A training method of an object value prediction model and a related device

Abstract

The application discloses a training method and related device of an object value prediction model, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and auxiliary driving, obtains a training data set, the training data set includes historical object data of a historical registered object, performs feature extraction on the historical object data through a feature extraction module in an initial network model to obtain a historical object feature vector, outputs a predicted conversion behavior of the historical registered object according to the historical object feature vector through a time sequence module in the initial network model, optimizes model parameters of the initial network model based on differences between the predicted conversion behavior and a calibrated conversion behavior of different time nodes in a target time period, and obtains an object value prediction model. The object value prediction model can output a conversion behavior of a to-be-predicted object at different time nodes in the target time period, reflect a continuous conversion potential of the to-be-predicted object, and further evaluate whether it is beneficial to improve the ROI of a product by putting a recommended content.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a training method and related apparatus for an object value prediction model. Background Technology

[0002] In scenarios where user acquisition is targeted at specific content (such as advertising), when it is necessary to determine whether to purchase a particular ad space to run an ad, the value of the ad space is usually evaluated to make a decision on whether to purchase the ad space for running the ad, thereby maximizing the return on investment (ROI).

[0003] The revenue generated by placing an ad in a particular ad placement will vary depending on the target audience. Therefore, the value of an ad placement can be reflected through the target audience value, which refers to the payment behavior of the target audience (such as users) after registering through the ad. Thus, predicting the target audience value is crucial for improving ROI.

[0004] However, the object value predicted by related technologies cannot fully characterize the object's payment behavior and payment habits, and it is difficult to explicitly describe the object's potential value, thus affecting the prediction effect of the value of the ad space. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a training method and related apparatus for an object value prediction model. The resulting object value prediction model can predict the conversion behavior of the object to be predicted at different time points within a target time period, thus providing a more comprehensive and accurate characterization of the potential value of the object to be predicted. Based on this object value assessment, it can be determined whether delivering product recommendations to the object to be predicted is beneficial to improving the product's ROI.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] On one hand, embodiments of this application provide a training method for an object value prediction model, the method comprising:

[0008] Obtain a training dataset, which includes historical object data of historical registered objects. The historical object data has a corresponding set of calibration conversion behaviors. The set of calibration conversion behaviors includes the calibration conversion behaviors of the historical registered objects at different time points within the target time period after registration is completed. The calibration conversion behaviors at different time points are arranged in chronological order.

[0009] The feature extraction module in the initial network model extracts features from the historical object data to obtain the historical object feature vector.

[0010] Based on the feature vector of the historical object, the time series module in the initial network model outputs the predicted transformation behavior of the historical registered object at different time nodes within the target time period, and the predicted transformation behavior at different time nodes is arranged in the time order.

[0011] Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period, the model parameters of the initial network model are optimized to obtain the object value prediction model.

[0012] In another aspect, embodiments of this application provide a training apparatus for an object value prediction model, the apparatus comprising an acquisition unit, an extraction unit, an output unit, and an optimization unit:

[0013] The acquisition unit is used to acquire a training dataset, which includes historical object data of historical registered objects. The historical object data has a corresponding set of calibration conversion behaviors. The set of calibration conversion behaviors includes the calibration conversion behaviors of the historical registered objects at different time nodes within the target time period after registration is completed. The calibration conversion behaviors at different time nodes are arranged in chronological order.

[0014] The extraction unit is used to extract features from the historical object data through the feature extraction module in the initial network model to obtain the historical object feature vector;

[0015] The output unit is used to output the predicted transformation behavior of the historical registered object at different time nodes within the target time period based on the feature vector of the historical object and through the time sequence module in the initial network model. The predicted transformation behavior at different time nodes is arranged in the time order.

[0016] The optimization unit is used to optimize the model parameters of the initial network model based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time nodes within the target time period, so as to obtain the object value prediction model.

[0017] On the other hand, embodiments of this application provide a computer device, the computer device including a processor and a memory:

[0018] The memory is used to store computer programs and to transfer the computer programs to the processor;

[0019] The processor is configured to execute the method described in any of the foregoing aspects according to instructions in the computer program.

[0020] On the other hand, embodiments of this application provide a computer-readable storage medium for storing a computer program for performing the methods described in any of the foregoing aspects.

[0021] On the other hand, embodiments of this application provide a computer program product, including a computer program that, when run on a computer device, causes the computer device to execute the method described in any of the foregoing aspects.

[0022] As can be seen from the above technical solution, by acquiring a training dataset, which includes historical object data of historical registered objects, and then using the feature extraction module in the initial network model to extract features from the historical object data to obtain historical object feature vectors, the time series module in the initial network model can output the predicted transformation behavior of historical registered objects based on the historical object feature vectors. Since historical object data has a corresponding set of calibrated conversion behaviors, which includes the calibrated conversion behaviors of historical registered objects at different time points within the target time period after registration, representing the true value of historical registered objects at different time points within the target time period after registration, and the predicted conversion behaviors represent the predicted value of historical registered objects at different time points within the target time period after registration, as determined by the initial network model, the model parameters of the initial network model are optimized based on the differences between the predicted conversion behaviors and the calibrated conversion behaviors at different time points within the target time period. This yields an object value prediction model, which can predict the object value at different time points within the target time period and output the conversion behaviors of the object to be predicted at different time points within the target time period. By arranging the conversion behaviors at different time points in chronological order, the object value of the object to be predicted can be obtained. The object value obtained in this way is a continuous value, which can reflect the continuous conversion potential of the object to be predicted within the target time period. It can more comprehensively and accurately depict the potential value of the object to be predicted. Therefore, based on this object value, it is beneficial to evaluate whether to deliver product recommendations to the object to be predicted to improve the ROI of the product. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A schematic diagram of a training scenario for an object value prediction model provided in an embodiment of this application;

[0025] Figure 2 A flowchart illustrating a training method for an object value prediction model provided in an embodiment of this application;

[0026] Figure 3 A schematic diagram illustrating a training method for an object value prediction model provided in an embodiment of this application;

[0027] Figure 4 A schematic diagram illustrating a training method for another object value prediction model provided in an embodiment of this application;

[0028] Figure 5 This is a schematic diagram illustrating the predicted transformation behavior of a time-series module outputting the i-th time node, as provided in an embodiment of this application.

[0029] Figure 6a A schematic diagram illustrating a training method for another object value prediction model provided in an embodiment of this application;

[0030] Figure 6b A schematic diagram illustrating a training method for another object value prediction model provided in an embodiment of this application;

[0031] Figure 7 A schematic diagram illustrating a method for constructing a training dataset according to an embodiment of this application;

[0032] Figure 8 This is a schematic diagram illustrating a product recommendation content delivery scenario provided in an embodiment of this application.

[0033] Figure 9 This is a schematic diagram illustrating another product recommendation content delivery scenario provided in an embodiment of this application.

[0034] Figure 10 A structural diagram of a training device for an object value prediction model provided in an embodiment of this application;

[0035] Figure 11 A structural diagram of a terminal device provided in an embodiment of this application;

[0036] Figure 12 This is a structural diagram of a server provided in an embodiment of this application. Detailed Implementation

[0037] The embodiments of this application will now be described with reference to the accompanying drawings.

[0038] For products to be recommended, product owners often purchase advertising space and then place recommended content for the product on the purchased advertising space, hoping to generate revenue through the recommended content displayed on the advertising space. The product can provide various services to the target audience or has a certain value relative to the target audience. This application does not limit the type of product; for example, it can be an application, a product, etc. An application could be a game application, a karaoke application, etc.

[0039] Ad placements are used to display recommended content and can be a specific area on the application's interface. The revenue generated by placing recommended content on an ad placement varies depending on the target audience. Therefore, the value of ad placements targeting different audiences must be evaluated before placement. Generally, the value of an ad placement can be reflected in its target audience value, which refers to the conversion behavior generated by the target audience through the recommended content displayed on the ad placement.

