Advertisement budget allocation method and system based on machine learning and user behaviors

Abstract

The invention discloses an advertisement budget distribution method and system based on machine learning and user behaviors, and the method comprises the following steps: collecting the historical behavior data of a user, and extracting the feature data of the historical behavior data through feature engineering; constructing a user probability prediction model, and training the model by using the feature data; in an advertisement putting process, collecting real-time behavior data of a user, inputting the real-time behavior data into the trained model, and predicting an advertisement clicking probability of the user and an advertisement commodity purchasing probability of the user; a fusion model is constructed, the fusion model calculates the conversion rate of the user according to the advertisement clicking probability of the user and the advertisement commodity purchasing probability of the user, and calculates the channel conversion rate according to the putting channel; distributing a preliminary advertisement budget according to the historical data of the putting channel; and dynamically adjusting the preliminary advertisement budget of the putting channel based on the channel conversion rate. According to the invention, precise distribution of the advertisement budget can be realized, and the overall conversion effect of the advertisement is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of advertisement delivery, and in particular relates to an advertisement budget allocation method and system based on machine learning and user behavior. Background Art

[0002] With the rapid development of the Internet and mobile Internet, the advertising industry has also undergone tremendous changes. The forms and delivery channels of digital advertising are becoming increasingly diversified, including social media, search engines, display ads, video ads, etc. When choosing delivery channels, advertisers hope to maximize advertising effects through reasonable budget allocation strategies, thereby increasing advertising exposure, click-through rate and conversion rate. However, in the face of numerous channels and complex user behavior data, traditional manual or rule-driven budget allocation methods have been unable to adapt to changing market demands and real-time changing user behavior patterns.

[0003] Existing advertising budget allocation methods usually rely on historical experience data or simple rule settings, which makes it difficult to accurately predict user behavior and channel conversion rates. These methods cannot dynamically adjust budget allocation during the advertising process, and thus cannot fully adapt to the rapid changes in user behavior, resulting in poor advertising effects and waste of resources. In modern Internet advertising, how to efficiently use advertising budgets and improve advertising conversion rates has become a key issue of concern to advertisers and platforms.

[0004] A Chinese invention patent application with publication number CN118278990A discloses a method, device, electronic device and medium for allocating an advertising delivery budget, and its implementation scheme is: obtaining a total budget for advertising delivery; allocating the total budget to multiple objects at a first ratio within a first time period to deliver advertising to the multiple objects; obtaining target indicators corresponding to the multiple objects within the first time period, wherein the target indicators include the consumed budgets of the multiple objects; in response to a preset condition, predicting a predicted ratio for allocating the remaining budget for the next time period for the multiple objects based on time information corresponding to the next time period, wherein the remaining budget is the difference between the total budget and the budget consumed in the first time period; and determining a second ratio for allocating the remaining budget for the multiple objects in the next time period based on the predicted ratio.

[0005] However, this solution only relies on time periods and budget consumption to adjust budget allocation, without fully considering individual user behavior data; in actual advertising, the behavior patterns of different user groups vary significantly, and it is difficult to distinguish these differences, which may lead to insufficient accuracy in budget allocation. In addition, the budget allocation of this solution is mainly based on the consumption data and remaining budget of the time period, which cannot accurately predict the user's click and conversion behavior; it cannot help advertisers better identify high-conversion channels and achieve more refined budget allocation. Moreover, there is no optimization for different delivery channels, and the budget allocation is adjusted only based on the overall budget consumption, which cannot achieve fine control of different channels. Summary of the invention

[0006] The application of machine learning technology has brought new solutions to advertising budget allocation. Through the application of machine learning models, features can be extracted from a large amount of historical user behavior data, the user's potential behavior can be predicted, and the advertising budget can be dynamically adjusted according to real-time data. The present invention provides an advertising budget allocation method and system based on machine learning and user behavior, aiming to solve the problems of insufficient budget allocation accuracy, inability to achieve refined budget allocation, and inability to achieve refined control of different channels in the prior art.

