Model training method, information promotion method and related products

Abstract

The embodiment of the invention provides a model training method, an information promotion method and a related product, and relates to the technical field of machine learning. The method comprises the following steps: firstly, acquiring target object data, semantic features for evaluating to-be-popularized information and input virtual resource quantity; secondly, a training label is constructed according to the target object data, the semantic features and the virtual resource quantity, the training label is used for identifying whether the target object inputs virtual resources for the to-be-popularized information or not and identifying the input virtual resource quantity, so that a machine learning model can process a dichotomy task and a regression task at the same time, and the flexibility of the machine learning model is improved; and finally, training a machine learning model according to the training label. Therefore, model training is carried out by utilizing the front-end data (including the target object data, the semantic features and the virtual resource quantity), and model training can be carried out even if the rear-end data is delayed or cannot be acquired, so that the accuracy of a pre-estimation popularization result output by the value pre-estimation model is improved, and the information popularization effect is improved.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a model training method, an information dissemination method, and related products. Background Technology

[0002] With the rapid development of the internet, online information promotion methods are becoming increasingly diversified. As advertisers hope to effectively promote their information to their target audience through online information promotion methods and obtain the maximum amount of virtual resources at the lowest cost, maximizing the return on investment (ROI) has become a key focus for advertisers.

[0003] In order to calculate ROI based on promotion costs and estimated promotion results, the features of the information to be promoted (such as click-through rate and conversion rate) and the features of the target object to be promoted need to be input into a pre-trained value prediction model to obtain the estimated promotion results.

[0004] However, the characteristics of the information to be promoted often require a large data return delay to obtain, or even cannot be obtained at all, thus making it impossible to accurately predict the promotion results and affecting the overall effectiveness of information promotion. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a model training method, an information promotion method, and related products, which can improve the accuracy of the predicted promotion results output by the value prediction model, thereby enhancing the effectiveness of information promotion.

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

[0007] In view of the above, the first aspect of this application discloses a model training method, the method comprising:

[0008] Acquire target object data, semantic features of the target object's evaluation of the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted;

[0009] Based on the target object data, the semantic features, and the amount of virtual resources, training labels are constructed, wherein the training labels are used to identify whether the target object has invested virtual resources in the information to be promoted, and to identify the amount of virtual resources invested by the target object in the information to be promoted;

[0010] The machine learning model is trained based on the training labels to obtain the value prediction model.

[0011] A second aspect of this application discloses an information promotion method, the method comprising:

[0012] Obtain semantic features of the target audience's data and their evaluation of the promotional information;

[0013] The value prediction model determines the predicted promotion results corresponding to the semantic features of the data of the object to be promoted and the evaluation of the information to be promoted by the object to be promoted. The predicted promotion results include the probability of the object to be promoted investing virtual resources in the information to be promoted and the amount of virtual resources invested by the object to be promoted in the information to be promoted. The value prediction model is trained by the model training method described in the first aspect.

[0014] Based on the estimated promotion results and promotion costs, the promotion value of the information to be promoted is determined.

[0015] A third aspect of this application discloses a model training device, the device comprising: an acquisition module, a construction module, and a training module;

[0016] The acquisition module is used to acquire target object data, semantic features of the target object's evaluation of the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted.

[0017] The construction module is used to construct training labels based on the target object data, the semantic features, and the amount of virtual resources. The training labels are used to identify whether the target object has invested virtual resources in the information to be promoted, and to identify the amount of virtual resources invested by the target object in the information to be promoted.

[0018] The training module is used to train the machine learning model based on the training labels to obtain a value prediction model.

[0019] A fourth aspect of this application discloses an information promotion device, the device comprising: a feature acquisition module, a result acquisition module, and a value determination module;

[0020] The feature acquisition module is used to acquire the data of the object to be promoted and the semantic features of the object's evaluation of the promotion information.

[0021] The result acquisition module is used to determine the estimated promotion result corresponding to the semantic features of the data of the object to be promoted and the evaluation of the information to be promoted by the object to be promoted through the value prediction model. The estimated promotion result includes the probability of the object to be promoted investing virtual resources in the information to be promoted and the amount of virtual resources invested by the object to be promoted in the information to be promoted. The value prediction model is trained by the model training method described in the first aspect.

[0022] The value determination module is used to determine the promotion value of the information to be promoted based on the estimated promotion results and promotion costs.

[0023] A fifth aspect of this application discloses a computer device, the device comprising a processor and a memory:

[0024] The memory is used to store program code and transmit the program code to the processor;

[0025] The processor is configured to execute the steps of the model training method described in the first aspect, or the information promotion method described in the second aspect, according to the instructions in the program code.

[0026] The sixth aspect of this application discloses a computer-readable storage medium for storing program code for performing the steps of the model training method described in the first aspect, or the information promotion method described in the second aspect.

[0027] Compared with the prior art, this application has the following beneficial effects:

[0028] This application discloses a model training method, an information promotion method, and related products. First, it acquires target object data, semantic features of the evaluation of the information to be promoted, and the amount of virtual resources invested. Second, it constructs training labels based on the target object data, semantic features, and virtual resource amount. These training labels identify whether the target object has invested virtual resources in the information to be promoted and the amount of virtual resources invested, enabling the machine learning model to handle both binary classification and regression tasks simultaneously, thus improving the flexibility of the machine learning model. Finally, it trains the machine learning model based on the training labels. Therefore, by using front-end data (including target object data, semantic features, and virtual resource amount) for model training, model training can be performed even if back-end data is delayed or unavailable, thereby improving the accuracy of the predicted promotion results output by the value prediction model and ultimately enhancing the information promotion effect. Attached Figure Description

[0029] 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.

[0030] Figure 1 This is a diagram illustrating an online information promotion method.

[0031] Figure 2 A scenario architecture diagram of a model training method provided in this application embodiment;

[0032] Figure 3A flowchart illustrating a model training method provided in this application embodiment;

[0033] Figure 4 A flowchart illustrating an information promotion method provided in this application embodiment;

[0034] Figure 5 A schematic diagram of a DeepFM model provided in an embodiment of this application;

[0035] Figure 6 A schematic diagram of a model training device provided in an embodiment of this application;

[0036] Figure 7 A schematic diagram of an information promotion device provided in an embodiment of this application;

[0037] Figure 8 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application;

[0038] Figure 9 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0040] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular 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 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.

[0041] With the rapid development of the internet, online information promotion methods are becoming increasingly diversified. (See also...) Figure 1 The image is a schematic diagram of an online information promotion method. Figure 1 The promotional information in the image refers to the game represented by the screen. Figure 1 The design aims to attract users to the software page by showing them the visuals, sparking their interest in the game represented by the visuals, and ultimately encouraging them to download and experience the game.