[0040] The object can be a user. The conversion behavior of the object can refer to the relevant behavior generated by the object based on the recommended content, which can bring certain revenue to the product provider. For example, the conversion behavior can be that the object purchases a product based on the recommended content, that is, the object makes a payment based on the recommended content.

[0041] Therefore, when placing ads, product developers predict the value of the target audience to assess the value of the ad space, and then decide whether to purchase the ad space in order to maximize the product's ROI. It is clear that predicting the value of the target audience is crucial for ad space purchase decisions, and consequently, it is also crucial for improving the product's ROI.

[0042] In related technologies, modeling is primarily based on the cumulative payments made by an object from its registration date to a specific day, thereby predicting the object's cumulative payment behavior as its value. However, this method fails to explicitly describe the object's potential value. For example, within seven days from the registration date, user A only paid 70 yuan on the first day and did not pay on the remaining six days, while user B paid 10 yuan per day for those seven days. Since user A and user B's cumulative payment behavior is consistent over those seven days, their predicted object values ​​using the cumulative payment method are the same. However, their payment habits during this period are clearly different. Based on this difference in payment habits, user B is more likely to continue paying, and consequently, user B's potential value can be considered higher than user A's. It is evident that this method's predicted object value deviates significantly from the object's true value, especially since cumulative payment behavior cannot fully characterize the object's payment behavior and habits. Therefore, it is difficult to explicitly describe the object's potential value, thus affecting the prediction of ad space value and ultimately impacting the product's ROI.

[0043] Therefore, this application provides a training method for an object value prediction model. The resulting object value prediction model can predict the conversion behavior of the object to be predicted at different time points within a target time period. Compared with the object value characterized by cumulative payment behavior, the conversion behavior of the object to be predicted at different time points within a target time period is a manifestation of continuous value. It can reflect the continuous conversion potential of the object to be predicted within the target time period and can more comprehensively and accurately characterize the potential value of the object to be predicted. Based on this object value, it is beneficial to evaluate whether to deliver product recommendation content to the object to be predicted to improve the ROI of the product.

[0044] The training method for the object value prediction model provided in this application can be implemented using a computer device, which can be a terminal device or a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminal devices include, but are not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, and vehicle terminals. Terminal devices and servers can be directly or indirectly connected via wired or wireless communication, which is not limited herein. This invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0045] This application can be applied to the field of artificial intelligence (AI). AI is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0046] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, smart customer service, vehicle networking, and intelligent transportation. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0047] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0048] This application primarily relates to machine learning. Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0049] For example, in the embodiments of this application, an initial network model can be trained based on machine learning to obtain an object value prediction model, which is used to predict the object value.

[0050] Figure 1 This illustration shows a model training scenario provided in an embodiment of this application. Figure 1 In the scenario shown, server 100 is used as an example of the aforementioned computer device for illustration:

[0051] For the initial network model, a training dataset for model training can be obtained through server 100. In this embodiment, the training dataset includes historical object data of historically registered objects. This historical object data has a corresponding set of calibrated conversion behaviors. The set of calibrated conversion behaviors includes the calibrated conversion behaviors of historically registered objects at different time points within a target time period after registration. The calibrated conversion behaviors at different time points are arranged in chronological order, and the calibrated conversion behaviors at different time points represent the true value of historically registered objects at different time points. Here, the historically registered object can be a user who has already completed registration, and the calibrated conversion behaviors can be the actual conversion behaviors of the historically registered object.

[0052] Furthermore, the feature extraction module in the initial network model can extract features from the historical object data in the training dataset to obtain historical object feature vectors. Then, the temporal module in the initial network model outputs the predicted transformation behavior of historical registered objects at different time points within the target time period based on these historical object feature vectors. Correspondingly, the predicted transformation behaviors at different time points are arranged in chronological order. The predicted transformation behavior is the transformation behavior of the historical registered objects predicted by the initial network model.

[0053] Since the calibration transformation behavior at different time points within the target time period is used to represent the true value of historically registered objects at different time points, and the predicted transformation behavior at different time points within the target time period is used to represent the predicted value of historically registered objects at different time points, the difference between the calibration transformation behavior and the predicted transformation behavior at the same time point can reflect the accuracy of the initial network model's prediction of the transformation behavior of historically registered objects at that time point. Based on this difference, the model parameters of the initial network model are optimized, and the object value prediction model is obtained after optimization.

[0054] The object value prediction model obtained based on the above training process can be used to predict the conversion behavior of the object to be predicted at different time nodes within the target time period. Correspondingly, it can output the conversion behavior of the object to be predicted at different time nodes within the target time period. By arranging the conversion behavior at different time nodes in chronological order, the object value of the object to be predicted can be obtained. The object value obtained in this way is a continuous value, which can reflect the continuous conversion potential of the object to be predicted within the target time period. It can more comprehensively and accurately depict the potential value of the object to be predicted. Then, based on this object value, it can be evaluated whether it is beneficial to deliver product recommendation content to the object to be predicted to improve the ROI of the product.

[0055] Figure 2 A flowchart illustrating a training method for an object value prediction model provided in this application embodiment, using a server as an example of the aforementioned computer device, shows the method comprising S201-S204:

[0056] S201: Obtain the training dataset.

[0057] For a specific product, after purchasing advertising space, the product's recommended content is placed in that space. The goal is for potential customers to view the recommended content through the ad space and convert their purchases, thus generating revenue for the product. To improve the product's ROI, the decision to run ad campaigns is made by predicting the value of the target audience. Generally, the more stable the target audience's value, the more stable the ROI from running the product's recommended content. Therefore, predicting the target audience's value is crucial for improving the product's ROI.

[0058] Generally, object value prediction can be achieved using an object value prediction model, which can be obtained through model training. Specifically, during the model training phase, a training dataset is first acquired. This dataset includes historical object data of registered objects, which has a corresponding set of calibrated conversion behaviors. This set includes the calibrated conversion behaviors of historical registered objects at different time points within the target time period after registration, arranged chronologically. Each time point's calibrated conversion behavior represents the true value of the historical registered object at that time. Therefore, by arranging the calibrated behaviors at different time points chronologically, the continuous true value of historical registered objects within the target time period can be obtained. This comprehensively and accurately depicts the conversion behavior, conversion habits, and continuous conversion status of historical registered objects within the target time period. Thus, using this training dataset to train the initial network model enables the final object value prediction model to predict the continuous predicted value of the object within the target time period, thereby supporting the product's recommended content delivery decisions.

[0059] The target time period can be a certain period of time, and the time nodes can be the various time nodes determined by dividing the target time period according to certain time division units. The specific settings can be set according to the actual forecasting needs, and this application does not impose any restrictions on them. For example, the seven-day period from the registration date can be set as the target time period, and correspondingly, it can be divided into seven time nodes, namely the first day, the second day... the seventh day, using days as the time division unit.

[0060] It should be noted that the historical registration targets can be selected based on the actual situation, and this application does not impose any restrictions on this. For example, for product A that requires recommended content delivery, the historical registration targets can be determined based on the product stage that product A is in. For ease of understanding, the following determination method is provided as an example in the embodiments of this application:

[0061] If product A is in a stable operational phase, meaning it has been online for some time and has a certain number of registered users, then to acquire new users by promoting product A, you can select these registered users as your historical registration targets. If product A is in a cold start phase, meaning it's a new product, and you want to acquire registered users by promoting it, since no users have used product A yet, you can first identify product B, which is related to product A (e.g., similar product functions, similar target user groups), and then select registered users of product B as your historical registration targets. Alternatively, you can select a portion of registered users of product A and a portion of registered users of product B as your historical registration targets.