[0007] To solve the above technical problems, the method proposed by the present invention comprises the following steps: Collecting historical behavior data of users, and extracting feature data of the historical behavior data through feature engineering; Constructing a user probability prediction model, and using the feature data to train the user probability prediction model; During the advertising process, real-time user behavior data is collected and input into a trained user probability prediction model to predict the probability of a user clicking on an ad and the probability of a user purchasing an advertised product. Constructing a fusion model, wherein the fusion model calculates the user conversion rate based on the probability of the user clicking on the advertisement and the probability of the user purchasing the advertised product, and calculates the channel conversion rate based on the delivery channel; Allocate preliminary advertising budget based on historical data of delivery channels; Dynamically adjust the initial advertising budget of the delivery channel based on the channel conversion rate.

[0008] The calculation method of the channel conversion rate is as follows:

[0009] In the formula, Indicates Channel conversion rate of each channel, Represents the total number of users in all channels, Indicates The number of users of each channel, Indicates Channel The conversion rate of each user.

[0010] Before the feature engineering, the historical behavior data is cleaned, deduplicated, and missing value filled.

[0011] Preferably, the user probability prediction model is a gated recurrent unit model, and the output layer of the gated recurrent unit model includes two independent neurons, which are used to predict the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product, respectively. The structure of the output layer is expressed as follows:

[0012] In the formula, is the probability of a user clicking on an ad predicted by the gated recurrent unit model, is the probability of a user purchasing an advertised product predicted by the gated recurrent unit model, is the weight matrix corresponding to the user clicking on the ad, is the weight matrix corresponding to the user purchasing the advertised product, is the bias term for the user click ad prediction task, is the bias item for the task of predicting whether users purchase advertised products. is the final hidden state after the update gate and reset gate output in the gated recurrent unit model.

[0013] Preferably, the user probability prediction model is two lightweight gradient boosting machine models, which are used to predict the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product, respectively. The two probabilities are expressed as follows:

[0014]

[0015] In the formula, is the probability of a user clicking on an ad predicted by the lightweight gradient boosting model, is the probability of a user purchasing an advertised product predicted by the lightweight gradient boosting model, is the logarithmic probability output by the lightweight gradient boosting machine model in the user click ad prediction task, It is the logarithmic probability output by the lightweight gradient boosting machine model in the task of predicting whether a user will purchase an advertised product.

[0016] Preferably, the user probability prediction model is a random forest model, and the output of the random forest model is expressed as:

[0017]

[0018] In the formula, is the probability of a user clicking on an ad predicted by the random forest model, is the probability of a user purchasing an advertised product predicted by the random forest model, is the number of trees in the random forest model, The random forest model The probability of users clicking on ads predicted by Keshu, The random forest model The probability of a user purchasing an advertised product predicted by Keshu.

[0019] Preferably, the user probability prediction model is a hybrid expert model based on multiple gates, including an input layer, an expert network layer, a gated network layer, a task tower layer and an output layer. The structure and function of the model are as follows: In the input layer, the sequence representation is extracted from the user's real-time behavior data, and the user's static features and sequences are concatenated into a complete model input; In the expert network layer, multiple expert networks are defined. Each expert is responsible for extracting a set of common features. Each expert is a fully connected network or a complex network, and outputs a common high-dimensional feature representation. In the gated network layer, two gated networks are constructed to correspond to the probability prediction of users clicking on ads and the probability prediction of users purchasing advertised products. One of them corresponds to the two prediction tasks, and the contribution weight of each expert is dynamically allocated. In the task tower layer, each task has an independent tower. The task tower for predicting the probability of users clicking on ads is used to optimize the cross entropy loss of click-through rate prediction, and the task tower for predicting the probability of users purchasing advertised products is used to optimize the cross entropy loss of purchase probability prediction. The output layer outputs two probability values: the probability that a user clicks on an ad and the probability that a user purchases the advertised product.