[0042] Since advertisers hope to promote their information to their target audience through effective online marketing methods and obtain the greatest revenue from the audience at the lowest cost, maximizing ROI has become a key focus for advertisers.

[0043] To maximize ROI, a more accurate prediction of promotion results is needed first, so that ROI can be precisely calculated based on promotion costs and predicted results. In related technologies, the features of the information to be promoted and the features of the target object are typically input into a pre-trained value prediction model, which then outputs the predicted promotion results for the information to be promoted.

[0044] However, the characteristics of promotional information, such as click-through rate and conversion rate, usually come from backend data, which often requires a large data return delay to obtain. This limits the accuracy of the value prediction model, resulting in lower accuracy of the predicted promotional results output by the value prediction model, and thus affecting the overall effectiveness of information promotion.

[0045] Through research, the inventors proposed a model training method, an information promotion method, and related products. First, by acquiring target object data, semantic features of the evaluation of the information to be promoted, and the amount of virtual resources invested, more comprehensive training data is provided for the machine learning model, helping it to understand the target object's preference for virtual resource investment more deeply. Second, training labels are constructed based on the target object data, semantic features, and virtual resource amount. These labels identify whether the target object has invested virtual resources (binary classification task) and the amount of virtual resources invested (regression task), enabling the machine learning model to handle both binary classification and regression tasks simultaneously, thus improving its flexibility. Finally, the machine learning model is trained based on the training labels. Therefore, by using front-end data (including target object data, semantic features, and virtual resource amount) for model training, even if back-end data is delayed or unavailable, model training can still be performed, thereby improving the accuracy of the predicted promotion results output by the value prediction model and ultimately enhancing the effectiveness of information promotion.

[0046] Next, the implementing entities of the model training method and information propagation method provided in the embodiments of this application will be specifically introduced:

[0047] The execution entity of the model training method and information promotion method provided in this application embodiment can be a computer device with data processing capabilities, specifically a terminal device or a server. As examples, terminal devices can include, but are not limited to, mobile phones, desktop computers, tablet computers, laptops, handheld computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. Furthermore, the server can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. In addition, the model training method and information promotion method provided in this application embodiment can also be executed collaboratively by the terminal device and the server. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. Therefore, this application embodiment does not limit the implementation entity of the technical solution of this application.

[0048] Next, the scenarios for the model training method provided in the embodiments of this application will be described in detail:

[0049] See Figure 2 This figure is a scenario architecture diagram of a model training method provided in an embodiment of this application, including a computer device 100. The computer device 100 can be one of the various types of terminal devices or servers described above. Specifically, the following embodiments use computer device 100 as server A for illustrative purposes.

[0050] Server A is used to acquire target object data, semantic features of the target object's evaluation of the promotional information, and the amount of virtual resources invested by the target object in the promotional information. Thus, by collecting target object data (such as historical payment characteristics, historical click characteristics, etc.), semantic features of the target object's evaluation of the promotional information, and the amount of virtual resources invested by the target object in the promotional information, Server A provides a comprehensive and rich sample dataset for subsequent machine learning model training. This helps the machine learning model to more deeply understand the object's preferences and willingness to invest virtual resources, thereby improving the prediction accuracy and generalization ability of the machine learning model.

[0051] Server A is also used to construct training labels based on target object data, semantic features, and virtual resource quantity. These training labels identify whether a target object receives virtual resources for the information to be promoted, and the amount of virtual resources received by the target object for that information. Thus, by integrating target object data, semantic features, and virtual resource quantity to construct specific training labels, Server A helps the machine learning model to focus more effectively during training, thereby improving the efficiency and effectiveness of model training. Furthermore, comprehensive training guidance based on target object data, semantic features, and virtual resource quantity helps the machine learning model better capture the relationship between behavior, semantics, and virtual resource quantity, improving the predictive accuracy of the machine learning model.

[0052] Server A is also used to train a machine learning model based on training labels to obtain a value prediction model. Thus, Server A uses front-end data (including target object data, semantic features, and virtual resource quantity) to train the machine learning model. Even if there is a delay or unavailability of back-end data, the machine learning model can still be trained, thereby avoiding the problem of low accuracy in predicting promotion results due to back-end data issues. This improves the accuracy of the predicted promotion results and ultimately enhances the overall effectiveness of information promotion.

[0053] Next, taking server A as the execution subject, the model training and generalization method provided in this application embodiment will be described in detail:

[0054] See Figure 3 This figure is a flowchart of a model training method provided in an embodiment of this application. Figure 3 The model training method shown includes the following steps:

[0055] S301: Obtain target object data, semantic features of the target object's evaluation of the promotional information, and the amount of virtual resources invested by the target object in the promotional information.

[0056] Target audience data can include the target audience's (any user's) historical click characteristics and historical engagement characteristics. Historical click characteristics refer to the target audience's clicking behavior towards promotional information, including the number of clicks, click frequency, and click path. Historical click characteristics reflect the target audience's level of interest and focus on promotional information, helping to identify potential target audience groups. Historical engagement characteristics can include the target audience's engagement tier (e.g., classifying target audiences into different tiers based on the amount of virtual resources invested), engagement frequency (how often the target audience invests virtual resources, such as once a month, once a week, etc.), engagement preferences (which information-related products the target audience prefers to invest virtual resources in, e.g., whether the target audience prefers to invest virtual resources in virtual products or physical products), and engagement records.

[0057] The semantic features of a target audience's evaluation of promotional information represent their attitudes and preferences towards it. For example, these semantic features may include their evaluations of the promotional information's functionality, performance, appearance, value, cost-effectiveness, etc., as well as their emotional tendencies towards it (such as positive, negative, or neutral evaluations).

[0058] The amount of virtual resources invested by the target audience in promoting the information is a specific numerical value that directly reflects the target audience's ability and willingness to invest. For example, virtual resources could be promotional platform points or in-game currency.

[0059] It should be noted that when acquiring target data, the semantic characteristics of the target object's promotional information, and the amount of virtual resources invested by the target object in the promotional information, the server will strictly adhere to information security and privacy protection principles. This ensures that the acquired target data, semantic characteristics of the target object's promotional information, and the amount of virtual resources invested by the target object are all information and data authorized by the target object or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. Additionally, necessary anonymization processing must be performed on the collected target data, semantic characteristics of the target object's promotional information, and the amount of virtual resources invested by the target object in the promotional information. Anonymization refers to the technical processing of the target data, semantic characteristics of the target object's promotional information, and the amount of virtual resources invested by the target object in the promotional information, reducing or eliminating the sensitivity of personal identification without altering the meaning of these elements, thereby preventing the leakage of personal privacy.

