Training method and device of target prediction model, equipment and storage medium

Abstract

The application discloses a target prediction model training method and device, equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining an initial prediction result through an initial target prediction network; obtaining word embedding of a consumer object and word embedding of a consumed object through a feature pre-training network; obtaining an adjustment factor for fitting the difference between the initial prediction result and statistical results corresponding to the training samples based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, industry domain knowledge corresponding to the training samples and the initial prediction result through a target calibration network; and calibrating the initial prediction result through the adjustment factor to obtain a calibrated prediction result. The application directly calibrates the initial prediction result based on the adjustment factor learned based on the industry domain knowledge, reduces the difference between the calibrated prediction result and the statistical result, and improves the prediction accuracy of the target (such as the conversion rate of an advertisement).

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for training a target prediction model. Background Technology

[0002] Conversion rate prediction is a crucial task in the advertising industry, with significant applications in achieving precise ad targeting and estimating the value of ad placements. Conversion rate, in particular, refers to the probability that a consumer will convert an ad based on its message.

[0003] Currently, artificial intelligence technology has begun to be applied to advertising conversion rate prediction. In this technology, deep learning models are used to build conversion rate prediction models, which are then trained with a large number of training samples (such as consumer characteristics and advertising features) to obtain a conversion rate prediction model that can be used to predict conversion rates.

[0004] However, since deep models are relatively black boxes, there will inevitably be a certain gap between the conversion behavior predicted by the trained deep model and the actual conversion behavior, resulting in insufficient accuracy in conversion rate prediction. Summary of the Invention

[0005] This application provides a training method, apparatus, device, and storage medium for a target prediction model, which can reduce the difference between the prediction results and statistical results of a target (such as conversion rate) and improve the prediction accuracy of the target. The technical solution may include the following.

[0006] According to one aspect of the embodiments of this application, a method for training a target prediction model is provided, the target prediction model including an initial target prediction network, a feature pre-training network, and a target calibration network; the method includes:

[0007] Obtain training samples for the target prediction model, wherein the training samples include consumer object features, consumed object features, and contextual features of the consumed object;

[0008] The initial prediction network obtains an initial prediction result based on the training samples, and the initial prediction result is used to characterize the probability that the consumed object will achieve the goal.

[0009] The feature pre-training network obtains the word embeddings of the consumer object and the consumer object based on the features of the consumer object and the features of the consumed object; wherein, the word embeddings of the consumer object learn the potential association between the features of the consumer object and the target intent of the consumer object, and the word embeddings of the consumer object learn the potential association between the features of the consumed object and the target intent of the consumer object;

[0010] The target calibration network obtains an adjustment factor based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, the industry domain knowledge corresponding to the training samples, and the initial prediction results. The adjustment factor is used to fit the difference between the initial prediction results and the statistical results corresponding to the training samples.

[0011] The initial prediction result is calibrated using the adjustment factor to obtain the calibrated prediction result;

[0012] Based on the calibrated prediction results and the statistical results, the target prediction model is trained to obtain the trained target prediction model.

[0013] According to one aspect of the embodiments of this application, a training apparatus for a target prediction model is provided, the target prediction model including an initial target prediction network, a feature pre-training network, and a target calibration network; the apparatus includes:

[0014] The training sample acquisition module is used to acquire training samples of the target prediction model, wherein the training samples include consumer object features, consumed object features and context features of the consumed object;

[0015] The prediction result acquisition module is used to obtain an initial prediction result based on the training samples through the initial target prediction network. The initial prediction result is used to characterize the probability that the consumed object will achieve the target.

[0016] The feature pre-training module is used to obtain the word embeddings of the consumer object and the consumer object based on the features of the consumer object and the features of the consumed object through the feature pre-training network; wherein the word embeddings of the consumer object learn a potential association between the features of the consumer object and the target intent of the consumer object, and the word embeddings of the consumer object learn a potential association between the features of the consumed object and the target intent of the consumer object;

[0017] The adjustment factor acquisition module is used to obtain an adjustment factor through the target calibration network based on the word embedding of the consumer object, the word embedding of the consumed object, the training sample, the industry domain knowledge corresponding to the training sample, and the initial prediction result. The adjustment factor is used to fit the difference between the initial prediction result and the statistical result corresponding to the training sample.

[0018] The prediction result calibration module is used to calibrate the initial prediction result using the adjustment factor to obtain the calibrated prediction result;

[0019] The prediction model training module is used to train the target prediction model based on the calibrated prediction results and the statistical results, so as to obtain the trained target prediction model.

[0020] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the training method of the target prediction model described above.

[0021] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the storage medium, the computer program being loaded and executed by a processor to implement the training method of the above-described target prediction model.

[0022] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, which includes 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 training method for the target prediction model described above.

[0023] The technical solutions provided in this application include at least the following beneficial effects.

[0024] By learning industry-specific knowledge from the training samples and fitting an adjustment factor that accounts for the difference between the initial prediction and statistical results, the initial predictions obtained by the initial target prediction network based on the training samples are calibrated, resulting in calibrated predictions. This direct calibration of the initial predictions through adjustment factors reduces the impact of the inevitable discrepancy between the implementation process fitted by the initial target prediction network and the actual implementation process. This effectively reduces the difference between the calibrated predictions and statistical results, thereby improving the accuracy of target predictions (such as conversion rates, click-through rates, and exposure rates). Furthermore, since the calibration is based directly on industry-specific knowledge, the accuracy of target predictions can be further enhanced.

[0025] Furthermore, by learning the word embeddings of consumer objects that have potential associations between consumer object characteristics and consumer object target intentions (such as purchasing, subscribing, browsing, etc.), and learning the word embeddings of consumed objects that have potential associations between consumed object characteristics and consumer object target intentions, the initial prediction results can be calibrated. This can provide the correct optimization direction for the calibration process of the initial prediction results and help to better leverage the role of industry domain knowledge, thereby effectively improving the calibration effect of the initial prediction results and further improving the prediction accuracy of the target. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0027] Figure 1 This is a schematic diagram of the implementation environment of a solution provided in one embodiment of this application;

[0028] Figure 2 This is a schematic diagram of a feature pre-trained network provided in one embodiment of this application;

[0029] Figure 3 This is a schematic diagram of a target calibration network provided in one embodiment of this application;

[0030] Figure 4 This is a flowchart of a training method for a target prediction model provided in one embodiment of this application;

[0031] Figure 5 This is a schematic diagram of a target prediction model provided in one embodiment of this application;

[0032] Figure 6 This is a flowchart of a method for obtaining calibrated prediction results according to an embodiment of this application;

[0033] Figures 7 to 10 This is a schematic diagram illustrating the training effect of the target prediction model provided in the embodiments of this application;

[0034] Figure 11 This is a block diagram of a training apparatus for a target prediction model provided in one embodiment of this application;

[0035] Figure 12 This is a block diagram of a training apparatus for a target prediction model provided in another embodiment of this application;

[0036] Figure 13 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0038] Artificial intelligence (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.