[0020] Preferably, the conversion rate is calculated as follows:

[0021] In the formula, To calculate the conversion rate of the corresponding user, It is a preset weighting coefficient that determines the contribution of click probability and purchase probability to the final conversion rate. is the probability that a user clicks on an ad, is the probability that the user purchases the advertised product.

[0022] Preferably, the conversion rate is calculated as follows:

[0023] In the formula, Indicates that at a given click probability and purchase probability Under the condition of Likelihood distribution, Indicates that at a given conversion rate Under the condition of, the probability of users clicking on ads is, Indicates that at a given conversion rate Under the condition of, the probability of users purchasing the advertised product is, is the prior probability, indicating the user's conversion rate The distribution of , estimated using historical data or expert knowledge, is the marginal probability of click probability, which represents the probability distribution of all users clicking ads. is the marginal probability of purchase probability, which represents the probability distribution of all users purchasing the advertised product.

[0024] Preferably, the dynamic adjustment of the initial advertising budget of the delivery channel is adjusted based on the conversion rate difference, and the adjustment method is: For each channel, calculate the difference between the channel conversion rate of the corresponding channel and the preset channel conversion rate when allocating the initial advertising budget:

[0025] If the channel conversion rate of a channel is higher than the preset conversion rate, the advertising budget of the group is increased; if it is lower than the preset conversion rate, the advertising budget of the channel is reduced:

[0026] In the formula, It is the adjustment coefficient used to control the sensitivity of budget adjustment.

[0027] Another aspect of the present invention further provides an advertising budget allocation system based on machine learning and user behavior, wherein the system is used to implement the above method, and comprises: Data collection module, used to collect historical behavior data and real-time behavior data of users; The data processing module processes the data submitted by the data collection module. The processing methods include cleaning, deduplication and missing value filling. A probability prediction model receives the processed real-time behavior data and predicts the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product, wherein the probability prediction model is trained using the processed historical behavior data; The fusion model receives the probability of users clicking on ads and the probability of users purchasing advertised products output by probability prediction, calculates the user conversion rate, and calculates the channel conversion rate based on the delivery channel; The budget allocation module is used to allocate preliminary advertising budgets based on the historical data of the delivery channels; and dynamically adjust the preliminary advertising budgets of the delivery channels based on the channel conversion rate.

[0028] Compared with the prior art, the present invention has the following technical effects: 1. The advertising budget allocation method proposed in the present invention collects and analyzes historical behavior data of users by introducing machine learning technology, builds a probability prediction model for users to click on advertisements and purchase goods, and achieves a refined understanding of user behavior. This enables advertising budgets to be more accurately allocated to users who are likely to convert, improves the personalization and accuracy of advertising delivery, and achieves precise allocation of advertising budgets.

[0029] 2. The fusion model constructed by the advertising budget allocation method proposed in the present invention can comprehensively consider the probability of users clicking on advertisements and purchasing products, thereby calculating the user conversion rate, and further calculating the channel conversion rate based on the data of the delivery channel. This calculation of channel conversion rate can identify advertising delivery channels with high conversion rates, so that the budget can be more efficiently allocated to channels with better effects, thereby improving the overall conversion effect of advertising.

[0030] 3. The advertising budget allocation method proposed in the present invention extracts features from historical data and predicts user behavior, and no longer relies on simple rules or experience. With the help of data-driven prediction models, advertisers can intelligently allocate budgets based on user behavior trends and channel effects, achieve scientific and intelligent decision-making, and avoid the problem of strong subjectivity in traditional methods.

[0031] 4. In the advertising budget allocation method proposed in the present invention, the advertising budget is not only initially allocated based on the channel conversion rate, but can also be dynamically adjusted according to real-time feedback during the delivery process. This flexible budget control method enables advertisers to quickly optimize budget allocation according to the effect, improves the controllability and flexibility of the advertising budget, and ensures that the funds can maximize the improvement of advertising effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the process of the method of the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in combination with specific embodiments of the present application and with reference to the accompanying drawings.