[0060] S302: Construct training labels based on target object data, semantic features, and virtual resource quantity. The training labels are used to identify whether the target object has virtual resources invested in the information to be promoted, and to identify the amount of virtual resources invested in the target object for the information to be promoted.

[0061] Training labels serve as guidance for machine learning models, helping them understand the behavior of target objects. In the model training method provided in this application, training labels are used to instruct the machine learning model to learn two aspects of information: first, whether the target object invests virtual resources in the information to be promoted; and second, the amount of virtual resources invested by the target object in the information to be promoted.

[0062] The label "Whether the target object invests virtual resources in the information to be promoted" is a binary label, which can be "yes" or "no", or "1" or "0". For example, if the target object invests virtual resources in the information to be promoted, the binary label is "yes" or "1"; if the target object does not invest virtual resources in the information to be promoted, the binary label is "no" or "0". By providing "Whether the target object invests virtual resources in the information to be promoted" as a binary label, machine learning models can understand which factors (such as target object data and semantic features) influence the target object's decision and learn the decision boundary accordingly.

[0063] "The amount of virtual resources invested by the target audience in promoting the information" is a regression label, which can be a specific numerical value. By providing "the amount of virtual resources invested by the target audience in promoting the information" as a regression label, machine learning models can learn the distribution pattern of the amount of virtual resources invested by the target audience, thereby enabling the prediction of the amount of virtual resources invested.

[0064] It should be noted that the training label mentioned above can be a single complete label or composed of two sub-labels, each used for different prediction tasks. The first sub-label identifies the probability of a target object investing virtual resources in the information to be promoted, while the second sub-label identifies the amount of virtual resources invested by the target object. In other words, when the target object data and semantic features are input into the machine learning model, the first sub-label is used to train the first module of the machine learning model to identify the probability of the target object investing virtual resources in the information to be promoted; and the second sub-label is used to train the second module of the machine learning model to identify the amount of virtual resources invested by the target object. By using two sub-labels, the learning of the machine learning model can be more accurately guided, improving the accuracy of predictions.

[0065] Understandably, the construction of training labels is crucial for training machine learning models. Accurate training labels help machine learning models learn the potential relationships between target object data, semantic features, and the amount of virtual resources invested, thereby achieving accurate predictions of the invested virtual resources. If the training labels are incorrect or inconsistent, the machine learning model will learn incorrect information, thus affecting its predictive performance.

[0066] S303: Train a machine learning model based on the training labels to obtain a value prediction model.

[0067] The constructed training labels are used as input to the machine learning model to guide its training process. The goal of the machine learning model is to learn the potential relationship between target object data, semantic features, and the amount of virtual resources invested based on the training labels, so as to accurately predict in the future whether the target object is willing to invest virtual resources in the information to be promoted, and the amount of virtual resources that the target object may invest in the information to be promoted.

[0068] During the training of a machine learning model, iterative optimization algorithms (such as gradient descent and stochastic gradient descent) are used to continuously adjust the model parameters to minimize the difference between the estimated virtual resource quantity and the actual training labels. This difference is usually measured by a loss function; the smaller the value of the loss function, the more accurate the prediction results of the trained value prediction model.

[0069] Through continuous iteration and optimization, machine learning models will gradually learn the relationship between target object data, semantic features, and the amount of virtual resources invested. Based on the target object's data and semantic features, they can predict whether the target object is willing to invest virtual resources in the information to be promoted, and the amount of virtual resources that might be invested (i.e., estimated promotion results). These estimated promotion results can provide strong support for enterprises in formulating marketing strategies, product pricing, and advertising placement.

[0070] In summary, this application discloses a model training method, which includes: first, acquiring target object data, semantic features of the evaluation of the information to be promoted, and the amount of virtual resources invested, providing more comprehensive training data for the machine learning model and helping the machine learning model to understand the target object's preference for the amount of virtual resources invested more deeply; second, constructing training labels based on the target object data, semantic features, and virtual resource amount, the training labels are used to identify whether the target object has invested virtual resources (binary classification task) and the amount of virtual resources invested (regression task), enabling the machine learning model to handle both binary classification and regression tasks simultaneously, thus improving the flexibility of the machine learning model; finally, training the machine learning model based on the training labels. Therefore, by using front-end data (including target object data, semantic features, and virtual resource amount) for model training, even if there is a delay or unavailability of back-end data, model training can still be performed, thereby improving the accuracy of the predicted promotion results output by the value prediction model, and thus improving the information promotion effect.

[0071] See Figure 4 This figure is a flowchart of an information promotion method provided in an embodiment of this application. It should be noted that the information promotion method provided in this embodiment is applied to a demand-side platform (DSP). The DSP side can be a server. This application does not limit the specific DSP-side device. Figure 4 The information promotion method shown includes the following steps:

[0072] S401: The DSP side acquires target object data.

[0073] Target audience refers to a subset of objects selected from the database managed by the DSP or from partner data providers. Target audiences may be selected based on certain criteria (such as virtual characters in games, game levels, etc.) to represent the target audience of the information to be promoted. By selecting representative target audiences, the DSP can more accurately target its market, thereby improving the effectiveness and efficiency of information promotion.

[0074] Target object data includes basic information about the target object, but typically does not directly relate to the target object's virtual resource investment behavior. For example, basic information about the target object may include the target object's level / rank and login time preferences.

[0075] Target audience data also includes data that directly reflects the target audience's preferences and habits, such as historical click characteristics and historical spending characteristics. Historical click characteristics refer to the target audience's clicking behavior towards promotional information, including the number of clicks, click frequency, and click path. Historical spending characteristics can include the target audience's spending tiers, spending frequency, spending preferences, and spending records.

[0076] It should be noted that when acquiring target data, the DSP side strictly adheres to information security and privacy protection principles, ensuring that all acquired target data is authorized by the target or fully authorized by all parties. Furthermore, the collection, use, and processing of such data must comply with relevant national and regional laws, regulations, and standards. Necessary anonymization processing is also required for the acquired target data. Anonymization refers to the technical processing of target data to reduce or eliminate the sensitivity of personally identifiable information without altering its original meaning, thereby preventing the leakage of personal privacy.

[0077] S402: The DSP side obtains the semantic features of the target object's evaluation of the promotion information.

[0078] First, the DSP needs to obtain the target audience's textual evaluation of the promotional information, which may include the target audience's evaluation of the promotional information's functions, performance, appearance, value, cost-effectiveness, etc., as well as the target audience's emotional tendency towards the promotional information (such as positive evaluation, negative evaluation, neutral evaluation, etc.).