[0039] 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, and machine learning / deep learning.

[0040] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. 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 learn-by-doing.

[0041] The technical solutions provided in this application involve machine learning technology in artificial intelligence. Machine learning technology is used to train a target prediction model (such as an initial target prediction network, a feature pre-training network, a target calibration network, etc.), and then the target prediction is achieved through the trained target prediction model (that is, the calibration of the initial prediction result).

[0042] The method provided in this application can be executed by a computer device, which refers to an electronic device with data computing, processing, and storage capabilities. This computer device can be a terminal such as a PC (Personal Computer), tablet computer, smartphone, wearable device, intelligent robot, or vehicle-mounted device; it can also be a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0043] The technical solutions provided in this application are applicable to any target prediction scenario, such as advertising placement, advertising value estimation, conversion rate prediction, product recommendation, news recommendation, and video placement. The technical solutions provided in this application can improve the accuracy of target prediction. The training method of the target prediction model provided in this application will be described in detail below.

[0044] Please refer to Figure 1 The diagram illustrates an implementation environment for a solution provided in one embodiment of this application. This implementation environment may include a model training device 10 and a model usage device 20.

[0045] The model training device 10 can be an electronic device such as a PC, computer, tablet computer, server, intelligent robot, vehicle terminal, or other electronic devices with strong computing power. The model training device 10 is used to train the target prediction model 30.

[0046] In this embodiment, the target prediction model 30 is a neural network model used to predict targets. For example, in an advertising scenario, the target prediction model 30 can be used to predict at least one of the following: conversion rate (number of conversions / number of clicks), click-through rate (number of clicks / number of impressions), click-to-conversion rate (number of conversions / number of impressions), and duration (e.g., browsing time). For instance, the target prediction model 30 can predict only the conversion rate, or it can predict both the conversion rate and the click-through rate simultaneously; this embodiment does not limit this. In a product placement scenario, the target prediction model 30 can predict at least one of the following: conversion rate, click-through rate, click-to-conversion rate, and duration, for products such as games, items, food, and equipment. In a news placement scenario, the target prediction model 30 can predict at least one of the following: conversion rate, click-through rate, click-to-conversion rate, and duration, for news, articles, and notifications; this embodiment does not limit the industries to which the target prediction model 30 can be applied.

[0047] Optionally, the model training device 10 can use machine learning to train the target prediction model 30 to achieve better performance.

[0048] The trained target prediction model 30 can be deployed on the model-using device 20 to provide target prediction services. The model-using device 20 can be a terminal device such as a mobile phone, computer, smart TV, multimedia playback device, wearable device, medical device, or a server; this application does not limit this.

[0049] In some embodiments, such as Figure 1As shown, the target prediction model 30 may include an initial target prediction network 31, a feature pre-training network 32, and a target calibration network 33.

[0050] The initial target prediction network 31 can be a neural network, which can be used to predict targets and obtain initial prediction results corresponding to the targets. These initial prediction results characterize the probability that the consumed object will achieve the target. For example, when the initial target prediction network 31 is used to predict conversion rates, the initial prediction result can refer to the predicted probability value of the conversion rate corresponding to the consumed object, used to characterize the probability that the consumer object will generate conversion behavior towards the consumed object. Optionally, the range of the predicted probability value can be [0, 1]. The consumer object can refer to the user, consumer, or browser of the consumed object, etc., and this embodiment does not limit this. For example, when the consumed object is an advertisement, the consumer object can refer to the browser corresponding to the advertisement.

[0051] In this embodiment, the initial target prediction network 31 can refer to a trained target prediction network, meaning that the initial target prediction network 31 already possesses a certain target prediction capability. Optionally, the initial target prediction network 31 can also be a multi-task prediction neural network, which can be used to predict multiple targets simultaneously, meaning the initial prediction result can include the prediction results corresponding to each of the multiple targets. For example, the initial target prediction network 31 can be used to predict conversion rate, click-through rate, and duration simultaneously, then the initial prediction result can include the prediction results corresponding to conversion rate, click-through rate, and duration respectively. The initial target prediction network 31 can also be used to predict the conversion rate of different tasks, which is not limited in this embodiment.

[0052] The feature pre-training network 32 can be a neural network that can be used to predict the target intent of a consumer, obtaining a prediction result corresponding to the target intent. This prediction result is used to characterize the likelihood of the consumer achieving its target intent. The target intent can include purchasing, viewing, browsing, clicking, forwarding, etc. In conversion rate prediction scenarios, since purchasing accounts for a large proportion, purchasing can be set as the target intent.

[0053] In this embodiment, the feature pre-training network 32 can refer to a trained neural network, meaning that the feature pre-training network 32 already possesses a certain target intent prediction capability. Optionally, during the training process of the target prediction network, it is necessary to obtain the word embeddings corresponding to the training samples obtained through the feature pre-training network 32, and it is not necessary to obtain the prediction results corresponding to the target intent.

[0054] Compared to the sparse features corresponding to the training samples, the word embeddings corresponding to the training samples have relatively dense features, which is beneficial for the target prediction model 30 to fit the target. At the same time, the word embeddings corresponding to the training samples learn the target intent of the consumer, which can provide an optimized training direction for the target prediction model 30, thereby improving the training effect of the target prediction model 30.

[0055] In one example, refer to Figure 2 The feature pre-training network 32 can be a multi-layer neural network, which may include an input layer, an embedding layer, a hidden layer and an output layer.

[0056] The input layer of the feature pre-training network 32 is used to acquire training samples, the embedding layer of the feature pre-training network 32 is used to acquire word embeddings corresponding to the training samples, the hidden layers (which can be multiple) of the feature pre-training network 32 are used to learn features (such as densification) based on the word embeddings corresponding to the training samples, and the output layer of the feature pre-training network 32 (such as the sigmoid function) is used to predict the prediction result corresponding to the target intent based on the learned features. The output of the embedding layer of the feature pre-training network 32 (i.e., the word embeddings corresponding to the training samples) is used as the input of the target calibration network 33.