[0034] Embodiment 1 With the accumulation of online user behavior data, especially the continuous growth of user click, browse, purchase and other behavior data, advertising platforms gradually rely on accurate user portraits and data analysis methods to optimize advertising delivery strategies and maximize the efficiency of advertising budget utilization. In addition, the application of machine learning technology has brought new solutions to advertising budget allocation. Through the application of machine learning models, features can be extracted from a large amount of historical user behavior data, predicting the user's potential behavior, and dynamically adjusting the advertising budget based on real-time data.

[0035] To this end, this embodiment is an advertising budget allocation method based on machine learning and user behavior, such as Figure 1 As shown, the following steps are included: Collecting historical behavior data of users, and extracting feature data of the historical behavior data through feature engineering; Constructing a user probability prediction model, and using the feature data to train the user probability prediction model; During the advertising process, real-time user behavior data is collected and input into a trained user probability prediction model to predict the probability of a user clicking on an ad and the probability of a user purchasing an advertised product. Constructing a fusion model, wherein the fusion model calculates the user conversion rate based on the probability of the user clicking on the advertisement and the probability of the user purchasing the advertised product, and calculates the channel conversion rate based on the delivery channel; Allocate preliminary advertising budget based on historical data of delivery channels; Dynamically adjust the initial advertising budget of the delivery channel based on the channel conversion rate.

[0036] In the data collection step, historical user behavior data can be collected from a variety of sources, including ad clicks, browsing history, purchase history, dwell time, bounce rate, etc., and social interaction data, device usage data, etc. During collection, ensure that each user behavior data record has an accurate timestamp to capture the time series characteristics of user behavior.

[0037] Before the feature engineering, the historical behavior data is cleaned, deduplicated, and missing value filled to remove duplicate, missing, and abnormal data to ensure data integrity and consistency. The numerical features are normalized to make the data range consistent and facilitate model training. Then, features related to user probability prediction are extracted from the original data through feature engineering. For example, user click behavior is converted into time series data.

[0038] In one embodiment of the present invention, the user probability prediction model is a gated recurrent unit (GRU) model. When using a gated recurrent unit, first collect the user's click behavior, browsing behavior, purchase behavior, dwell time, bounce rate and other data, these data will form a time series over time, for example, for a certain user, collect the time series features of each browsing and clicking of the advertisement.

[0039] For numerical features (such as the number of clicks, dwell time, number of views, etc.), they can be directly input into the model, but they need to be normalized or standardized so that each feature value is within a similar numerical range to prevent the feature difference from affecting the model learning. This embodiment normalizes the numerical features and scales the data to between 0 and 1. For categorical data (such as user device type, advertisement type, etc.), it needs to be encoded first and converted into a vector representation to make it suitable for gated recurrent unit processing. The encoding of categorical data can use one-hot encoding or embedded layer encoding.

[0040] When using gated recurrent units for prediction, the user's behavior data also needs to be converted into a format suitable for time series model processing.

[0041] In the gated recurrent unit, the update gate and reset gate control the flow of information. Output Calculate by following the steps below: Update Gate The update gate controls the information flow from the previous moment to the current moment. The calculation formula is:

[0042] in, It is the output of the update gate, which determines how much information is passed from the previous moment to the current moment; It is the input feature at the current moment, such as the user's current behavior data; is the hidden state of the previous moment; is the learning parameter; is the Sigmoid activation function, which is used to limit the output to the range [0,1].

[0043] Reset Gate The reset gate controls the degree to which the current input is combined with the memory of the previous moment. The calculation formula is:

[0044] in, It is the output of the reset gate, which determines the influence of the hidden state at the previous moment on the current moment; are the learned parameters.