[0079] Subsequently, the DSP, leveraging the natural language processing capabilities of the Large Language Model (LLM), extracts semantic features relevant to the information to be promoted from the collected textual reviews. The LLM model is a machine learning model capable of understanding and generating natural language text, possessing the ability to handle complex language structures and semantic relationships. In the information promotion method provided in this application embodiment, the DSP can utilize the LLM model to deeply analyze the textual content of the textual reviews, identifying keywords, phrases, and their semantic connections related to the information to be promoted, thereby extracting valuable semantic features.

[0080] It should be noted that after extracting the semantic features of the evaluation of the information to be promoted using the LLM model, these semantic features also need to be converted into an embedding representation that can be understood by the subsequent machine learning model. An embedding representation is a form that transforms high-dimensional data into a low-dimensional vector space, enabling the machine learning model to more accurately understand the implicit meaning in the semantic features and providing strong support for subsequent model training.

[0081] It should also be noted that the LLM model can interact with the target object through natural language processing, collecting real-time feedback. Based on this feedback, the LLM model can further analyze the target object's preferences and needs, dynamically generating more personalized semantic features. These personalized semantic features can more accurately reflect the target object's current state and preferences, providing strong support for subsequent model training.

[0082] It should also be noted that when processing the semantic features of promotional information from target audiences, the DSP will strictly adhere to information security and privacy protection principles. This ensures that all semantic features obtained from target audiences regarding promotional information are information and data authorized by the target audience or fully authorized by all parties involved. Furthermore, the collection, use, and processing of such data must comply with relevant national and regional laws, regulations, and standards. Additionally, necessary anonymization processing must be performed on the collected semantic features of promotional information from target audiences. Anonymization refers to the technical processing of the semantic features of promotional information from target audiences to reduce or eliminate the sensitivity of personal identification without altering the original meaning of the semantic features, thereby preventing the leakage of personal privacy.

[0083] S403: The DSP side constructs training labels based on the target object data, semantic features, and the amount of virtual resources invested by the target object in the information to be promoted. The training labels are used to identify whether the target object has invested virtual resources in the information to be promoted, and to identify the amount of virtual resources invested by the target object in the information to be promoted.

[0084] Training labels are constructed based on the target object's data and semantic features (hereinafter referred to as multi-dimensional features) and the amount of virtual resources invested by the target object in the information to be promoted. Training labels are the foundation for machine learning model learning and can be divided into first sub-labels and second sub-labels.

[0085] The first sub-tag is the Lifetime Value (LTV) tag, used to indicate whether the target audience has invested virtual resources in the promotional information within the target time period. This is a binary classification problem, with the tag divided into "invested" and "not invested". Understandably, the first sub-tag can also be a tag used to identify the probability of the target audience investing virtual resources in the promotional information. In the subsequent information promotion process, if the probability of the target audience investing is greater than or equal to the investment threshold (usually 0.5), the target audience is classified as "invested"; if the probability of the target audience investing is less than the investment threshold, the target audience is classified as "not invested". This application does not limit this aspect.

[0086] The second sub-label is the Zero-Inflated Lognormal (ZLIN) label, used to indicate the amount of virtual resources invested by the target object in the information to be promoted. This is a regression problem, and the machine learning model needs to predict a specific value. Understandably, if the first sub-label indicates that the target object did not invest any virtual resources in the information to be promoted within the target time period, or if the first sub-label indicates that the probability of the target object investing virtual resources in the information to be promoted is 0, then the amount of virtual resources invested as predicted by the second sub-label can be directly determined to be 0.

[0087] In the information propagation method provided in this application embodiment, the machine learning model can be a DeepFM model. See also Figure 5 This figure is a schematic diagram of a DeepFM model provided in an embodiment of this application. Figure 5 The DeepFM model shown integrates multiple modules, including a Logistic Regression (LR) module, a Factor Machine (FM) module, an Embedding layer module, a Full Connect (FC) layer module, a Gated Recurrent Unit (GRU) module, and a feature fusion (concat) module.

[0088] Specifically, such as Figure 5 The LR module shown is a linear module suitable for binary classification problems. The LR module identifies whether the target object has invested virtual resources within a target time period based on the target object's multi-dimensional characteristics (i.e., binary classification into "invested" and "not invested").

[0089] like Figure 5 The FM module shown captures second-order interactions between multi-dimensional features of a target object. Specifically, first, the FM module maps all multi-dimensional features into a low-dimensional, continuous embedding space, forming embedding vectors. Then, the FM module pairs these embedding vectors together, calculating the inner product of each pair to measure their similarity or relevance. By performing inner product operations on all multi-dimensional feature pairs, the FM module generates an interaction term matrix. In addition to the interaction terms, the FM module also considers linear terms of the multi-dimensional features. The linear terms are the sum of the products of the multi-dimensional features and their corresponding weights, reflecting the direct impact of the multi-dimensional features on the predicted value. Finally, the FM module adds the strengths of all second-order interactions to the weights of the linear terms to obtain a final predicted value. This predicted value can be used to identify both the probability of the target object investing virtual resources in the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted.

[0090] like Figure 5 The FC module shown receives the embedding vectors output by the Embedding module, mines and fuses the nonlinear relationships between these embedding vectors through a multi-layer neural network structure, and finally outputs a prediction value. This prediction value can be used to identify both the probability of a target object investing virtual resources in the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted.

[0091] like Figure 5 The GRU module shown receives the embedding vectors output by the Embedding module and captures the temporal dependencies between these embedding vectors, thereby outputting a prediction value. This prediction value can be used to identify both the probability of a target object investing virtual resources in the information to be promoted and the amount of virtual resources invested by the target object in the information to be promoted.

[0092] like Figure 5 The concat module shown is responsible for concatenating and fusing predictions from different modules to construct training labels. In the DeepFM model, the concat module can concatenate predictions from the LR, FM, FC, and GRU modules to construct the final training labels.

[0093] It's important to note that when constructing training labels using the GRU module, temporal features can be selected first from multi-dimensional features. Temporal features are those that change over time, reflecting the target audience's behavior and preferences at different points in time. For example, temporal features corresponding to the target audience's data could include: click-through rate, browsing time, and search keywords over a past period (e.g., the past 7 days). Temporal features corresponding to semantic features could include the target audience's evaluation of promotional information and their sentiment tendencies over a past period. Selecting temporal features aims to capture dynamic changes in behavior or semantics, thereby more accurately predicting future paid behavior.

[0094] Secondly, to more accurately capture timeliness, temporal features are weighted according to attention weights. Attention weights are a mechanism used to weight different temporal features, allowing the machine learning model to focus more on important features during training. The attention weight is negatively correlated with the time interval between the feature and the current time. That is, temporal features closer to the current time are given higher weights because they better reflect the current state and preferences of the object. This weighting helps the machine learning model pay more attention to recent temporal features during training, thereby improving prediction accuracy. For example, suppose the click-through rate data for the target object over the past week is: 10 clicks on day 1, 15 clicks on day 2, 15 clicks on day 3, 25 clicks on day 4, 20 clicks on day 5, 40 clicks on day 6, and 40 clicks on day 7, and the current time is day 8. Then, the click-through rate data for day 7 will be given the highest attention weight because it is the closest to the current time.