[0062] For ease of understanding, this embodiment assumes that the target time period includes k time nodes, where k is an integer greater than 1. Correspondingly, the labeling and conversion behaviors of historically registered objects at different time nodes within the target time period are arranged in chronological order, resulting in the following: Figure 3 Shown:

[0063]

[0064] Among them, y k This represents the calibration and transformation behavior of a historically registered object at the k-th time node, where k is an integer greater than 1.

[0065] S202: Extract features from historical object data using the feature extraction module in the initial network model to obtain historical object feature vectors.

[0066] After obtaining the training dataset, the feature extraction module in the initial network model can be used to extract features from the historical object data of the historical registered objects included in the training dataset, obtaining historical object feature vectors. This allows the time-series module in the initial network model to output corresponding prediction and transformation behaviors based on these historical object feature vectors. The initial network model can be found in [link to initial network model]. Figure 3 The feature extraction module shown in 101 can be found in [reference needed]. Figure 3 The timing module shown in 1011 can be found in [reference needed]. Figure 3 The number 1012 is shown.

[0067] The feature extraction module is used to extract features from historical object data. It mainly includes extracting key object features and representation object features. Key object features may include the historical conversion behavior of historical registered objects (e.g., the payment status of historical registered objects within 7 days after registration). Representation object features may include the natural person features of historical registered objects (e.g., the age and gender of historical registered objects).

[0068] It should be noted that this application does not impose any limitations on the network structure of the feature extraction module. For ease of understanding, a feature extraction module consisting of a two-layer fully connected network is used as an example. Specifically, feature extraction of historical object data can be implemented as follows:

[0069] h = σ(W1x + b1)

[0070] e(x)=σ(W2h+b2)

[0071] Where x is the vector representation of historical object data, σ is an arbitrary nonlinear activation function, W1, b1 and W2, b2 are the parameters of the first and second fully connected networks, respectively, h is the vector after processing x using the first fully connected network, and e(x) is the historical object feature vector. Specifically, floating-point features can be normalized, categorical features can be randomly initialized as floating-point vectors, and then the floating-point features and categorical features can be concatenated into a feature vector x. x is then input into the first fully connected network, which processes it and outputs h. h is then input into the second fully connected network, which processes it and outputs a higher-order feature vector representation e(x) as the historical object feature vector.

[0072] S203: Based on the feature vector of historical objects, the predicted transformation behavior of historical registered objects at different time nodes within the target time period is output through the time series module in the initial network model.

[0073] For the feature vectors of historical objects, the temporal module in the initial network model can output the predicted transformation behavior of historical registered objects at different time points within the target time period. The predicted transformation behavior at each time point represents the predicted value of the historical registered object at that time point, as predicted by the initial network model. Accordingly, when the target time period includes k time points, there are k corresponding predicted transformation behaviors. Arranging the predicted transformation behaviors at different time points within the target time period in chronological order yields the following... Figure 3 Shown:

[0074]

[0075] in, This represents the predicted transformation behavior of a historically registered object at the k-th time point, where k is an integer greater than 1.

[0076] S204: Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period, optimize the model parameters of the initial network model to obtain the object value prediction model.

[0077] Since the calibration transformation behavior at each time point represents the true value of a historically registered object at that time point, while the prediction transformation behavior at that time point represents the predicted value of the historically registered object at that time point, the difference between the two can reflect the difference between the predicted value predicted by the initial network model and the true value. That is, the difference reflects the accuracy of the initial network model in predicting the prediction transformation behavior of historically registered objects. Therefore, the model parameters of the initial network model can be optimized based on this difference, resulting in an object value prediction model. The optimization objective is to minimize the difference. Specifically, when the difference between the two satisfies the minimum difference value corresponding to the model training termination condition, optimization can be considered complete, and model training ends to obtain the object value prediction model.

[0078] Generally, a model loss function can be constructed, and then the optimization objective of minimizing the difference can be achieved by minimizing the model loss function. The model loss function can be... Let y represent the labeled conversion behavior. To represent the predicted conversion behavior, the optimization process can be represented by the following formula:

[0079]

[0080] Where W = {W1, W2, b1, b2}. Specifically, when the model loss function satisfies the minimum loss function value corresponding to the model training termination condition, the difference between the predicted conversion behavior and the labeled conversion behavior can be considered to have reached its minimum. At this point, optimization can be considered complete, and model training can be terminated to obtain the object value prediction model. The object value prediction model can be found in [reference needed]. Figure 3 As shown in Figure 102, it can be seen that the entire initial network model can be optimized by directly optimizing L during optimization.

[0081] Accordingly, in one possible implementation, S204 may include the following steps:

[0082] For each time node, a node loss function is constructed based on the difference between the predicted conversion behavior and the labeled conversion behavior at each time node;

[0083] The model loss function is obtained by weighted summation of the node loss functions at different time points;

[0084] The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

[0085] For details, please refer to Figure 3 As shown, for the b-th time node, the corresponding predicted transformation behavior can be used. With the labeled conversion behavior y bThe differences between them construct the node loss function:

[0086]

[0087] Where b is an integer between [1, k]. The node loss function can be the cross-entropy loss function, and correspondingly,

[0088] Therefore, the model loss function can be obtained by weighted summation of the node loss functions of each node:

[0089]

[0090] Among them, a b This represents the weight coefficient corresponding to the node loss function at the b-th time node; This represents the model loss function.

[0091] It should be noted that for a b The value of can be selected and set according to the actual needs of model training, and this application does not impose any limitations on it. For ease of understanding, the following examples are provided in the embodiments of this application:

[0092] In one possible implementation, to more quickly determine the model loss function based on the node loss function at each time point, the transformation behavior corresponding to each time point within the target time period can be treated equally. Specifically, the weight coefficients corresponding to the node loss function at each time point can be set to a fixed weight coefficient C, i.e., a b =C, thus the model loss function can be determined by summing the loss functions of the k nodes corresponding to the k time points. Based on this, once the node loss functions are determined, the model loss function can be determined relatively quickly, which facilitates the optimization of model parameters. It should be noted that the fixed weight coefficient C can be set according to the actual training situation of the model. This application does not impose any restrictions on this; for example, C can be set to 1.

[0093] In another possible implementation, since the target time period represents a period after registration, and within this target time period, the k time nodes have a temporal order, to characterize the differences in the conversion behavior of the object at different time nodes, different weight coefficients can be set for the node loss function of each time node. That is, when b takes any integer value in [1, k] in the above node loss function formula, its corresponding a... b They can be different. Specifically, weighting coefficients can be set based on the conversion behavior at each time point, i.e., a. b It can be based on b and y bThe model loss function is determined based on the node loss function and its corresponding weight coefficient at each time point, reflecting the differences in the transformation behavior of the object at different time points within the target time period.

[0094] After obtaining the aforementioned object value prediction model, its effectiveness can be verified through offline experiments. Specifically, taking the delivery of recommended content for a game application (a certain product) as an example, we collected historical registration data from the game's advertising and user acquisition scenarios over a week as the training dataset, and the data from the following two days as the test dataset. The task is to predict the payment amount (conversion behavior) of the target audience for each day within 7 days (i.e., the target time period is set to 7 days):

[0095] In the experiments, the feature extraction module used a three-layer fully connected neural network structure, the sampling module used random sampling, the temporal model used a transformer, and the baseline method used a traditional three-layer fully connected network structure. The Norm Gini coefficient was used as the performance metric on the training dataset, and the results are shown in Table 1 below.

[0096] Table 1

[0097] Measurement criteria Baseline This application Norm Gini 0.6834 0.7040

[0098] As can be seen, the object value prediction model provided in this application embodiment improves the Norm Gini index by 3.02% compared to the baseline method.