[0045] Candidate hidden states Use the reset gate to reset the hidden state of the previous moment, and then generate the candidate hidden state of the current moment :

[0046] in, is a candidate hidden state, Represents the hidden state of the previous moment after the reset gate is applied, and ⊙ is element-wise multiplication.

[0047] Final hidden state The fusion ratio between the candidate hidden state and the hidden state at the previous moment is determined by the update gate to obtain the final hidden state at the current moment. :

[0048] in, is the hidden state at the current moment and the memory of the model.

[0049] In this embodiment, in order to output two predicted probabilities: the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product, the output layer of the gated recurrent unit model described in this embodiment is designed with two independent neurons, which are respectively used to predict the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product.

[0050] The output layer structure of the gated recurrent unit in this embodiment is expressed as:

[0051] In the formula, is the probability of a user clicking on an ad predicted by the gated recurrent unit model, is the probability of a user purchasing an advertised product predicted by the gated recurrent unit model, is the weight matrix corresponding to the user clicking on the ad, is the weight matrix corresponding to the user purchasing the advertised product, is the bias term for the user click ad prediction task, is the bias item for the task of predicting whether users purchase advertised products. is the final hidden state after the update gate and reset gate output in the gated recurrent unit model.

[0052] The gated recurrent unit model can effectively predict user behavior (such as clicks, purchases, etc.) in future advertising by learning the temporal characteristics of user behavior. Compared with traditional feature engineering methods, the gated recurrent unit model can automatically extract deep patterns and rules from time series and adapt to complex dynamic changes in user behavior.

[0053] In addition, by designing the output layer of the gated recurrent unit model to have two output nodes, predicting the click probability and purchase probability respectively, the two tasks of click-through rate prediction and purchase prediction can be solved at the same time. This multi-task learning method can not only improve the accuracy of the model, but also provide more accurate personalized recommendations and budget optimization strategies in the actual advertising process.

[0054] In other embodiments of the present invention, the user probability prediction model uses two Light Gradient Boosting Machine (LightGBM) models, which are used to predict the probability of a user clicking on an ad and the probability of a user purchasing an advertised product. When using the Light Gradient Boosting Machine model, there are two prediction targets: the probability of clicking an ad and the probability of a user purchasing an advertised product. , the probability of purchasing the advertised product The goal of each prediction task is to predict whether a user will click on an ad or buy an advertised product based on historical user behavior data (such as clicks, views, dwell time, purchase behavior, and other features). The output of each task is a probability value, which indicates the likelihood of the event occurring.

[0055] For each model, the lightweight gradient boosting machine trains a set of weak learners (i.e., decision trees) through gradient boosting decision trees (GBDT), where each tree is fitted on the residual of the previous tree (i.e., the difference between the actual value and the predicted value). Each tree divides the data by maximizing the information gain, and finally forms a tree that can minimize the error. The final prediction result is obtained by weighted averaging the outputs of these weak learners.

[0056] In this embodiment, the two probabilities output by the two lightweight gradient boosting machines are expressed as follows:

[0057]

[0058] In the formula, is the probability of a user clicking on an ad predicted by the lightweight gradient boosting model, is the probability of a user purchasing an advertised product predicted by the lightweight gradient boosting model, is the logarithmic probability output by the lightweight gradient boosting machine model in the user click ad prediction task, It is the logarithmic probability output by the lightweight gradient boosting machine model in the task of predicting whether a user will purchase an advertised product.

[0059] The lightweight gradient boosting machine model is a model based on the gradient boosting decision tree, which can efficiently process large-scale data and is suitable for scenarios in which nonlinear relationships need to be captured in this embodiment, and can improve the accuracy of probability prediction.

[0060] In other embodiments of the present invention, the user probability prediction model uses a random forest model, and prediction is performed by constructing multiple decision trees. Each tree makes a prediction on the data, and finally the prediction results of all trees are integrated to obtain the final prediction. For binary classification problems, the output of each tree of the random forest is a category label (0 or 1). The random forest will eventually give a classification label based on the voting results of all trees. This embodiment requires a predicted probability value, and the output probability of each tree is used to calculate the average value of the probabilities output by all trees.