[0095] Finally, training labels are constructed based on the weighted temporal characteristics and the amount of virtual resources invested. It is understandable that, through the above steps, the GRU module can more accurately capture the temporal characteristics of the target object during the construction of training labels, thereby improving the prediction accuracy of the value prediction model. This method not only considers the historical behavior of the target object but also emphasizes the importance of recent behavior, helping to better reflect the current state and preferences of the target object.

[0096] It should also be noted that choosing an appropriate time period (i.e., the target time period) is crucial for improving the predictive accuracy of the value prediction model when constructing training labels. The target time period refers to the specific time frame that the machine learning model focuses on; a more recent time frame is usually chosen because recent behavior and preferences better reflect the current state of the target object (e.g., choosing the past 7 days, 30 days, or 90 days as the target time period). By limiting the target time period, the value prediction model can be ensured to focus more on recent behavior and preferences and reduce interference from historical data, thereby improving the predictive accuracy of the value prediction model.

[0097] Specifically, when performing steps S401 and S402 to acquire target object data, semantic features, and virtual resource quantities, only the target object data, semantic features, and virtual resource quantities within the target time period can be acquired, and training labels can be constructed solely based on the target object data, semantic features, and virtual resource quantities within the target time period. Thus, the training labels are used to identify whether the target object has invested virtual resources in the information to be promoted within the target time period, and to identify the amount of virtual resources invested by the target object in the information to be promoted within the target time period.

[0098] S404: The DSP side trains a machine learning model based on the training labels to obtain a value prediction model.

[0099] During the training of a machine learning model, it is necessary to adjust the model parameters to minimize the error between the predicted generalization result output by the value prediction model and the actual training labels.

[0100] In some specific implementations, the first step is to obtain the semantic features of the test object's data (such as number of views, click-through rate, dwell time, etc.), the test object's evaluation of the promotional information (such as the object's sentiment tendency, degree of attention to the information's characteristics), and the actual amount of virtual resources invested by the test object in the promotional information.

[0101] Secondly, the test subject data and the semantic features of the test subjects' evaluation of the information to be promoted are input into the value prediction model to obtain the predicted probability and predicted amount of virtual resources that the test subjects will invest in the information to be promoted. The predicted probability is a value between 0 and 1, representing the likelihood that the test subjects will invest virtual resources in the information to be promoted, as output by the value prediction model.

[0102] Subsequently, if the predicted input probability indicates that the test object will invest virtual resources in the information to be promoted (e.g., the predicted input probability is greater than the input threshold, such as 0.5), then based on the predicted virtual resource quantity and the actual virtual resource quantity, the loss value of the value prediction model is obtained by calculating the loss function of the value prediction model.

[0103] Finally, based on the calculated loss value, the model parameters of the value prediction model are adjusted using optimization algorithms such as backpropagation and gradient descent to reduce the loss value of the loss function.

[0104] It should be noted that minimizing the loss function can optimize the model parameters of the value prediction model and improve its prediction accuracy. In the information generalization method provided in this application, the loss function includes a mean loss function and a variance loss function. The mean loss function measures the average difference between the predicted virtual resource quantity and the actual virtual resource quantity. The variance loss function measures the uncertainty or volatility of the predicted virtual resource quantity.

[0105] Specifically, firstly, based on the predicted and actual virtual resource quantities, the mean loss function of the value estimation function is calculated to obtain the mean loss value of the value estimation function; and secondly, the variance loss function of the value estimation function is calculated to obtain the variance loss value of the value estimation function. It should be noted that this application does not limit the specific mean loss function and variance loss function.

[0106] For example, the mean loss function can be represented by the following formula (1):

[0107]

[0108] Among them, y i This refers to the actual amount of virtual resources. It predicts the amount of virtual resources, where N is the number of samples.

[0109] For example, the variance loss function can be represented by the following formula (2):

[0110]

[0111] in, It is the average value of the predicted virtual resource quantity. It predicts the amount of virtual resources, where N is the number of samples.

[0112] Subsequently, the mean loss and variance loss values ​​are weighted and summed (to balance the effects of mean loss and variance loss) to obtain the total loss value of the value prediction model.

[0113] In some specific implementations, to prevent excessively large variance σ (i.e., excessively drastic fluctuations in virtual resource quantity), the variance loss function also includes a boundary-ZLIN (z-linked upper bound) on σ. This helps ensure that the prediction results of the trained value prediction model are more stable and reliable. In other words, if the variance loss value exceeds a preset variance loss threshold, a weighted sum of the mean loss value and the variance loss value is required. By introducing the variance loss value, the prediction uncertainty of the value prediction model can be better controlled, improving the robustness of the value prediction model.

[0114] For example, the formula for boundary-ZLIN can be shown in formula (3) below:

[0115]

[0116] in, It is the Binary Cross-Entropy Loss, used to handle binary classification problems. In the information promotion method provided in this application embodiment, it is used to predict whether the target variable y is greater than 0, that is, whether the target object is to invest virtual resources in the information to be promoted. It is a loss function that follows a log-normal distribution and is used to handle cases where y is greater than 0, i.e., the target object is the amount of virtual resources invested in the information to be promoted.

[0117] S405: The semantic features of the data of the target object to be promoted and the evaluation of the target object on the promotion information are obtained by the DSP side.

[0118] It should be noted that when acquiring semantic features of the target audience's data and their evaluations of the promotional information, the server strictly adheres to information security and privacy protection principles. This ensures that all acquired data and semantic features are authorized by the target audience or fully authorized by all parties involved. Furthermore, the collection, use, and processing of this data must comply with relevant national and regional laws, regulations, and standards. Additionally, necessary anonymization processing is performed on the collected data and semantic features of the target audience's data and their evaluations of the promotional information. Anonymization refers to the technical processing of the semantic features of the target audience's data and their evaluations of the promotional information to reduce or eliminate the sensitivity of personally identifiable information without altering their meaning, thereby preventing the leakage of personal privacy.

[0119] S406: The DSP side will input the semantic features of the target object data and the target object's evaluation of the promotion information into the value assessment model to obtain the estimated promotion results of the target object.