[0099] As can be seen from the above technical solution, by acquiring a training dataset, which includes historical object data of historical registered objects, and then using the feature extraction module in the initial network model to extract features from the historical object data to obtain historical object feature vectors, the time series module in the initial network model can output the predicted transformation behavior of historical registered objects based on the historical object feature vectors. Since historical object data has a corresponding set of calibrated conversion behaviors, which includes the calibrated conversion behaviors of historical registered objects at different time points within the target time period after registration, representing the true value of historical registered objects at different time points within the target time period after registration, and the predicted conversion behaviors represent the predicted value of historical registered objects at different time points within the target time period after registration, as determined by the initial network model, the model parameters of the initial network model are optimized based on the differences between the predicted conversion behaviors and the calibrated conversion behaviors at different time points within the target time period. This yields an object value prediction model, which can predict the object value at different time points within the target time period and output the conversion behaviors of the object to be predicted at different time points within the target time period. By arranging the conversion behaviors at different time points in chronological order, the object value of the object to be predicted can be obtained. The object value obtained in this way is a continuous value, which can reflect the continuous conversion potential of the object to be predicted within the target time period. It can more comprehensively and accurately depict the potential value of the object to be predicted. Therefore, based on this object value, it is beneficial to evaluate whether to deliver product recommendations to the object to be predicted to improve the ROI of the product.

[0100] Furthermore, related technologies employ delayed conversion models to predict the object's first conversion behavior within a certain period after the registration date, using this as the object's value. However, delayed conversion models can only output the object's conversion behavior from the first R days until the Rth day, meaning they can only predict the object's first conversion behavior within a certain period after registration. Therefore, the determined object value still fails to explicitly describe the object's potential value. Compared to the delayed conversion models in related technologies, the object value prediction model provided in this application can output the conversion behavior of the object to be predicted at different time points within the target time period. The resulting object value is a continuous value, explicitly reflecting the object's sustained conversion potential within the target time period.

[0101] It is understandable that an object's conversion behavior at the next time node is strongly dependent on its conversion behavior at the previous time node. For example, when the conversion behavior represents an object's payment behavior for a product, and the target time period is Δt days, dividing Δt days into k time nodes using days as the time unit, the object's payment behavior at the (k-1)th time node has a strong influence on whether it will still pay at the kth time node. Specifically, if the object made a payment at the (k-1)th time node, it is more likely to make a payment at the kth time node. Therefore, in one possible implementation, the target conversion behavior at the previous time node can be used as the input to the temporal module of the initial network model, so that the temporal module can output the predicted conversion behavior for the next time node based on the target conversion behavior and the historical object feature vector. For ease of understanding, this example still assumes that the target time period includes k time nodes, where k is an integer greater than 1. Accordingly, S203 may include the following steps:

[0102] S2031: For the first time point, based on the historical object feature vector, the predicted transformation behavior of the first time point is output through the time series module in the initial network model;

[0103] S2032: For the i-th time node, obtain the target transformation behavior of the (i-1)-th time node through the time sequence module, where i is an integer greater than 1 and less than or equal to k;

[0104] S2033: Based on the historical object feature vector and the target transformation behavior at the (i-1)th time node, output the predicted transformation behavior at the ith time node.

[0105] In other words, it is possible Figure 3 Based on this, the target transformation behavior of the previous time point is used as the input to the temporal module of the initial network model. For details, please refer to... Figure 4 For the first time point, the predicted transformation behavior for that time point can be output by the time series module 1012 in the initial network model based on the historical object feature vector. For the i-th time node, where i is an integer in the range (1, k), the target transformation behavior G for the (i-1)-th time node can be obtained first through the time sequence module 1012. i-1 Then, the time-series module 1012 uses the historical object feature vector and G... i-1 Output the predicted transformation behavior at the i-th time point. Understandably, starting from the second time point, repeating S2032 and S2033 will output the predicted transformation behavior of the historical registered object at the [2, k] time points.

[0106] Among them, G i-1This is used to represent the transformation behavior of historically registered objects at the previous time point, and is input into the time series module as a prediction. The auxiliary parameters enable the time-series module to comprehensively learn the transformation behavior of the previous time node based on the feature vectors of historical objects, and output based on this. Compared to outputting only the feature vectors of historical objects By incorporating the transformation behavior of historically registered objects at the previous time point, the deterministic nature of the data is enriched. The feature dimensions used are beneficial for improving the output. The accuracy.

[0107] For ease of understanding, the timing module output in this embodiment can be represented by the following formula. The process:

[0108]

[0109] Among them, G i-1 Let e(x) be the target transformation behavior at the (i-1)th time node, e(x) be the historical object feature vector, θ be the model parameters of the time series module, and F be any recurrent neural network (RNN).

[0110] Taking F as the RNN units as an example, the corresponding output of the timing module is... The process can be as follows Figure 5 As shown, e(x) and G i-1 Input timing module, through timing module according to e(x) and G i-1 Output

[0111] It should be noted that G i-1 Used to represent the transformation behavior of historically registered objects at the previous time point, for G i-1 The method for determining this is not limited in the embodiments of this application. For ease of understanding, the following method is provided as an example:

[0112] Example 1: Since predicted conversion behavior is a form that can reflect the conversion behavior characteristics of historically registered objects, the predicted conversion behavior at the (i-1)th time node can be directly represented by... The target transformation behavior G at the (i-1)th time node i-1 ,Right now In other words, the prediction behavior of the (i-1)th time node output by the time series module is transformed. The input is fed into the time series module so that the time series module can use the historical object feature vector and Determine the output

[0113] Example 2: Given It refers to the predictive transformation behavior output by the time-series module, especially in the early stages of model training when the initial network model's loss function has not yet converged. There may be deviations. To avoid the propagation of such deviations due to dependencies between time points, in one possible implementation, the initial network model may also include a sampling module, and further, before S2032, the following steps may be included:

[0114] The sampling module selects one of the predicted transformation behavior and the calibrated transformation behavior at the (i-1)th time node as the target transformation behavior.

[0115] In other words, it is possible Figure 4 Based on this, a sampling module is added to transform the prediction behavior from the (i-1)th time node. and the calibration transformation behavior y at the (i-1)th time node i-1 Select one as the target conversion behavior G. i-1 For details, please refer to Figure 6a The initial network model 101 may also include a sampling module 1013, based on the feature extraction module 1011 and the temporal module 1012.

[0116] Among them, the conversion behavior y is labeled. i-1 This represents the true value at the (i-1)th time point, used to predict conversion behavior. This represents the predicted value at the (i-1)th time node, from which one is selected by the sampling module as G. i-1 The input is placed into the time series module so that the time series module can perform operations based on the historical object feature vector and G. i-1 Determine the output Due to the G input to the timing module i-1 Capable of being used for time series module prediction The auxiliary parameters enrich the determination The feature dimensions utilized. In other words, the time-series module can comprehensively learn the transformation behavior features of the previous time point based on the feature vectors of historical objects, thereby providing the output of the time-series module. The process plays an optimization role; at the same time, due to G i-1 yes or y i-1 This makes G i-1 It possesses randomness, thus enabling it to maintain a balance throughout the entire model training phase. As G i-1 The effect of extending the training time of the model.

[0117] The above model training process can also be found in [reference needed]. Figure 6b As shown, the feature extraction module extracts features from historical object data to obtain the historical object feature vector e(x). Then, for the first time node, the time series module can output the historical feature vector. Starting from the second time node, the sampling module can first randomly select one of the predicted transformation behavior and the calibrated transformation behavior from the previous time node as the target transformation behavior for the previous time node, and input the target transformation behavior into the time series module. Then, the time series module uses the historical object feature vector and the target transformation behavior from the previous time node to output the predicted transformation behavior for the next time node. For example, for the second time node, the sampling module can be used to select the target transformation behavior from (y1, The system randomly selects one of the historical object feature vectors as the input to the time series module for the target transformation behavior at the first time node. Then, the time series module outputs the target transformation behavior based on the historical object feature vector and the target transformation behavior at the first time node. For the k-th time node, the sampling module can be used to sample from (y k-1 , The system randomly selects one of the historical object features as the input to the time series module for the target transformation behavior at the (k-1)th time node. Then, the time series module outputs the target transformation behavior based on the historical object feature vector and the target transformation behavior at the (k-1)th time node. Throughout the process, the aforementioned objective of minimizing the difference is achieved through target optimization. It is understood that this essentially corresponds to the method implementation embodiment, so relevant details can be found in the description of the method implementation embodiment. Furthermore, it should be noted that target optimization only exists during the training phase of the object value prediction model to achieve the objective of minimizing the difference; similarly, the sampling module only exists during the training phase of the object value prediction model.

