 8, characterized in that, Before obtaining the click-through rate prediction result by performing multimedia information click-through rate prediction based on the optimized target click-through rate prediction model and the data to be detected, the method further includes: The target click-through rate prediction model is optimized by knowledge distillation based on the long window model.

10. The click-through rate prediction method as described in claim 9, characterized in that, The step of performing knowledge distillation on the target click-through rate (CTR) prediction model based on the long-term window model to obtain an optimized target CTR prediction model includes: The soft tag loss function is determined based on the first click-through rate prediction result corresponding to the long window model and the second click-through rate prediction result corresponding to the target click-through rate prediction model. Determine the hard tag loss function based on the second click-through rate prediction result; The overall loss function is determined based on the hard label loss function and the soft label loss function; Based on the overall loss function, knowledge distillation is performed on the target click-through rate prediction model to obtain an optimized target click-through rate prediction model.

11. The click-through rate prediction method as described in claim 10, characterized in that, The step of determining the soft tag loss function based on the first click-through rate (CTR) prediction result corresponding to the long-term window model and the second CTR prediction result corresponding to the target CTR prediction model includes: The first click-through rate (CTR) estimate is determined based on the first CTR estimate result corresponding to the long window model. The second click-through rate (CTR) estimate is determined based on the second CTR estimate result corresponding to the target CTR estimate model. Calculate the mean squared error based on the first and second click-through rate estimates; The soft label loss function is determined based on the mean square error.

12. The click-through rate prediction method as described in claim 10, characterized in that, The step of determining the overall loss function based on the hard-label loss function and the soft-label loss function includes: Obtain the first weight value corresponding to the hard label loss function, and obtain the second weight value corresponding to the soft label loss function; The weighted calculation is performed based on the hard label loss function, the soft label loss function, the first weight value, and the second weight value. The overall loss function is determined based on the weighted calculation results.

13. A click-through rate prediction device, characterized in that, The click-through rate prediction device includes: The data acquisition module is used to acquire the data to be detected. The feature extraction module is used to extract features from the data to be detected to obtain the features to be detected. The feature detection module is used to determine whether the feature to be detected has a time window difference problem; The click-through rate (CTR) prediction module is used to predict the CTR of multimedia information by using the target CTR prediction model and the data to be detected if there is a time window difference problem in the feature to be detected, and to obtain the CTR prediction result. The click-through rate prediction device further includes: a model training module and a transfer learning module; The data acquisition module is also used to acquire first time window data and second time window data, wherein the amount of data in the first time window data is greater than that in the second time window data. The model training module is used to train the initial model based on the first time window data and the second time window data respectively, to obtain a long time window model and a short time window model. The transfer learning module is used to perform transfer learning based on the long-time window model and the short-time window model to obtain a target click-through rate prediction model.

14. The click-through rate prediction device as described in claim 13, characterized in that, The target click-through rate prediction model is obtained by transfer learning from a long-time window model and a short-time window model. The accuracy of the long-time window model is higher than that of the short-time window model.

15. The click-through rate prediction device as described in claim 13, characterized in that, The transfer learning module is further configured to use the sparse model corresponding to the long time window model as a candidate sparse model; use the dense model corresponding to the short time window model as a candidate dense model; and generate a target click-through rate prediction model based on the candidate sparse model and the candidate dense model.

16. The click-through rate prediction device as described in claim 15, characterized in that, The transfer learning module is further configured to use the candidate sparse model as the hot-start model of the short-time window model; and to obtain the target click-through rate prediction model by jointly training the short-time window model based on the candidate sparse model and the candidate dense model after the hot start.

17. A click-through rate (CTR) prediction device, characterized in that, The click-through rate (CTR) prediction device includes: a memory, a processor, and a CTR prediction program stored in the memory and executable on the processor, wherein the CTR prediction program, when executed by the processor, implements the CTR prediction method as described in any one of claims 1 to 12.

18. A storage medium, characterized in that, The storage medium stores a click-through rate (CTR) prediction program, which, when executed by a processor, implements the CTR prediction method as described in any one of claims 1 to 12.

Images