[0120] By inputting the semantic features of the target audience's data and their evaluation of the promotional information into the value assessment model, the predicted promotion results corresponding to these semantic features are determined. The predicted promotion results include whether the target audience will invest virtual resources in the promotional information within the target time period (e.g., 7 days, 14 days, etc.) (or the probability of the target audience investing virtual resources in the promotional information within the target time period), and the amount of virtual resources invested by the target audience. Whether the target audience will invest virtual resources in the promotional information within the target time period is a binary classification result, i.e., "yes" or "no," "1" or "0." The probability of the target audience investing virtual resources in the promotional information within the target time period is a probability value between 0 and 1, representing the likelihood of the target audience investing virtual resources (e.g., 0.8 indicates an 80% probability of investing virtual resources). If the probability value is higher than the investment threshold, it means that the target audience will invest virtual resources in the information to be promoted; if the probability value is lower than or equal to the preset threshold, it means that the target audience will not invest virtual resources in the information to be promoted.

[0121] It is important to note that if the estimated promotion result indicates that the target audience will not invest virtual resources in the information to be promoted within the target time period, or if the probability of the target audience investing virtual resources in the information to be promoted within the target time period is less than the investment threshold, the value assessment model may not output the amount of virtual resources invested, or may only provide a default value (such as 0). For example, if the investment probability of the target audience is 0.2 (less than the preset threshold of 0.5), the amount of virtual resources will default to 0.

[0122] S407: The DSP side determines the promotional value of the information to be promoted based on the estimated promotional results and promotional costs of the target audience.

[0123] Specifically, firstly, based on the estimated promotion results of the target audience (i.e., the probability of investment and the amount of virtual resources invested), the probability of investment P and the amount of virtual resources invested V are multiplied to obtain the price adjustment coefficient F. For example, if the probability P of the target audience making a payment within a specific time period is 0.8, and the amount of virtual resources invested V is 1000, then the price adjustment coefficient F is 800.

[0124] Secondly, the adjusted revenue per thousand impressions (CPM) is determined based on the price adjustment factor and the revenue per thousand impressions. Specifically, assuming the media-provided CPM is CPM, then the adjusted CPM' = CPM × F. For example, assuming the media-provided CPM is 10 yuan and the price adjustment factor F is 800, then the adjusted CPM' would be 8000 yuan.

[0125] Subsequently, the promotion value is determined based on the adjusted revenue per thousand impressions and promotion cost. Specifically, assuming the promotion cost is C, the promotion value can be represented by the following formula (4):

[0126] Price = 0.001CPM' + C (4)

[0127] For example, if the promotion cost C is 50 yuan and the adjusted revenue per thousand impressions CPM' is 8000 yuan, then the promotion value Price is 8 + 50 = 58 yuan.

[0128] S408: The DSP side sends the promotional value to the media side through the real-time interface API, so that the media side can promote the information based on the promotional value.

[0129] A Real-Time API (RTA) is a technology that allows media outlets to obtain key strategic information (such as promotional value) by requesting the DSP's real-time API during the ad display or placement decision-making process. This real-time interface allows media outlets to obtain key strategic information, such as promotional value, from the DSP's real-time API every time they make an ad display or placement decision.

[0130] Understandably, RTA allows media outlets to obtain the latest advertising strategy information in real time with each ad request, ensuring the timeliness and accuracy of ad delivery. Furthermore, by assessing the activity level or other relevant attributes of the target audience in real time, it avoids duplicate targeting of previously promoted audiences or other specific groups. Moreover, compared to traditional methods (such as manually uploading audience packages), the RTA model reduces the storage and transmission of sensitive information, enhancing privacy protection.

[0131] In summary, this application discloses an information promotion method. This method utilizes front-end data (including target object data, semantic features, and virtual resource quantity) for model training, effectively avoiding limitations caused by back-end data delays or missing data. This not only improves the accuracy of the predicted promotion results output by the value prediction model but also significantly enhances the effectiveness of information promotion. Furthermore, this application allows the media side to obtain the latest advertising strategy information in real time with each advertising request through the RTA mode, ensuring the timeliness and accuracy of advertising delivery, which also significantly enhances the effectiveness of information promotion.

[0132] Based on the model training method provided in the preceding embodiments, this application also provides a corresponding model training apparatus. The model training apparatus provided in this application will be described in detail below:

[0133] See Figure 6 The figure is a schematic diagram of a model training device provided in an embodiment of this application. The model training device 600 includes: an acquisition module 601, a construction module 602, and a training module 603.

[0134] The acquisition module 601 is used to acquire target object data, semantic features of the target object's evaluation of the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted.

[0135] The construction module 602 is used to construct training labels based on target object data, semantic features and virtual resource quantity. The training labels are used to identify whether the target object is the information to be promoted and to identify the amount of virtual resources invested in the target object for the information to be promoted.

[0136] Training module 603 is used to train a machine learning model based on training labels to obtain a value prediction model.

[0137] In some specific implementations, the training label includes a first sub-label and a second sub-label. The first sub-label is used to identify the probability of the target object investing virtual resources in the information to be promoted, and the second sub-label is used to identify the amount of virtual resources invested by the target object in the information to be promoted.

[0138] The training module 603 is specifically used for: inputting target object data and semantic features into a machine learning model, training the machine learning model using a first sub-label to identify the probability of the target object investing virtual resources in the information to be promoted, and a second module that uses a second sub-label to train the machine learning model to identify the amount of virtual resources invested by the target object in the information to be promoted.

[0139] In some specific implementations, the model training device 600 also includes: a test acquisition module, a test input module, a loss acquisition module, and a parameter update module;

[0140] The test acquisition module is used to acquire test object data, semantic features of the test object's evaluation of the information to be promoted, and the actual amount of virtual resources invested by the test object in the information to be promoted.

[0141] The test input module is used to input the test object data and the semantic features of the test object to be promoted into the value prediction model to obtain the predicted probability of the test object investing virtual resources in the information to be promoted and the predicted amount of virtual resources.

[0142] The loss acquisition module is used to determine the loss value of the value estimation model based on the predicted amount of virtual resources, the actual amount of virtual resources, and the loss function of the value estimation model if the predicted investment probability characterization test object will invest virtual resources for the information to be promoted.

[0143] The parameter update module is used to update the model parameters of the value prediction model based on the loss value.

[0144] In some specific implementations, the loss acquisition module is specifically used to: obtain the mean loss value of the value prediction function by calculating the mean loss function of the value prediction model based on the predicted virtual resource quantity and the actual virtual resource quantity; and obtain the variance loss value of the value prediction function by calculating the variance loss function of the value prediction model; if the variance loss value is greater than the variance loss threshold, then the mean loss value and the variance loss threshold are weighted and summed to obtain the loss value of the value prediction model.