[0118] In one possible implementation, the sampling module could be based on random sampling from... and y i-1 One of them is randomly selected as G. i-1 Specifically, it can be determined using the following formula:

[0119] p ~ Uniform(0,1)

[0120]

[0121] Here, p represents a Uniform(0,1) random variable uniformly distributed from 0 to 1. Based on this, in the actual model training process, for any previous time node (here, the (i-1)th time node is used to represent the previous time node), there is a corresponding calibration transformation behavior y. i-1 It also has its corresponding predictive transformation behavior. Then, the sampling module can randomly select one of them as the target transformation behavior G. i-1 The auxiliary time series module completes the prediction transformation behavior for the i-th time node. The output.

[0122] It is understandable that the transformation behavior of an object at the next time point is strongly dependent on the transformation behavior of that object at the previous time point. Furthermore, in the model training process provided in the above embodiments, the time-series module is based on the historical object feature vectors and the target transformation behavior G at the previous time point. i-1 Predictive transformation behavior for determining the next time point of output and With y i The difference between them can represent the node loss function corresponding to the i-th time node, which shows the target transformation behavior G of the previous time node. i-1 This will be passed to the next time node, thus affecting the node loss function corresponding to the next time node. During model training, the optimization of model parameters is carried out in the direction of minimizing the model loss function. The model loss function is determined based on the node loss function of each time node.

[0123] Therefore, to avoid errors caused by dependencies between time points, which could affect the convergence of the model's loss function to the minimum and ultimately prolong the model training time, one possible implementation is to divide the target time period into N consecutive time periods in chronological order, where N is an integer greater than 1. Then, in the process of training the initial network model to obtain the object value prediction model, the initial network model can be trained first using the training dataset corresponding to the first time period to obtain the first pre-trained model. Then, the first pre-trained model can be trained using the training dataset corresponding to the second time period to obtain the second pre-trained model, and so on, segment by segment. For the N-1 pre-trained model obtained by training the N-2 pre-trained model using the training dataset corresponding to the N-1 time period, the N-1 pre-trained model can be trained using the training dataset corresponding to the Nth time period to obtain the final object value prediction model.

[0124] It is understood that the value of N can be set according to the actual model training requirements, and this application does not impose any limitations on it. For ease of understanding, this application takes N=2 as an example. In this case, the target time period includes a first time period and a second time period, and the first time period is earlier than the second time period. The first time period includes m time nodes, and the second time period includes km time nodes. Accordingly, S204 may include the following steps:

[0125] Based on the differences between the predicted conversion behavior and the labeled conversion behavior at different time points within the first time period, the model parameters of the initial network model are optimized to obtain the pre-trained model.

[0126] Based on the feature vectors of historical objects, the predicted conversion behavior of historical registered objects at different time nodes in the second time period is output through the temporal module in the pre-trained model. The predicted conversion behavior at different time nodes in the second time period is arranged in chronological order.

[0127] Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the second time period, the model parameters of the pre-trained model are optimized to obtain the object value prediction model.

[0128] Specifically, firstly, the predicted conversion behaviors of the [1, m] time nodes included in the first time period are output using the initial network model. Based on the differences between the predicted conversion behaviors and the labeled conversion behaviors of each time node in the first time period, the model parameters of the initial network model are optimized, and the pre-trained model is obtained after optimization. Then, the predicted conversion behaviors of the [m+1, k] time nodes included in the second time period are output using the pre-trained model. Based on the differences between the predicted conversion behaviors and the labeled conversion behaviors of each time node in the second time period, the model parameters of the pre-trained model are optimized, and the object value prediction model is obtained after optimization.

[0129] It is understandable that the process of optimizing the model parameters of the initial network model based on the difference between the predicted transformation behavior and the calibrated transformation behavior at each time node in the first time period, and optimizing the model parameters of the pre-trained model based on the difference between the predicted transformation behavior and the calibrated transformation behavior at each time node in the second time period, can be referred to in the aforementioned process of optimizing model parameters based on the model loss function.

[0130] Furthermore, in related technologies that use a cumulative conversion model to predict the value (conversion behavior) of objects within a target time period after object registration, the training dataset for the cumulative conversion model can only be constructed using historical object data of registered objects registered before the target time period. Since the final prediction output is the cumulative conversion behavior within the target time period, object data of registered objects registered most recently within the target time period cannot be used. For easier understanding, see [link to relevant documentation]. Figure 7 Taking t as the registration date and Δt days as the target time period as an example, the training phase of the cumulative conversion model can only use... Figure 7Historical object data marked in black cannot use historical object data marked in white. This shows that this method cannot utilize the latest object data of the most recently registered objects, resulting in low data utilization. Furthermore, the latest object data, compared to historical object data from Δt days ago, more closely reflects the actual distribution of object data. Therefore, not only is data utilization low, but the predictive performance of the final model is also reduced.

[0131] In this embodiment of the application, the prediction of object value outputs the transformation behavior of each time node within the target time period. Based on this characteristic, for any given time node, the historical object data of historically registered objects registered before that time node can be used as training data. Specifically, in one possible implementation, S201 may include the following steps:

[0132] Obtain training data subsets for different window periods. Each window period is a time period before different time nodes within the target time period. The training data subsets include historical object data of historical registered objects registered before the corresponding time nodes.

[0133] The training dataset is constructed by combining subsets of training data from different window periods.

[0134] Specifically, for any time node within the target time period, the time period before that time node is divided into a window period. Each window period corresponds to a subset of training data. Each subset of training data includes the historical object data of historical registered objects registered before that time node. Then, the training dataset corresponding to the entire model training process is constructed based on the training data subsets of each window period.

[0135] Taking a target time period of Δt days and predicting the value (conversion behavior) of objects within Δt days as an example, Δt is divided into Δt time nodes using days as the time unit. Correspondingly, the training dataset D... train ={D1,D2…D n …D Δt}, where D1, D2…D n …,D Δt Let D represent the training data subsets corresponding to the first window period, ..., the nth window period, ..., the Δtth window period, respectively. Specifically, D n The following represents the historical object data corresponding to the historical registration object registered n days ago:

[0136] D n ={(x j y j )} n

[0137] Where n represents the nth window period, and n is an integer between [1, Δt], x j This represents the historical object data of historically registered object j. This represents the transformation behavior of historically registered object j on the k-th day after registration.

[0138] For ease of understanding, Δt is set to 7 days here, which is the predicted value (conversion behavior) of objects within 7 days after registration. Dividing this into 7 time nodes using days as the unit of time, the corresponding window period can be set to 7. See [link / reference] for details. Figure 7 As shown, the corresponding training dataset D train = {D1, D2, D3, ..., D7}. Thus, for the first window period, its corresponding training data subset can be constructed based on the historical object data of the historical registered objects registered before the first time point to construct the training data subset D1 for the first window period. For the second window period, its corresponding training data subset can be constructed based on the historical object data of the historical registered objects registered before the second time point to construct the training data subset D2 for the second window period. Similarly, the training data subset corresponding to each window period is constructed, and then the training data subsets of multiple window periods are used to form the training dataset for the entire model training.

[0139] Based on this, when predicting the value (conversion behavior) of an object within Δt days after registration, the latest object data of the most recently registered object within Δt days can be utilized. In other words, the object value prediction model provided in this application embodiment can utilize the latest object data of the most recently registered object, thereby improving data utilization. Furthermore, since the latest object data is closer to the actual distribution of object data, the object value prediction model trained in this way has a better prediction effect than the prediction model in related technologies.