[0057] The target calibration network 33 is also a neural network. It can be used to obtain an adjustment factor that fits the difference between the initial prediction result and the statistical result corresponding to the training samples. The initial prediction result is then calibrated using this adjustment factor to obtain a calibrated prediction result, the accuracy of which is higher than that of the initial prediction result. In this embodiment, the target calibration network 33 is an incompletely trained neural network. Its network parameters need to be adjusted based on the training samples, the word embeddings corresponding to the training samples, and the initial prediction result to complete the training of the target calibration network 33.

[0058] In one example, refer to Figure 3 The target calibration network 33 may include an input layer ( Figure 3 (Not shown in the image), embedding layer, domain factorization layer, task layer, and output layer.

[0059] The input layer of the target calibration network 33 is used to obtain training samples and the word embeddings corresponding to the training samples. The embedding layer of the target calibration network 33 is used to obtain the optimized word embeddings corresponding to the training samples. The domain factor decomposition layer of the target calibration network 33 is used to obtain the latent vector representations corresponding to the training samples based on the optimized word embeddings corresponding to the training samples. The task layer of the target calibration network 33 is used to obtain the output factors corresponding to the task based on the latent vector representations. The output layer of the target calibration network 33 is used to obtain the adjustment factors based on the output factors and the initial prediction results, and to calibrate the initial prediction results through the adjustment factors to obtain the calibrated prediction results.

[0060] The domain factorization layer can adopt a FwFFM (Field-Weighted Field-Ware Factorization Machine) structure, which can fully leverage the cross-feature interactions to learn more domain knowledge. Latent vector representations can be used to characterize the cross-features between the features corresponding to the training samples, i.e., the influence of interactions between features. The output factor refers to the prediction result corresponding to the target calibration network, such as the predicted probability value for the task. Optionally, the task layer can be a single-task or multi-task structure, and the number of tasks corresponding to this task layer corresponds to the number of targets corresponding to the initial target prediction network 31. For example, if the initial target prediction network 31 is used to predict conversion rate and click-through rate, then the output of the task layer of the target calibration network 33 corresponds to the predicted conversion rate and click-through rate (i.e., the output factor).

[0061] The technical solution of this application will be described below through method embodiments.

[0062] Please refer to Figure 4 The diagram illustrates a flowchart of a training method for a target prediction model provided in one embodiment of this application. The execution entity for each step of this method can be the model training device described above. The method may include the following steps (401-406).

[0063] Step 401: Obtain training samples for the target prediction model. The training samples include consumer object features, consumed object features, and contextual features of the consumed object.

[0064] Among them, the characteristics of the consumer object are used to characterize the consumer object. For example, the characteristics of the consumer object may include the basic attribute characteristics of the consumer object, such as name, age, gender, address, behavioral data, interests, etc. The characteristics of the consumed object are used to characterize the consumed object. For example, the characteristics of the consumed object may include the basic attribute characteristics of the consumed object (such as name, version, content, etc.) and statistical characteristics (such as statistical conversion rate, statistical click-through rate, etc.). The contextual characteristics of the consumed object are used to characterize the relationship between the consumed object and the application environment, such as the environment, scene, industry, and related things corresponding to the consumed object. For example, taking an advertisement as an example, the contextual characteristics of the advertisement can be used to characterize the environment, scene, industry, etc., corresponding to the advertisement.

[0065] Optionally, the target prediction model can be used to predict different targets in different industries, such as advertising, games, goods, news, e-commerce, etc. These different targets may include conversion rate, click-through rate, click-to-conversion rate, duration, etc., which are not limited in this embodiment. The target prediction model, consumer, and consumed object in this embodiment are the same as those described in the above embodiments. Content not described in this embodiment can be referred to the above embodiments, and will not be repeated here.

[0066] Optionally, the characteristics of the consumer object, the characteristics of the consumed object, and the contextual characteristics of the consumed object can refer to the encoded data (such as one-hot encoding).

[0067] Step 402: Based on the training samples, the initial target prediction network obtains an initial prediction result, which is used to characterize the probability of the consumed object achieving the target.

[0068] Optionally, the characteristics of the consumer object, the characteristics of the consumed object, and the contextual characteristics of the consumed object can be input into the initial target prediction network to obtain the initial prediction result.

[0069] For example, refer to Figure 5 The target prediction model 500 is primarily used to predict conversion rates. The initial target prediction network 503 within the target prediction model 500 is used to obtain initial prediction results corresponding to the training samples. These initial prediction results mainly include the predicted probability value of the initial conversion rate (for example, the initial prediction result may also include the predicted probability value of the initial click-through rate, the predicted probability value of the initial exposure rate, etc.). For example, by inputting the consumer's features, the features of the consumed object, and the contextual features of the consumed object into the initial target prediction network 503, the initial prediction result can be obtained, and this initial prediction result can be provided to the output layer of the target calibration network 501.

[0070] Step 403: Based on the features of the consumer object and the features of the consumed object, the word embeddings of the consumer object and the consumed object are obtained through the feature pre-training network; wherein, the word embedding of the consumer object learns the potential association between the features of the consumer object and the target intent of the consumer object, and the word embedding of the consumed object learns the potential association between the features of the consumed object and the target intent of the consumer object.

[0071] In this embodiment, a feature pre-trained network is used to predict the target intent of a consumer, which may include at least one of purchase, view, browse, click, forward, etc. The word embeddings obtained by the feature pre-trained network include information about the target intent of the consumer.

[0072] For example, refer to Figure 5 The features of the consumer object and the consumed object are input into the feature pre-training network 502 of the target prediction model 500. Through the embedding layer of the feature pre-training network 502, the word embeddings of the consumer object and the consumed object are obtained. Then, the word embeddings of the consumer object and the consumed object can be provided to the input layer of the target calibration network 501. Figure 5 (Not shown in the image).

[0073] Step 404: The target calibration network obtains an adjustment factor based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, the industry domain knowledge corresponding to the training samples, and the initial prediction results. This adjustment factor is used to fit the difference between the initial prediction results and the statistical results corresponding to the training samples.