[0145] In some specific implementations, the construction module 602 includes: a first construction submodule, a second construction submodule, and a third construction submodule;

[0146] The first construction submodule is used to select temporal features from target object data and semantic features;

[0147] The second construction submodule is used to perform weighted processing on temporal features based on attention weights, where attention weights are negatively correlated with time difference, and time difference is the difference between the current time and the time corresponding to the temporal feature.

[0148] The third submodule is used to construct training labels based on the weighted temporal features and the amount of virtual resources.

[0149] In some specific implementations, the construction module 602 includes: a fourth construction submodule and a fifth construction submodule;

[0150] The fourth submodule is used to filter target object data, semantic features, and virtual resource quantity within the target time period;

[0151] The fifth submodule is used to construct training labels based on the target object data, semantic features, and virtual resource quantity within the target time period. The training labels are used to identify whether the target object is the information to be promoted and to invest virtual resources within the target time period, and to identify the amount of virtual resources invested by the target object as the information to be promoted within the target time period.

[0152] In summary, this application discloses a model training device that utilizes front-end data (including target object data, semantic features, and virtual resource quantity) for model training, effectively avoiding limitations caused by back-end data delays or missing data. This not only improves the accuracy of the predicted promotion results output by the value prediction model but also significantly enhances the effectiveness of information promotion.

[0153] Based on the information promotion method provided in the foregoing embodiments, this application also provides an information promotion device. The information promotion device provided in this application will be described in detail below:

[0154] See Figure 7 The figure is a schematic diagram of an information promotion device provided in an embodiment of this application. The information promotion device 700 includes: a feature acquisition module 701, a result acquisition module 702, and a value determination module 703.

[0155] The feature acquisition module 701 is used to acquire the data of the object to be promoted and the semantic features of the object's evaluation of the promotion information;

[0156] The result acquisition module 702 is used to determine the estimated promotion results corresponding to the semantic features of the data of the object to be promoted and the evaluation of the information to be promoted by the object to be promoted through the value prediction model. The estimated promotion results include the probability of the object to be promoted investing virtual resources in the information to be promoted and the amount of virtual resources invested by the object to be promoted in the information to be promoted. The value prediction model is trained by the model training method.

[0157] The value determination module 703 is used to determine the promotion value of the information to be promoted based on the estimated promotion results and promotion costs.

[0158] In some specific implementations, the information promotion device 700 is applied to the demand-side platform. The information promotion device 700 also includes a value delivery module.

[0159] The value delivery module is used to send the promotional value of the information to be promoted to the media through a real-time interface.

[0160] In summary, this application discloses an information promotion device. This application utilizes front-end data (including target object data, semantic features, and virtual resource quantity) for model training, effectively avoiding limitations caused by back-end data delays or missing data. This not only improves the accuracy of the predicted promotion results output by the value prediction model but also significantly enhances the effectiveness of information promotion. Furthermore, this application allows the media side to obtain the latest advertising strategy information in real time with each advertising request through the RTA mode, ensuring the timeliness and accuracy of advertising delivery, which also significantly enhances the effectiveness of information promotion.

[0161] See Figure 8 , Figure 8 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. For example... Figure 8 As shown, for ease of explanation, only the parts related to the embodiments of this application are shown. For specific technical details not disclosed, please refer to the method section of the embodiments of this application. The terminal can be any terminal device including mobile phones, tablets, personal digital assistants, point-of-sale (POS) terminals, in-vehicle computers, etc. Taking a computer as an example:

[0162] Figure 8 This is a block diagram illustrating a portion of the structure of a computer associated with the terminal provided in an embodiment of this application. (Reference) Figure 8 The computer includes: a radio frequency (RF) circuit 1210, a memory 1220, an input unit 1230 (including a touch panel 1231 and other input devices 1232), a display unit 1240 (including a display panel 1241), a sensor 1250, an audio circuit 1260 (which can connect to a speaker 1261 and a microphone 1262), a wireless fidelity (WiFi) module 1270, a processor 1280, and a power supply 1290, etc. Those skilled in the art will understand that... Figure 8 The computer architecture shown does not constitute a limitation on the computer and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0163] The memory 1220 can be used to store software programs and modules. The processor 1280 executes various computer functions and data processing by running the software programs and modules stored in the memory 1220. The memory 1220 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs 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 computer (such as audio data, telephone directory, etc.). In addition, the memory 1220 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.

[0164] The processor 1280 is the control center of the computer, connecting various parts of the computer through various interfaces and lines. It performs various computer functions and processes data by running or executing software programs and / or modules stored in the memory 1220, and by calling data stored in the memory 1220. Optionally, the processor 1280 may include one or more processing units; preferably, the processor 1280 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 also not be integrated into the processor 1280.

[0165] In this embodiment of the application, the processor 1280 included in the terminal also has the following functions:

[0166] Acquire target object data, semantic features of the target object's evaluation of the promotional information, and the amount of virtual resources invested by the target object in the promotional information;

[0167] Based on the target object data, semantic features, and virtual resource quantity, training labels are constructed. The training labels are used to identify whether the target object is the information to be promoted that has received virtual resources, and to identify the amount of virtual resources that the target object has received for the information to be promoted.

[0168] A machine learning model is trained based on the training labels to obtain a value prediction model;

[0169] or,

[0170] Obtain semantic features of the target audience's data and their evaluation of the promotional information;

[0171] The value prediction model determines the predicted promotion results corresponding to the semantic features of the data of the target object and the evaluation of the target object on the information to be promoted. The predicted promotion results include the probability of the target object investing virtual resources in the information to be promoted and the amount of virtual resources invested by the target object in the information to be promoted. The value prediction model is trained by the model training method.

[0172] The promotional value of the information to be promoted is determined based on the estimated promotion results and promotion costs.

[0173] See Figure 9 , Figure 9 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 1300 can vary significantly due to different configurations or performance, and may include one or more central processing units (CPUs) 1322 (e.g., one or more processors) and memory 1332, and one or more storage media 1330 (e.g., one or more mass storage devices) for storing application programs 1342 or data 1344. The memory 1332 and storage media 1330 can be temporary or persistent storage. The program stored in the storage media 1330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server. Furthermore, the CPU 1322 may be configured to communicate with the storage media 1330 and execute the series of instruction operations in the storage media 1330 on the server 1300.

[0174] Server 1300 may also include one or more power supplies 1326, one or more wired or wireless network interfaces 1350, one or more input / output interfaces 1358, and / or one or more operating systems, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0175] The steps performed by the server in the above embodiments can be based on this Figure 9 The server structure shown.

[0176] CPU1322 is used to perform the following steps:

[0177] Acquire target object data, semantic features of the target object's evaluation of the promotional information, and the amount of virtual resources invested by the target object in the promotional information;

[0178] Based on the target object data, semantic features, and virtual resource quantity, training labels are constructed. The training labels are used to identify whether the target object is the information to be promoted that has received virtual resources, and to identify the amount of virtual resources that the target object has received for the information to be promoted.