[0140] Accordingly, for the historical object feature vector corresponding to each window period, the temporal module in the initial network model outputs the predicted conversion behavior of the historical registered object at different time nodes within the target time period. Then, for each window period, based on the difference between the predicted conversion behavior and the labeled conversion behavior at different time nodes within the target time period, a window period loss function is constructed, and the window period loss functions of different window periods are weighted and summed to obtain the model loss function. Finally, the model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

[0141] Specifically, for each window period, the window period loss function can be determined based on the difference between the predicted conversion behavior and the labeled conversion behavior at each time point within the target time period. For the nth window period, the window period loss function can be constructed using the following formula:

[0142]

[0143] in, Let n be the window period loss function. and Let c represent the calibrated conversion behavior and the predicted conversion behavior of the historical registered object j at the k-th time node within the target time period, respectively, where c is an integer between [1, n].

[0144] Furthermore, the model loss function can be obtained by weighted summing of the window period loss functions for each window period. The model loss function can be constructed using the following formula:

[0145]

[0146] Among them, a n This represents the weighting coefficients corresponding to the window loss function for the nth window period. It should be noted that 'a' varies depending on the window period. n The value of can be set according to the actual needs of model training, and this application does not impose any restrictions on it. For example, the weight coefficients corresponding to the window loss function for each window period can be set to the same weight coefficient; or, according to the window period order, different weight coefficients can be set for the window loss function of each window period.

[0147] In a campaign delivery scenario, the display location is used to showcase recommended content for the product being advertised. Generally, the display location can be the display area corresponding to the terminal device used by the user. This display area can be a specific area on the running interface of the application on these terminal devices. Similarly, in an advertising campaign for user acquisition, the display location can be an ad placement, where the recommended product content is displayed as an advertisement in the purchased ad placement to generate revenue. It should be noted that the terminal device used by the user can be a smartphone, tablet, computer, or other electronic device with content display capabilities. Correspondingly, the display area can be a specific area on the running interface of the application on these electronic devices. For ease of understanding, the following examples are provided in this application:

[0148] See Figure 8 The aforementioned object is a smartphone as the terminal device, and a social software X1 running on the smartphone is the aforementioned application. Accordingly, interface 801 is the running interface of the social software X1 when it is running, and the display area 802 on the running interface is the display position.

[0149] See Figure 9The aforementioned object uses a smartphone as the terminal device, and a short video software X2 running on the smartphone is used as the aforementioned application. Accordingly, interface 901 is the running interface of the short video software X2 when it is running, and the display area 902 on the running interface is the display position.

[0150] Generally, delivering product recommendations refers to placing recommended content on the display area of ​​the application interface of the application running on the user's terminal device. This allows the user to view the recommended content while using the application, learn about the product, and potentially convert. Accordingly, in actual delivery decisions, an object value prediction model can be used to support the decision-making process. Here, the object whose value needs to be predicted is designated as the object to be predicted, and the corresponding display position is designated as the display position to be delivered. The object value model then predicts the object value of the object to be predicted, and based on this object value, a decision is made on whether to purchase the display position for delivering the recommended content. Specifically, the following example illustrates this:

[0151] First, the object data of the object to be predicted corresponding to the display location can be obtained. Then, the feature extraction module in the object value prediction model can be used to extract features from the object data to obtain the object feature vector. Based on the object feature vector, the time series module in the object value prediction model can be used to output the conversion behavior of the object to be predicted at different time nodes within the target time period. The conversion behavior of the object to be predicted at different time nodes within the target time period is arranged in chronological order.

[0152] First, for any object to be predicted, the feature extraction module in the object value prediction model can extract features from the object data of the object to be predicted to obtain the object feature vector. Then, the time series module in the object value prediction model can output the transformation behavior of the object to be predicted at different time nodes within the target time period. By arranging the transformation behavior at different time nodes in chronological order, the continuous prediction value of the object to be predicted within the target time period can be obtained.

[0153] Since the object value prediction model provided in this application provides a continuous predicted value of the object to be predicted within a target time period, it can comprehensively and accurately depict the conversion behavior, conversion habits, and continuous conversion status of the object to be predicted within the target time period. Based on this, the continuous conversion potential of the object to be predicted can be evaluated, and then a decision can be made based on the evaluation results of the continuous conversion potential to determine whether it is necessary to purchase the corresponding display position for the object to be predicted, that is, whether to deliver product recommendation content to the object to be predicted.

[0154] Specifically, those with stable and consistent conversion potential can be designated as the product's target audience for promotion. Then, placement slots can be purchased for these target audiences to display recommended content, aiming to generate conversions based on the displayed content. Conversely, those without consistent conversion potential can be designated as potential promotion targets. Further, the decision on whether to distribute recommended content to these potential targets can be made based on actual circumstances. For example, if the number of potential promotion targets is insufficient to reach the planned campaign size, a subset with better conversion potential can be selected for content distribution.

[0155] As can be seen from the above technical solution, by training the initial network model, an object value prediction model is obtained. This object value prediction model can predict the object value at different time points within a target time period and output the conversion behavior of the object to be predicted at different time points within the target time period. By arranging the conversion behavior at different time points in chronological order, the object value of the object to be predicted can be obtained. Based on this, the object value obtained is a continuous value, which can reflect the continuous conversion potential of the object to be predicted within the target time period. It can more comprehensively and accurately depict the potential value of the object to be predicted. Based on this, objects with stable and continuous conversion potential can be selected as the objects to be promoted for the product. Then, product recommendation content can be delivered to the objects to be promoted, thereby optimizing and improving the product's ROI.

[0156] It should be noted that, based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods.

[0157] based on Figure 2 Corresponding to the training method for the object value prediction model provided in the embodiments, this application also provides a training device for the object value prediction model, the structural diagram of which is shown below. Figure 10 As shown, the device includes an acquisition unit 1001, an extraction unit 1002, an output unit 1003, and an optimization unit 1004.

[0158] The acquisition unit 1001 is used to acquire a training dataset, which includes historical object data of historical registered objects. The historical object data has a corresponding set of calibration conversion behaviors. The set of calibration conversion behaviors includes the calibration conversion behaviors of the historical registered objects at different time nodes within the target time period after registration is completed. The calibration conversion behaviors at different time nodes are arranged in chronological order.

[0159] The extraction unit 1002 is used to extract features from the historical object data through the feature extraction module in the initial network model to obtain the historical object feature vector;

[0160] The output unit 1003 is used to output the predicted transformation behavior of the historical registered object at different time nodes within the target time period based on the feature vector of the historical object and through the time sequence module in the initial network model. The predicted transformation behavior at different time nodes is arranged in the time order.

[0161] The optimization unit 1004 is used to optimize the model parameters of the initial network model based on the difference between the predicted transformation behavior and the calibrated transformation behavior at different time nodes within the target time period, so as to obtain the object value prediction model.

[0162] In one possible implementation, the target time period includes k time nodes, where k is an integer greater than 1, and the output unit is specifically used for:

[0163] For the first time point, based on the historical object feature vector, the predicted transformation behavior of the first time point is output through the time series module in the initial network model;

[0164] For the i-th time node, the target transformation behavior of the (i-1)-th time node is obtained through the time sequence module, where i is an integer greater than 1 and less than or equal to k;

[0165] Based on the historical object feature vector and the target transformation behavior at the (i-1)th time node, the predicted transformation behavior at the ith time node is output.

[0166] In one possible implementation, the initial network model includes a sampling module, and the device further includes a selection unit:

[0167] The selection unit is used to select one of the predicted transformation behavior and the calibrated transformation behavior at the (i-1)th time node as the target transformation behavior through the sampling module.