[0074] Industry domain knowledge refers to the domain knowledge of the industry corresponding to the training samples. This industry domain knowledge can be used to represent the characteristics of various domains within that industry. For example, taking the e-commerce industry as an example, this industry domain knowledge can be used to represent the characteristics of various domains within the e-commerce industry, such as industry name, items, and consumer relationships.

[0075] In this embodiment, industry domain knowledge may include the basic attribute feature domain of the consumer object, the industry feature domain of the consumed object, and the context feature domain of the context features. For example, the e-commerce industry can be used as an example. The basic attribute feature domain of the consumer object represents the domain knowledge corresponding to the consumer object within a certain industry, the industry feature domain represents the domain knowledge corresponding to the consumed object within a certain industry, and the context feature domain represents the context knowledge corresponding to the consumed object within a certain industry. By setting the number of industry domains, this embodiment can significantly reduce the complexity of the network parameters of the target calibration network while ensuring that the target calibration network can learn industry domain knowledge.

[0076] Statistical results refer to the true values ​​corresponding to the target, which can be labeled. Taking conversion rate as an example, if the consumer converts from the target, the statistical result can be recorded as 1; if the consumer does not convert, the statistical result can be recorded as 0. There will inevitably be a difference between the conversion rate predicted by the model and the statistical result, denoted as bias. A positive bias indicates that the model overestimates the value, while a negative bias indicates that the model underestimates the value. The adjustment factor in the embodiments of this application is used to fit this bias.

[0077] In one example, refer to Figure 6 Step 404 may include the following sub-steps:

[0078] Step 404a: The target calibration network obtains the output factor based on the word embedding of the consuming object, the word embedding of the consumed object, the training samples, and the industry domain knowledge corresponding to the training samples; where the output factor refers to the probability that the consumed object will achieve the target under the industry domain knowledge.

[0079] For example, refer to Figure 5 The target calibration network 501 of the target prediction model 500 includes an embedding layer, a domain factorization layer (i.e., FwFFM), and a task layer. The embedding layer obtains optimized word embeddings of the consumer object, the consumed object, and context features based on the word embeddings of the consumer object, the consumed object, and training samples. The domain factorization layer obtains latent vector representations based on the optimized word embeddings of the consumer object, the consumed object, the context features, and industry domain knowledge corresponding to the training samples. These latent vector representations characterize the cross-features between the consumer object, the consumed object, and the context features. The task layer obtains the output factor based on the latent vector representations.

[0080] Here, the latent vector representation refers to the output of the domain factorization layer. The domain factorization layer learns the latent vector representation by fully cross-referencing the optimized word embeddings of the consumer object, the optimized word embeddings of the consumed object, and the word embeddings of the context features, using the industry domain knowledge corresponding to the training samples.

[0081] Optionally, the process of obtaining the latent vector representation can be as follows: First parameters are obtained through the domain factorization layer based on the interaction vector between the optimized word embedding of the consumer object and the industry feature domain, and the interaction vector between the optimized word embedding of the consumer object and the context feature domain; second parameters are obtained through the domain factorization layer based on the interaction vector between the optimized word embedding of the consumed object and the basic attribute feature domain of the consumer object, and the interaction vector between the optimized word embedding of the consumed object and the context feature domain; third parameters are obtained through the domain factorization layer based on the interaction vector between the word embedding of the context features and the basic attribute feature domain of the consumer object, and the interaction vector between the word embedding of the context features and the industry feature domain; fourth parameters are obtained by weighted summation of the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, and the word embedding of the context features by the domain factorization layer; and the latent vector representation is obtained through the domain factorization layer based on the first, second, third, and fourth parameters.

[0082] This implicit vector representation can be expressed by the following formula:

[0083]

[0084] Where, x i This refers to the i-th word embedding (such as the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, and the word embedding of the context features), x j This refers to the embedding of the j-th word, w i It is a word embedding x i The weight values ​​are scalars, w0 is the initial weight parameter, and F(i) refers to the word embedding x. i The corresponding feature domain (such as the basic attribute feature domain of the consumer object, the industry feature domain, and the context feature domain), v(i, F(j)) refers to the word embedding x i For the interaction vector of the feature domain F(j), v(j, F(i)) refers to the word embedding x. j For the interaction vector of the feature domain F(i), m is the total number of word embeddings.

[0085] based on The first, second, and third parameters can be obtained. This corresponds to the fourth parameter mentioned above.

[0086] Traditional FFM (Field-Ware Factorization Machines) have a large number of network parameters, O(m+m*n*k), where n is the number of feature domains and k is the dimension of word embeddings. In this embodiment, FwFFM abstracts the feature domains and divides them into three types: basic attribute feature domain of the consumer object, industry feature domain, and context feature domain. In this case, n is only 3, which can learn industry features and significantly reduce the complexity of network parameters.

[0087] Step 404b: Calculate the adjustment factor based on the output factor and the initial prediction results.

[0088] In one example, the adjustment factor can be obtained by factor product, and this adjustment factor can also be called the product factor. When the adjustment factor is a product factor, the process of obtaining the adjustment factor can be as follows: multiply the output factor and the initial prediction result to obtain the initial product factor; use the first function to map the range of the initial product factor to the target range to obtain the transition product factor; use the first adjustment parameter to scale the range of the transition product factor to obtain the adjustment factor.

[0089] Alternatively, the initial product factor can be expressed as follows:

[0090] f(p, factor) = p * factor;

[0091] Where p is the initial prediction result (e.g., initial prediction conversion rate), and factor is the output factor. Optionally, in a multi-objective prediction scenario, an initial product factor needs to be calculated for each objective.

[0092] Since the adjustment factor obtained by factor product may fluctuate too much, leading to large bias fluctuations as mentioned above, this application embodiment designs a first function on the model side to control the range of the initial output factor in order to obtain a transition product factor. This transition product factor can be expressed as follows:

[0093]

[0094] Here, tanh() is the first function, and x is the initial output factor. This maps the range of the initial output factor to the target range, such as (-1, 1). The target range can be adjusted according to actual usage requirements, and the first function can also be adjusted according to actual usage requirements.

[0095] Then, using the first adjustment parameter, the range of the transition product factor is scaled to obtain the adjustment factor, which can be expressed as follows:

[0096] g(x) = λ * tanh(x) + 1;

[0097] Where λ is the first adjustment parameter, representing the range of the transition product factor. For example, if λ is set to 0.7, the final range of the adjustment factor is (0.3, 1.7).