[0179] A machine learning model is trained based on the training labels to obtain a value prediction model;

[0180] or,

[0181] Obtain semantic features of the target audience's data and their evaluation of the promotional information;

[0182] The value prediction model determines the predicted promotion results corresponding to the semantic features of the data of the target object and the evaluation of the target object on the information to be promoted. The predicted promotion results include the probability of the target object investing virtual resources in the information to be promoted and the amount of virtual resources invested by the target object in the information to be promoted. The value prediction model is trained by the model training method.

[0183] The promotional value of the information to be promoted is determined based on the estimated promotion results and promotion costs.

[0184] This application also provides a computer-readable storage medium for storing a computer program that executes any one of the implementation methods of a model training method or an information dissemination method described in the foregoing embodiments.

[0185] This application also provides a computer program product or computer program, 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 any one of the implementation methods of the model training method or product promotion method described in the foregoing embodiments.

[0186] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0187] 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.

[0188] 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.

[0189] 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.

[0190] 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 personal 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 computer programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0192] 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 model training method, characterized in that, The method includes: Acquire target object data, semantic features of the target object's evaluation of the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted; Based on the target object data, the semantic features, and the amount of virtual resources, training labels are constructed, wherein the training labels are used to identify whether the target object has invested virtual resources in the information to be promoted, and to identify the amount of virtual resources invested by the target object in the information to be promoted; The machine learning model is trained based on the training labels to obtain the value prediction model.

2. The method according to claim 1, characterized in that, The training label includes a first sub-label and a second sub-label. The first sub-label is used to identify the probability that the target object invests virtual resources in the information to be promoted, and the second sub-label is used to identify the amount of virtual resources that the target object invests in the information to be promoted. Training the machine learning model based on the training labels includes: By inputting the target object data and the semantic features into the machine learning model, the machine learning model is trained using the first sub-label to identify the probability of the target object investing virtual resources in the information to be promoted, and a second module is trained using the second sub-label to identify the amount of virtual resources invested by the target object in the information to be promoted.

3. The method according to claim 1, characterized in that, The method further includes: Acquire test object data, semantic features of the test object's evaluation of the information to be promoted, and the actual amount of virtual resources invested by the test object in the information to be promoted; The test object data and the semantic features of the test object to be promoted information are input into the value prediction model to obtain the predicted probability of the test object investing virtual resources in the information to be promoted and the predicted amount of virtual resources. If the predicted investment probability indicates that the test object will invest virtual resources in the information to be promoted, then the loss value of the value prediction model is determined based on the predicted virtual resource amount, the actual virtual resource amount, and the loss function of the value prediction model. The model parameters of the value prediction model are updated based on the loss value.

4. The method according to claim 3, characterized in that, The step of determining the loss value of the value estimation model based on the predicted virtual resource quantity, the actual virtual resource quantity, and the loss function of the value estimation model includes: Based on the predicted virtual resource quantity and the actual virtual resource quantity, the mean loss value of the value prediction function is obtained by calculating the mean loss function of the value prediction model, and the variance loss value of the value prediction function is obtained by calculating the variance loss function of the value prediction model. If the variance loss value is greater than the variance loss threshold, then the mean loss value and the variance loss threshold are weighted and summed to obtain the loss value of the value prediction model.

5. The method according to claim 1, characterized in that, The step of constructing training labels based on the target object data, the semantic features, and the virtual resource quantity includes: Select temporal features from the target object data and the semantic features; The temporal features are weighted according to attention weights, wherein the attention weights are negatively correlated with the time difference, and the time difference is the difference between the current time and the time corresponding to the temporal feature. Training labels are constructed based on the weighted temporal features and the amount of virtual resources.

6. The method according to claim 1, characterized in that, The step of constructing training labels based on the target object data, the semantic features, and the amount of virtual resources, wherein the training labels are used to identify whether the target object has invested virtual resources in the information to be promoted, and to identify the amount of virtual resources invested by the target object in the information to be promoted, including: Filter the target object data, the semantic features, and the virtual resource quantity within a target time period; Based on the target object data, semantic features, and virtual resource quantity within the target time period, training labels are constructed, wherein the training labels are used to identify whether the target object has invested virtual resources in the information to be promoted within the target time period, and to identify the amount of virtual resources invested by the target object in the information to be promoted within the target time period.

7. An information promotion method, characterized in that, The method includes: Obtain semantic features of the target audience's data and their evaluation of the promotional information; The value prediction model determines the predicted promotion result corresponding to the semantic features of the data of the object to be promoted and the evaluation of the information to be promoted by the object to be promoted. The predicted promotion result includes the probability of the object to be promoted investing virtual resources in the information to be promoted and the amount of virtual resources invested by the object to be promoted in the information to be promoted. The value prediction model is trained by the model training method described in any one of claims 1 to 6. Based on the estimated promotion results and promotion costs, the promotion value of the information to be promoted is determined.

8. The method according to claim 7, characterized in that, The method is applied to the demand-side platform, and the method further includes: The promotional value of the information to be promoted is sent to the media through a real-time interface.

9. A model training device, characterized in that, The device includes: an acquisition module, a construction module, and a training module; The acquisition module is used to acquire target object data, semantic features of the target object's evaluation of the information to be promoted, and the amount of virtual resources invested by the target object in the information to be promoted. The construction module is used to construct training labels based on the target object data, the semantic features, and the amount of virtual resources. The training labels are used to identify whether the target object has invested virtual resources in the information to be promoted, and to identify the amount of virtual resources invested by the target object in the information to be promoted. The training module is used to train the machine learning model based on the training labels to obtain a value prediction model.

10. An information promotion device, characterized in that, The information promotion device includes: a feature acquisition module, a result acquisition module, and a value determination module; The feature acquisition module is used to acquire the data of the object to be promoted and the semantic features of the object's evaluation of the promotion information. The result acquisition module is used to determine the estimated promotion result corresponding to the semantic features of the data of the object to be promoted and the evaluation of the information to be promoted by the object to be promoted through the value prediction model. The estimated promotion result includes the probability of the object to be promoted investing virtual resources in the information to be promoted and the amount of virtual resources invested by the object to be promoted in the information to be promoted. The value prediction model is trained by the model training method according to any one of claims 1 to 6. The value determination module is used to determine the promotion value of the information to be promoted based on the estimated promotion results and promotion costs.

11. A computer device, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the steps of the model training method according to any one of claims 1 to 6, or the information propagation method according to claim 7 or 8, according to the instructions in the program code.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for performing the steps of the model training method of any one of claims 1 to 6, or the information promotion method of claim 7 or 8.

Images