[0168] In one possible implementation, the target time period includes a first time period and a second time period, the first time period being earlier than the second time period, the first time period including m time nodes, and the second time period including km time nodes, wherein the optimization unit is specifically used for:

[0169] Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the first time period, the model parameters of the initial network model are optimized to obtain a pre-trained model.

[0170] Based on the feature vector of the historical object, the predicted conversion behavior of the historical registered object at different time nodes in the second time period is output through the time sequence module in the pre-trained model. The predicted conversion behavior at different time nodes in the second time period is arranged in the order of time.

[0171] Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the second time period, the model parameters of the pre-trained model are optimized to obtain the object value prediction model.

[0172] In one possible implementation, the optimization unit is specifically used for:

[0173] For each time node, a node loss function is constructed based on the difference between the predicted conversion behavior and the labeled conversion behavior at each time node;

[0174] The model loss function is obtained by weighted summation of the node loss functions at different time points;

[0175] The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

[0176] In one possible implementation, the acquisition unit is specifically used for:

[0177] Obtain training data subsets for different window periods, each window period being a time period before different time nodes within the target time period, and the training data subsets include historical object data of historical registered objects registered before the corresponding time nodes;

[0178] The training dataset is composed of subsets of training data from different window periods.

[0179] In one possible implementation, the output unit is specifically used for:

[0180] For each window period, the predicted transformation behavior of the historical registered object at different time nodes within the target time period is output through the time series module in the initial network model.

[0181] The optimization unit is specifically used for:

[0182] For each window period, a window period loss function is constructed based on the difference between the predicted conversion behavior and the labeled conversion behavior at different time points within the target time period.

[0183] The model loss function is obtained by weighted summation of the window period loss functions for different window periods;

[0184] The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

[0185] In one possible implementation, the acquisition unit is further configured to acquire object data of the object to be predicted corresponding to the display location to be deployed;

[0186] The extraction unit is also used to extract features from the object data through the feature extraction module in the object value prediction model to obtain an object feature vector;

[0187] The output unit is further configured to output the transformation behavior of the object to be predicted at different time nodes within the target time period through the time series module in the object value prediction model, based on the object feature vector, wherein the transformation behavior of the object to be predicted at different time nodes within the target time period is arranged in the time order.

[0188] As can be seen from the above technical solution, by acquiring a training dataset, which includes historical object data of historical registered objects, and then using the feature extraction module in the initial network model to extract features from the historical object data to obtain historical object feature vectors, the time series module in the initial network model can output the predicted transformation behavior of historical registered objects based on the historical object feature vectors. Since historical object data has a corresponding set of calibrated conversion behaviors, which includes the calibrated conversion behaviors of historical registered objects at different time points within the target time period after registration, representing the true value of historical registered objects at different time points within the target time period after registration, and the predicted conversion behaviors represent the predicted value of historical registered objects at different time points within the target time period after registration, as determined by the initial network model, the model parameters of the initial network model are optimized based on the differences between the predicted conversion behaviors and the calibrated conversion behaviors at different time points within the target time period. This yields an object value prediction model, which can predict the object value at different time points within the target time period and output the conversion behaviors of the object to be predicted at different time points within the target time period. By arranging the conversion behaviors at different time points in chronological order, the object value of the object to be predicted can be obtained. The object value obtained in this way is a continuous value, which can reflect the continuous conversion potential of the object to be predicted within the target time period. It can more comprehensively and accurately depict the potential value of the object to be predicted. Therefore, based on this object value, it is beneficial to evaluate whether to deliver product recommendations to the object to be predicted to improve the ROI of the product.

[0189] This application also provides a computer device, which can be a terminal device, taking a smartphone as an example:

[0190] Figure 11 The diagram shown is a block diagram of a portion of the structure of a smartphone provided in an embodiment of this application. (Reference) Figure 11 The smartphone includes components such as a radio frequency (RF) circuit 1110, a memory 1120, an input unit 1130, a display unit 1140, a sensor 1150, an audio circuit 1160, a Wi-Fi module 1170, a processor 1180, and a power supply 1190. The input unit 1130 may include a touch panel 1131 and other input devices 1132, the display unit 1140 may include a display panel 1141, and the audio circuit 1160 may include a speaker 1161 and a microphone 1162. Those skilled in the art will understand that... Figure 11 The smartphone structure shown does not constitute a limitation on smartphones and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0191] The memory 1120 can be used to store software programs and modules. The processor 1180 executes various functions and data processing of the smartphone by running the software programs and modules stored in the memory 1120. The memory 1120 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the smartphone (such as audio data, phonebook, etc.). In addition, the memory 1120 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0192] The processor 1180 is the control center of the smartphone, connecting various parts of the smartphone via various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 1120 and calling data stored in the memory 1120. Optionally, the processor 1180 may include one or more processing units; preferably, the processor 1180 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 1180.

[0193] In this embodiment, the processor 1180 in the smartphone can perform the following steps:

[0194] Obtain a training dataset, which includes historical object data of historical registered objects. The historical object data has a corresponding set of calibration conversion behaviors. The set of calibration conversion behaviors includes the calibration conversion behaviors of the historical registered objects at different time points within the target time period after registration is completed. The calibration conversion behaviors at different time points are arranged in chronological order.

[0195] The feature extraction module in the initial network model extracts features from the historical object data to obtain the historical object feature vector.

[0196] Based on the feature vector of the historical object, the time series module in the initial network model outputs the predicted transformation behavior of the historical registered object at different time nodes within the target time period, and the predicted transformation behavior at different time nodes is arranged in the time order.

[0197] Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period, the model parameters of the initial network model are optimized to obtain the object value prediction model.

[0198] The computer device provided in this application embodiment can also be a server. Please refer to [link / reference]. Figure 12 As shown, Figure 12 This is a structural diagram of the server 1200 provided in this application embodiment. The server 1200 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 1222 and memory 1232, and one or more storage media 1230 (e.g., one or more mass storage devices) for storing application programs 1242 or data 1244. The memory 1232 and storage media 1230 can be temporary or persistent storage. The program stored in the storage media 1230 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the server. Furthermore, the CPU 1222 may be configured to communicate with the storage media 1230 and execute the series of instruction operations in the storage media 1230 on the server 1200.

[0199] Server 1200 may also include one or more power supplies 1226, one or more wired or wireless network interfaces 1250, one or more input / output interfaces 1258, and / or one or more operating systems 1241, such as Windows Server. TM Mac OS X TM Unix TM Linux TMFreeBSD TM etc.

[0200] In this embodiment, the steps executed by the central processing unit 1222 in the server 1200 can be based on Figure 12 The structure shown is implemented.

[0201] According to one aspect of this application, a computer-readable storage medium is provided for storing a computer program for implementing the training method of the object value prediction model described in the foregoing embodiments.

[0202] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in various optional implementations of the above embodiments.

[0203] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0204] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0205] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0206] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0207] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0208] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0209] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A training method for an object value prediction model, characterized in that, The method includes: Obtain a training dataset, which includes historical object data of historical registered objects. The historical object data has a corresponding set of calibration conversion behaviors. The set of calibration conversion behaviors includes the calibration conversion behaviors of the historical registered objects at different time points within the target time period after registration is completed. The calibration conversion behaviors at different time points are arranged in chronological order. The feature extraction module in the initial network model extracts features from the historical object data to obtain the historical object feature vector. Based on the feature vector of the historical object, the time series module in the initial network model outputs the predicted transformation behavior of the historical registered object at different time nodes within the target time period, and the predicted transformation behavior at different time nodes is arranged in the time order. Based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period, the model parameters of the initial network model are optimized to obtain the object value prediction model. The object value prediction model is used to predict the conversion behavior of the object to be predicted at different time points within the target time period. The conversion behavior is the paid behavior generated by the object based on recommended content, and the object value is the conversion behavior generated by the object through recommended content placed in the ad slot.