[0098] In another example, a residual approach, such as the GBDT (Gradient Boosting Decision Tree) residual approach, can be used to obtain the adjustment factor, which can also be called the residual factor. When the adjustment factor is a residual factor, the process of obtaining the adjustment factor can be as follows: Calculate the difference between 1 and the initial prediction result to obtain the first difference; divide the initial prediction result by the first difference to obtain the fifth parameter; calculate the adjustment factor based on the sum of the logarithm of the fifth parameter and the output factor.

[0099] Alternatively, the adjustment factor can be expressed as follows:

[0100]

[0101] Where p is the initial prediction result and factor is the output factor. This is the fifth parameter. Optionally, in multi-target prediction scenarios, an output factor needs to be calculated for each target.

[0102] Step 405: The initial prediction results are calibrated by adjusting the factors to obtain the calibrated prediction results.

[0103] Optionally, the adjustment factor can be multiplied with the initial prediction result to calibrate the initial prediction result and obtain the calibrated prediction result.

[0104] Step 406: Based on the calibrated prediction results and statistical results, train the target prediction model to obtain the trained target prediction model.

[0105] Optionally, the cross-entropy algorithm can be used to calculate the training loss of the target prediction model based on the calibrated prediction results and statistical results. This training loss is used to characterize the difference between the calibrated prediction results and statistical results. With the goal of converging the training loss, the network parameters of the target calibration network are adjusted to obtain the trained target prediction model.

[0106] During the training of the target prediction model, the training loss will first drop to a minimum, then gradually increase, and eventually stabilize. Therefore, in the initial stage of training, the Adam optimizer can be used, and the learning rate can be steadily accumulated from 0 to a set maximum learning rate (which can be set based on experience) according to the warmup and decay mechanisms. This avoids getting stuck in a local optimum due to an excessively large learning rate in the initial stage. In the later stages of training, the learning rate can be steadily decreased to complete the training of the target prediction model.

[0107] For example, refer to Figure 7 and Figure 8 , Figure 7 This diagram illustrates the training performance of a target prediction model provided in one embodiment of this application under a product factor. Figure 8 This diagram illustrates the training effect of a target prediction model provided in another embodiment of this application under a product factor. Figure 7 To determine the training effect corresponding to conventional training methods, Figure 8 As can be seen from the training effect corresponding to the training method provided in the embodiments of this application, Figure 8 The corresponding training effect is better than Figure 7 The corresponding training effect.

[0108] For example, refer to Figure 9 , Figure 9 This diagram illustrates the training performance of a target prediction model provided in one embodiment of this application under residual factors. Furthermore, existing experimental results show that training a target prediction model for the e-commerce industry based on residuals yields better results, while training a target prediction model for non-e-commerce industries based on product factors also yields better results. Optionally, target prediction models with different network parameters can be trained for different industries to achieve independent scaling by industry.

[0109] Optionally, during the use of the target prediction model, the input samples (including consumer object features, consumed object features, and contextual features of the consumed object) can be input into the initial target prediction network, the consumer object features and consumed object features can be input into the feature pre-training network, and the input samples can be input into the target calibration network. The target prediction model can then output the final prediction result corresponding to the target.

[0110] In summary, the technical solution provided in this application, by learning industry-specific knowledge corresponding to the training samples and fitting an adjustment factor that accounts for the difference between the initial prediction result and the statistical result, calibrates the initial prediction result obtained by the initial target prediction network based on the training samples, thus obtaining a calibrated prediction result. This direct calibration of the initial prediction result through the adjustment factor reduces the impact of the inevitable difference between the implementation process fitted by the initial target prediction network and the actual implementation process on the initial prediction result. This effectively reduces the difference between the calibrated prediction result and the statistical result, thereby improving the prediction accuracy of targets (such as conversion rate, click-through rate, exposure rate, etc.). Furthermore, since the calibration is based directly on industry-specific knowledge, the prediction accuracy of the target can be further improved.

[0111] Furthermore, by learning the word embeddings of consumer objects that have potential associations between consumer object characteristics and consumer object target intentions (such as purchasing, subscribing, browsing, etc.), and learning the word embeddings of consumed objects that have potential associations between consumed object characteristics and consumer object target intentions, the initial prediction results can be calibrated. This can provide the correct optimization direction for the calibration process of the initial prediction results and help to better leverage the role of industry domain knowledge, thereby effectively improving the calibration effect of the initial prediction results and further improving the prediction accuracy of the target.

[0112] In addition, by setting the feature domain to three feature domains—the basic attribute feature domain of the consumer object, the industry feature domain, and the context feature domain—the model can learn industry features while significantly reducing the complexity of the model's network parameters.

[0113] In an exemplary embodiment, taking the example of a target prediction model that can be used to predict conversion rates and the target consumer being an advertisement, the training method of the target prediction model is introduced, which may include the following.

[0114] Training samples for the target prediction model are obtained, including consumer characteristics, advertisement characteristics, and advertisement contextual characteristics. In this embodiment, the consumer can refer to the viewer corresponding to the advertisement.

[0115] The initial target prediction network of the target prediction model obtains the initial predicted conversion rate based on consumer characteristics, advertising characteristics, and advertising contextual characteristics. Optionally, when needed, the initial target prediction network can also be used to predict multiple targets, and can also be used to predict the conversion rate of different targets; this application embodiment does not limit this.

[0116] By pre-training a feature-based network for the target prediction model, word embeddings for the consumer object and advertisements are obtained based on consumer object features and advertisement features. Optionally, these word embeddings learn information about the consumer object's purchase intent. For example, the word embeddings for the consumer object learn a potential correlation between consumer object features and the consumer object's purchase intent, and the word embeddings for advertisements learn a potential correlation between advertisement features and the consumer object's purchase intent.

[0117] By using the target calibration network of the target prediction model, the optimized word embeddings of the consumer object, the optimized word embeddings of the advertisement, and the word embeddings of the context features are obtained based on the word embeddings of the consumer object, the word embeddings of the advertisement, the features of the consumer object, the features of the advertisement, and the context features of the advertisement.

[0118] By using the target calibration network of the target prediction model, and based on the industry domain knowledge corresponding to the e-commerce industry, the optimized word embeddings of consumer objects, optimized word embeddings of advertisements, and word embeddings of context features are fully cross-referenced to obtain the latent vector representations corresponding to the training samples.

[0119] By using the target calibration network of the target prediction model and based on the latent vector representation, the predicted conversion rate (i.e., the output factor) in the e-commerce industry is obtained.

[0120] An adjustment factor is calculated using the residual method, based on the predicted conversion rate and the initial predicted conversion rate in the e-commerce industry. This adjustment factor can be used to characterize the difference between the initial predicted conversion rate and the statistical conversion rate.

[0121] The adjusted factor is multiplied by the initial predicted conversion rate to obtain the calibrated conversion rate.

[0122] The cross-entropy algorithm is used to calculate the training loss of the target prediction model based on the calibrated conversion rate and statistical conversion rate. Optionally, in multi-target prediction scenarios, the training losses corresponding to multiple targets are added together to obtain the total training loss.

[0123] With the goal of converging the training loss of the target prediction model, the network parameters of the target calibration network are adjusted to obtain the trained target prediction model. This trained target prediction model can be used to predict the conversion rate of advertisements in the e-commerce industry.

[0124] For example, refer to Figure 10 This diagram illustrates the training effect of a target prediction model provided in another embodiment of this application under the residual factor. The training effect of the target prediction model used to predict conversion rate under the training method provided in this embodiment (corresponding to...) Figure 10 The new-pcvr model is superior to the target prediction model used to predict conversion rates, and its training performance under conventional training methods (corresponding to...) Figure 10(pcvr in the middle).

[0125] The technical solution provided in this application embodiment can increase GMV (Gross Merchandise Volume) and consumption in the e-commerce industry by more than 2%, and reduce bias by 3%. The experimental results are obvious. The technical solution provided in this application embodiment can obtain a more accurate conversion rate, which is conducive to the precise placement of advertisements in the e-commerce industry, thereby improving the advertising effect in the e-commerce industry.

[0126] In a feasible example, some features corresponding to the advertiser's bid can also be considered during the training process of the target prediction model. This is because the final effect of advertising is determined by the ranking and absolute value of ECPM (Effective Cost Per Miles, the advertising benefit achievable per thousand impressions); ECPM = pctr * pcvr * bid, where pctr is the predicted probability value of click-through rate, pcvr is the predicted probability value of conversion rate, and bid is the advertiser's bid. Therefore, if the advertiser's bid is fully considered during the pcvr calibration process, it will help improve the effectiveness of advertising.

[0127] In summary, the technical solution provided in this application, by learning domain knowledge of the e-commerce industry and fitting an adjustment factor that accounts for the difference between the initial predicted conversion rate and the statistical conversion rate, calibrates the initial predicted conversion rate obtained by the initial target prediction network based on training samples, resulting in a calibrated predicted conversion rate. This direct calibration of the initial predicted conversion rate through the adjustment factor reduces the inevitable difference between the conversion process fitted by the initial target prediction network and the actual conversion process, thereby effectively reducing the difference between the calibrated predicted conversion rate and the statistical conversion rate, and thus improving the accuracy of conversion rate prediction. Furthermore, since the calibration of the initial predicted conversion rate is based directly on domain knowledge of the e-commerce industry, the accuracy of conversion rate prediction can be further improved, which is beneficial for improving the rationality of advertising exposure.

[0128] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0129] Please refer to Figure 11This diagram illustrates a block diagram of a training apparatus for a target prediction model according to an embodiment of this application. The apparatus can be used to implement the training method for the aforementioned target prediction model. The apparatus 1100 may include: a training sample acquisition module 1101, a prediction result acquisition module 1102, a feature pre-training module 1103, an adjustment factor acquisition module 1104, a prediction result calibration module 1105, and a prediction model training module 1106.

[0130] The training sample acquisition module 1101 is used to acquire training samples of the target prediction model, wherein the training samples include consumer object features, consumed object features and context features of the consumed object.

[0131] The prediction result acquisition module 1102 is used to obtain an initial prediction result based on the training samples through the initial target prediction network. The initial prediction result is used to characterize the probability that the consumed object will achieve the target.

[0132] The feature pre-training module 1103 is used to obtain the word embeddings of the consumer object and the consumer object based on the features of the consumer object and the features of the consumed object through the feature pre-training network; wherein the word embeddings of the consumer object learn a potential association between the features of the consumer object and the target intent of the consumer object, and the word embeddings of the consumer object learn a potential association between the features of the consumed object and the target intent of the consumer object.

[0133] The adjustment factor acquisition module 1104 is used to obtain an adjustment factor through the target calibration network based on the word embedding of the consumer object, the word embedding of the consumed object, the training sample, the industry domain knowledge corresponding to the training sample, and the initial prediction result. The adjustment factor is used to fit the difference between the initial prediction result and the statistical result corresponding to the training sample.

[0134] The prediction result calibration module 1105 is used to calibrate the initial prediction result to obtain the calibrated prediction result.

[0135] The prediction model training module 1106 is used to train the target prediction model based on the calibrated prediction results and the statistical results, so as to obtain the trained target prediction model.

[0136] In one exemplary embodiment, such as Figure 12 As shown, the adjustment factor acquisition module 1104 includes: an output factor acquisition submodule 1104a and an adjustment factor acquisition submodule 1104b.

[0137] The output factor acquisition submodule 1104a is used to obtain an output factor through the target calibration network based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, and the industry domain knowledge corresponding to the training samples; wherein, the output factor refers to the probability that the consumed object achieves the target under the industry domain knowledge.

[0138] The adjustment factor acquisition submodule 1104b is used to calculate an adjustment factor based on the output factor and the initial prediction result. The adjustment factor is used to fit the difference between the initial prediction result and the statistical result.

[0139] In an exemplary embodiment, the target calibration network includes an embedding layer, a domain factorization layer, and a task layer; the output factor acquisition submodule 1104a is used for:

[0140] The embedding layer obtains the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, and the word embedding of the context features based on the word embedding of the consumer object, the word embedding of the consumed object, and the training samples.

[0141] The domain factor decomposition layer obtains a latent vector representation based on the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, the word embedding of the context features, and the industry domain knowledge corresponding to the training samples. The latent vector representation is used to characterize the cross features between the consumer object, the consumed object, and the context features.

[0142] The output factor is obtained by the task layer based on the implicit vector representation.