2. The method according to claim 1, characterized in that, The target time period includes k time nodes, where k is an integer greater than 1. Based on the historical object feature vector, the predicted transformation behavior of the historical registered object at different time nodes within the target time period is output through the time-series module in the initial network model, including: For the first time point, based on the historical object feature vector, the predicted transformation behavior of the first time point is output through the time series module in the initial network model; For the i-th time node, the target transformation behavior of the (i-1)-th time node is obtained through the time sequence module, where i is an integer greater than 1 and less than or equal to k; Based on the historical object feature vector and the target transformation behavior at the (i-1)th time node, the predicted transformation behavior at the ith time node is output.

3. The method according to claim 2, characterized in that, The initial network model includes a sampling module. Before obtaining the target transformation behavior at the (i-1)th time node through the time series module, the method further includes: The sampling module selects one of the predicted transformation behavior and the calibrated transformation behavior at the (i-1)th time node as the target transformation behavior.

4. The method according to claim 2, characterized in that, The target time period includes a first time period and a second time period, where the first time period is earlier than the second time period. The first time period includes m time nodes, and the second time period includes km time nodes. The model parameters of the initial network model are optimized based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time nodes within the target time period to obtain the object value prediction model, including: Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the first time period, the model parameters of the initial network model are optimized to obtain a pre-trained model. Based on the feature vector of the historical object, the predicted conversion behavior of the historical registered object at different time nodes in the second time period is output through the time sequence module in the pre-trained model. The predicted conversion behavior at different time nodes in the second time period is arranged in the order of time. Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the second time period, the model parameters of the pre-trained model are optimized to obtain the object value prediction model.

5. The method according to claim 1, characterized in that, The method of optimizing the model parameters of the initial network model based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period to obtain the object value prediction model includes: For each time node, a node loss function is constructed based on the difference between the predicted conversion behavior and the labeled conversion behavior at each time node; The model loss function is obtained by weighted summation of the node loss functions at different time points; The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

6. The method according to claim 1, characterized in that, The acquisition of the training dataset includes: Obtain training data subsets for different window periods, each window period being a time period before different time nodes within the target time period, and the training data subsets include historical object data of historical registered objects registered before the corresponding time nodes; The training dataset is composed of subsets of training data from different window periods.

7. The method according to claim 6, characterized in that, The step of outputting the predicted conversion behavior of the historical registered object at different time nodes within the target time period based on the historical object feature vector through the time-series module in the initial network model in chronological order includes: For each window period corresponding to the historical object feature vector, the time series module in the initial network model outputs the predicted transformation behavior of the historical registered object at different time nodes within the target time period; The method of optimizing the model parameters of the initial network model based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period to obtain the object value prediction model includes: For each window period, a window period loss function is constructed based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period; The model loss function is obtained by weighted summation of the window period loss functions for different window periods; The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

8. The method according to any one of claims 1-7, characterized in that, The method further includes: Obtain the object data of the object to be predicted corresponding to the display location to be displayed; The object feature vector is obtained by extracting features from the object data through the feature extraction module in the object value prediction model. Based on the object feature vector, the time series module in the object value prediction model outputs the transformation behavior of the object to be predicted at different time nodes within the target time period, and the transformation behavior of the object to be predicted at different time nodes within the target time period is arranged in the time order.

9. A training device for an object value prediction model, characterized in that, The device includes an acquisition unit, an extraction unit, an output unit, and an optimization unit: The acquisition unit is used to acquire a training dataset, which includes historical object data of historical registered objects. The historical object data has a corresponding set of calibration conversion behaviors. The set of calibration conversion behaviors includes the calibration conversion behaviors of the historical registered objects at different time nodes within the target time period after registration is completed. The calibration conversion behaviors at different time nodes are arranged in chronological order. The extraction unit is used to extract features from the historical object data through the feature extraction module in the initial network model to obtain the historical object feature vector; The output unit is used to output the predicted transformation behavior of the historical registered object at different time nodes within the target time period based on the feature vector of the historical object and through the time sequence module in the initial network model. The predicted transformation behavior at different time nodes is arranged in the time order. The optimization unit is used to optimize the model parameters of the initial network model based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period, to obtain the object value prediction model, so as to predict the conversion behavior of the object to be predicted at different time points within the target time period through the object value prediction model. The conversion behavior is the paid behavior generated by the object based on the recommended content, and the object value is the conversion behavior generated by the object through the recommended content placed in the advertising slot.

10. The apparatus according to claim 9, characterized in that, The target time period includes k time nodes, where k is an integer greater than 1. The output unit is specifically used for: For the first time point, based on the historical object feature vector, the predicted transformation behavior of the first time point is output through the time series module in the initial network model; For the i-th time node, the target transformation behavior of the (i-1)-th time node is obtained through the time sequence module, where i is an integer greater than 1 and less than or equal to k; Based on the historical object feature vector and the target transformation behavior at the (i-1)th time node, the predicted transformation behavior at the ith time node is output.

11. The apparatus according to claim 10, characterized in that, The initial network model includes a sampling module, and the device further includes a selection unit; The selection unit is used to select one of the predicted transformation behavior and the calibrated transformation behavior at the (i-1)th time node as the target transformation behavior through the sampling module.

12. The apparatus according to claim 10, characterized in that, The target time period includes a first time period and a second time period, wherein the first time period is earlier than the second time period, the first time period includes m time nodes, and the second time period includes km time nodes. The optimization unit is specifically used for: Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the first time period, the model parameters of the initial network model are optimized to obtain a pre-trained model. Based on the feature vector of the historical object, the predicted conversion behavior of the historical registered object at different time nodes in the second time period is output through the time sequence module in the pre-trained model. The predicted conversion behavior at different time nodes in the second time period is arranged in the order of time. Based on the differences between the predicted conversion behavior and the calibrated conversion behavior at different time points within the second time period, the model parameters of the pre-trained model are optimized to obtain the object value prediction model.

13. The apparatus according to claim 9, characterized in that, The optimization unit is specifically used for: For each time node, a node loss function is constructed based on the difference between the predicted conversion behavior and the labeled conversion behavior at each time node; The model loss function is obtained by weighted summation of the node loss functions at different time points; The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

14. The apparatus according to claim 9, characterized in that, The acquisition unit is specifically used for: Obtain training data subsets for different window periods, each window period being a time period before different time nodes within the target time period, and the training data subsets include historical object data of historical registered objects registered before the corresponding time nodes; The training dataset is composed of subsets of training data from different window periods.

15. The apparatus according to claim 14, characterized in that, The output unit is specifically used for: For each window period corresponding to the historical object feature vector, the time series module in the initial network model outputs the predicted transformation behavior of the historical registered object at different time nodes within the target time period; The optimization unit is specifically used for: For each window period, a window period loss function is constructed based on the difference between the predicted conversion behavior and the calibrated conversion behavior at different time points within the target time period; The model loss function is obtained by weighted summation of the window period loss functions for different window periods; The model parameters of the initial network model are optimized based on the model loss function to obtain the object value prediction model.

16. The apparatus according to any one of claims 9-15, characterized in that, The acquisition unit is also used to acquire object data of the object to be predicted corresponding to the display location to be displayed; The extraction unit is also used to extract features from the object data through the feature extraction module in the object value prediction model to obtain an object feature vector; The output unit is further configured to output the transformation behavior of the object to be predicted at different time nodes within the target time period through the time series module in the object value prediction model, based on the object feature vector, wherein the transformation behavior of the object to be predicted at different time nodes within the target time period is arranged in the time order.

17. A computer device, characterized in that, The computer device includes a processor and memory: The memory is used to store computer programs and to transfer the computer programs to the processor; The processor is configured to execute the method according to any one of claims 1-8 according to instructions in the computer program.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for implementing the method according to any one of claims 1-8.

19. A computer program product, comprising a computer program, characterized in that, When it is run on a computer device, it causes the computer device to perform the method according to any one of claims 1-8.

Images