[0143] In an exemplary embodiment, the industry domain knowledge includes the basic attribute feature domain of the consumer object corresponding to the consumer object, the industry feature domain corresponding to the consumed object, and the context feature domain corresponding to the context feature; the output factor acquisition submodule 1104a is further configured to:

[0144] The first parameter is obtained by the domain factor decomposition layer based on the interaction vector between the optimized word embedding of the consumer object and the industry feature domain, and the interaction vector between the optimized word embedding of the consumer object and the context feature domain.

[0145] The second parameter is obtained by the domain factor decomposition layer based on the interaction vector between the optimized word embedding of the consumed object and the basic attribute feature domain of the consumed object, and the interaction vector between the optimized word embedding of the consumed object and the context feature domain.

[0146] The third parameter is obtained by using the interaction vector between the word embedding of the context features and the basic attribute feature domain of the consumer object based on the domain factor decomposition layer, and the interaction vector between the word embedding of the context features and the industry feature domain.

[0147] The fourth parameter is obtained by weighted summation of the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, and the word embedding of the context features through the domain factor decomposition layer.

[0148] The latent vector representation is obtained through the domain factorization layer based on the first parameter, the second parameter, the third parameter, and the fourth parameter.

[0149] In an exemplary embodiment, when the adjustment factor is a product factor, the adjustment factor acquisition submodule 1104b is configured to:

[0150] The initial product factor is obtained by multiplying the output factor and the initial prediction result.

[0151] Using the first function, the range of the initial product factor is mapped to the target range to obtain the transition product factor;

[0152] The transition product factor is range-scaled using the first adjustment parameter to obtain the adjustment factor.

[0153] In an exemplary embodiment, when the adjustment factor is a residual factor, the adjustment factor acquisition submodule 1104b is further configured to:

[0154] The difference between 1 and the initial prediction result is calculated to obtain the first difference value;

[0155] Divide the initial prediction result by the first difference to obtain the fifth parameter;

[0156] The adjustment factor is calculated based on the sum of the logarithm of the fifth parameter and the output factor.

[0157] In one exemplary embodiment, the prediction model training module 1105 is configured to:

[0158] The training loss of the target prediction model is calculated using the cross-entropy algorithm based on the calibrated prediction results and the statistical results. The training loss is used to characterize the difference between the calibrated prediction results and the statistical results.

[0159] With the goal of converging the training loss, the network parameters of the target calibration network are adjusted to obtain the trained target prediction model.

[0160] In summary, the technical solution provided in this application, by learning industry-specific knowledge corresponding to the training samples and fitting an adjustment factor that accounts for the difference between the initial prediction result and the statistical result, calibrates the initial prediction result obtained by the initial target prediction network based on the training samples, thus obtaining a calibrated prediction result. This direct calibration of the initial prediction result through the adjustment factor reduces the impact of the inevitable difference between the implementation process fitted by the initial target prediction network and the actual implementation process on the initial prediction result. This effectively reduces the difference between the calibrated prediction result and the statistical result, thereby improving the prediction accuracy of targets (such as conversion rate, click-through rate, exposure rate, etc.). Furthermore, since the calibration is based directly on industry-specific knowledge, the prediction accuracy of the target can be further improved.

[0161] Furthermore, by learning the word embeddings of consumer objects that have potential associations between consumer object characteristics and consumer object target intentions (such as purchasing, subscribing, browsing, etc.), and learning the word embeddings of consumed objects that have potential associations between consumed object characteristics and consumer object target intentions, the initial prediction results can be calibrated. This can provide the correct optimization direction for the calibration process of the initial prediction results and help to better leverage the role of industry domain knowledge, thereby effectively improving the calibration effect of the initial prediction results and further improving the prediction accuracy of the target.

[0162] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0163] Please refer to Figure 13 This illustration shows a schematic diagram of a computer device according to an embodiment of this application. The computer device can be any electronic device with data computing, processing, and storage functions, and can be implemented as... Figure 1 The model training device 10 and / or model usage device 20 in the implementation environment of the scheme shown. Specifically, it may include the following:

[0164] The computer device 1300 includes a central processing unit (such as a CPU, GPU, or FPGA) 1301, a system memory 1304 including RAM (Random-Access Memory) 1302 and ROM (Read-Only Memory) 1303, and a system bus 1305 connecting the system memory 1304 and the central processing unit 1301. The computer device 1300 also includes a basic input / output system (I / O system) 1306 to facilitate information transfer between various devices within the server, and a mass storage device 1307 for storing the operating system 1313, application programs 1314, and other program modules 1315.

[0165] In some embodiments, the basic input / output system 1306 includes a display 1308 for displaying information and an input device 1309 for user input, such as a mouse or keyboard. Both the display 1308 and the input device 1309 are connected to the central processing unit 1301 via an input / output controller 1310 connected to the system bus 1305. The basic input / output system 1306 may also include the input / output controller 1310 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1310 also provides output to a display screen, printer, or other types of output devices.

[0166] The mass storage device 1307 is connected to the central processing unit 1301 via a mass storage controller (not shown) connected to the system bus 1305. The mass storage device 1307 and its associated computer-readable media provide non-volatile storage for the computer device 1300. That is, the mass storage device 1307 may include computer-readable media (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0167] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage medium is not limited to the above-mentioned types. The system memory 1304 and mass storage device 1307 described above can be collectively referred to as memory.

[0168] According to an embodiment of this application, the computer device 1300 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1300 can be connected to the network 1312 through the network interface unit 1311 connected to the system bus 1305, or the network interface unit 1311 can be used to connect to other types of networks or remote computer systems (not shown).

[0169] The memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the training method of the target prediction model described above.

[0170] In one exemplary embodiment, a computer-readable storage medium is also provided, wherein a computer program is stored therein, which, when executed by a processor, implements the training method for the target prediction model described above.

[0171] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random-Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0172] In one exemplary embodiment, a computer program product or computer program is also provided, the computer program product or computer program including 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 the processor executes the computer instructions, causing the computer device to perform the training method for the target prediction model described above.

[0173] It should be noted that all information (including but not limited to object device information, object personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the object or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, consumer characteristics, consumer characteristics, contextual characteristics of the consumer, industry knowledge, statistical results, etc. involved in this application were all obtained with full authorization.

[0174] It should be understood that "multiple" as used herein refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

[0175] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for training a target prediction model, characterized in that, The target prediction model includes an initial target prediction network, a feature pre-training network, and a target calibration network; the method includes: Obtain training samples for the target prediction model, wherein the training samples include consumer object features, consumed object features, and contextual features of the consumed object; The initial prediction network obtains an initial prediction result based on the training samples, and the initial prediction result is used to characterize the probability that the consumed object will achieve the goal. The feature pre-training network obtains the word embeddings of the consumer object and the consumer object based on the features of the consumer object and the features of the consumed object; wherein, the word embeddings of the consumer object learn the potential association between the features of the consumer object and the target intent of the consumer object, and the word embeddings of the consumer object learn the potential association between the features of the consumed object and the target intent of the consumer object; The target calibration network obtains an adjustment factor based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, the industry domain knowledge corresponding to the training samples, and the initial prediction results. The adjustment factor is used to fit the difference between the initial prediction results and the statistical results corresponding to the training samples. The initial prediction result is calibrated using the adjustment factor to obtain the calibrated prediction result; Based on the calibrated prediction results and the statistical results, the target prediction model is trained to obtain the trained target prediction model.

2. The method according to claim 1, characterized in that, The adjustment factor is obtained through the target calibration network based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, the industry domain knowledge corresponding to the training samples, and the initial prediction results, including: The target calibration network obtains an output factor based on the word embedding of the consumer object, the word embedding of the consumed object, the training samples, and the industry domain knowledge corresponding to the training samples; wherein, the output factor refers to the probability that the consumed object will achieve the target under the industry domain knowledge. The adjustment factor is calculated based on the output factor and the initial prediction result.

3. The method according to claim 2, characterized in that, The target calibration network includes an embedding layer, a domain factorization layer, and a task layer; The step of obtaining output factors through the target calibration network based on the word embeddings of the consumer object, the word embeddings of the consumed object, the training samples, and the industry domain knowledge corresponding to the training samples includes: The embedding layer obtains the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, and the word embedding of the context features based on the word embedding of the consumer object, the word embedding of the consumed object, and the training samples. The domain factor decomposition layer obtains a latent vector representation based on the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, the word embedding of the context features, and the industry domain knowledge corresponding to the training samples. The latent vector representation is used to characterize the cross features between the consumer object, the consumed object, and the context features. The output factor is obtained by the task layer based on the implicit vector representation.

4. The method according to claim 3, characterized in that, The industry domain knowledge includes the basic attribute feature domain of the consumer object corresponding to the consumer object, the industry feature domain corresponding to the consumed object, and the context feature domain corresponding to the context feature. The process of obtaining a latent vector representation through the domain factorization layer based on the optimized word embeddings of the consumer object, the optimized word embeddings of the consumed object, the word embeddings of the context features, and the industry domain knowledge corresponding to the training samples includes: The first parameter is obtained by the domain factor decomposition layer based on the interaction vector between the optimized word embedding of the consumer object and the industry feature domain, and the interaction vector between the optimized word embedding of the consumer object and the context feature domain. The second parameter is obtained by the domain factor decomposition layer based on the interaction vector between the optimized word embedding of the consumed object and the basic attribute feature domain of the consumed object, and the interaction vector between the optimized word embedding of the consumed object and the context feature domain. The third parameter is obtained by using the interaction vector between the word embedding of the context features and the basic attribute feature domain of the consumer object based on the domain factor decomposition layer, and the interaction vector between the word embedding of the context features and the industry feature domain. The fourth parameter is obtained by weighted summation of the optimized word embedding of the consumer object, the optimized word embedding of the consumed object, and the word embedding of the context features through the domain factor decomposition layer. The latent vector representation is obtained through the domain factorization layer based on the first parameter, the second parameter, the third parameter, and the fourth parameter.

5. The method according to claim 2, characterized in that, When the adjustment factor is a product factor, the calculation of the adjustment factor based on the output factor and the initial prediction result includes: The initial product factor is obtained by multiplying the output factor and the initial prediction result. Using the first function, the range of the initial product factor is mapped to the target range to obtain the transition product factor; The transition product factor is range-scaled using the first adjustment parameter to obtain the adjustment factor.

6. The method according to claim 2, characterized in that, When the adjustment factor is a residual factor, the calculation of the adjustment factor based on the output factor and the initial prediction result includes: The difference between 1 and the initial prediction result is calculated to obtain the first difference value; Divide the initial prediction result by the first difference to obtain the fifth parameter; The adjustment factor is calculated based on the sum of the logarithm of the fifth parameter and the output factor.

7. The method according to any one of claims 1 to 6, characterized in that, The step of training the target prediction model based on the calibrated prediction results and the statistical results to obtain the trained target prediction model includes: The training loss of the target prediction model is calculated using the cross-entropy algorithm based on the calibrated prediction results and the statistical results. The training loss is used to characterize the difference between the calibrated prediction results and the statistical results. With the goal of converging the training loss, the network parameters of the target calibration network are adjusted to obtain the trained target prediction model.

8. A training device for a target prediction model, characterized in that, The target prediction model includes an initial target prediction network, a feature pre-training network, and a target calibration network; the device includes: The training sample acquisition module is used to acquire training samples of the target prediction model, wherein the training samples include consumer object features, consumed object features and context features of the consumed object; The prediction result acquisition module is used to obtain an initial prediction result based on the training samples through the initial target prediction network. The initial prediction result is used to characterize the probability that the consumed object will achieve the target. The feature pre-training module is used to obtain the word embeddings of the consumer object and the consumer object based on the features of the consumer object and the features of the consumed object through the feature pre-training network; wherein the word embeddings of the consumer object learn a potential association between the features of the consumer object and the target intent of the consumer object, and the word embeddings of the consumer object learn a potential association between the features of the consumed object and the target intent of the consumer object; The adjustment factor acquisition module is used to obtain an adjustment factor through the target calibration network based on the word embedding of the consumer object, the word embedding of the consumed object, the training sample, the industry domain knowledge corresponding to the training sample, and the initial prediction result. The adjustment factor is used to fit the difference between the initial prediction result and the statistical result corresponding to the training sample. The prediction result calibration module is used to calibrate the initial prediction result using the adjustment factor to obtain the calibrated prediction result; The prediction model training module is used to train the target prediction model based on the calibrated prediction results and the statistical results, so as to obtain the trained target prediction model.

9. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the training method for the target prediction model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the training method for the target prediction model as described in any one of claims 1 to 7.

Images