[0061] When predicting, each decision tree will output a predicted label (0 or 1) based on the input user features. Since this embodiment requires prediction probability, random forest will calculate the prediction probability of each tree instead of directly giving a label. For each tree, its output is the probability of click or purchase, which can be calculated by the frequency of the leaf nodes of the tree, that is, the proportion of samples belonging to the click or purchase category on the leaf nodes.

[0062] Assume that each tree The output is the probability of clicking the ad and the probability of purchasing the advertised product , then for the final output of the random forest, we need to average the output probabilities of all trees to get the final probability value. Specifically, assuming there is The final output probability of the random forest is the average of all tree output probabilities, expressed as:

[0063]

[0064] In the formula, is the probability of a user clicking on an ad predicted by the random forest model, is the probability of a user purchasing an advertised product predicted by the random forest model, is the number of trees in the random forest model, The random forest model The probability of users clicking on ads predicted by Keshu, The random forest model The probability of a user purchasing an advertised product predicted by Keshu.

[0065] Random forests can effectively handle user behavior prediction problems and improve the stability and accuracy of predictions by integrating the outputs of multiple decision trees.

[0066] In other embodiments of the present invention, the user probability prediction model is a hybrid expert model based on multiple gates, including an input layer, an expert network layer, a gated network layer, a task tower layer and an output layer. The structure and function of the model are as follows: In the input layer, the sequence representation is extracted from the user's real-time behavior data, and the user's static features and sequences are concatenated into a complete model input; In the expert network layer, multiple expert networks are defined. Each expert is responsible for extracting a set of common features. Each expert is a fully connected network or a complex network, and outputs a common high-dimensional feature representation. In the gated network layer, two gated networks are constructed to correspond to the probability prediction of user clicking on advertisements and the probability prediction of user purchasing advertised products. One of them corresponds to the two prediction tasks, and the contribution weight of each expert is dynamically allocated. The gated network is usually implemented by a fully connected layer, with the input as the concatenated features and the output as the weight distribution for each expert. In the task tower layer, each task has an independent tower. The task tower for predicting the probability of users clicking on ads is used to optimize the cross entropy loss of click-through rate prediction, and the task tower for predicting the probability of users purchasing advertised products is used to optimize the cross entropy loss of purchase probability prediction. The output layer outputs two probability values: the probability that a user clicks on an ad and the probability that a user purchases the advertised product.

[0067] In the multi-gate based hybrid expert model, the loss function is the weighted sum of the probability of a user clicking on an ad and the probability of purchasing an advertised product, as expressed as follows:

[0068] in, is the weight hyperparameter, is the loss function of the probability of a user clicking an ad, The loss function for the probability of a user purchasing an advertised product.

[0069] Some other embodiments of the present invention may also use a DIN (Deep Interest Network) model or a DeepFM (Deep Factorization Machine) model.

[0070] The model structure of DIN is divided into the following parts: Input layer, receiving input user static features, advertisement static features, and user behavior sequence features; The attention layer calculates the relevance weight of each historical behavior in the behavior sequence and the current ad through the attention score, dynamically generates user interest expressions, and highlights the behavior that is most relevant to the current ad; A fully connected network is used to concatenate features and further model them through a multi-layer perceptron. The output layer outputs the probability of clicking on the ad and the probability of purchasing.

[0071] Among them, the loss function of the probability of users clicking on advertisements adopts the cross entropy loss function, and the loss function of the probability of users purchasing advertised products can be weighted combined or independently optimized.

[0072] The model structure of DeepFM consists of the following parts: Input layer, which receives the input sparse feature embedding (user features, advertising features, real-time behavior features); The factorization machine part learns low-order feature interactions (such as the interaction between user age and ad category); The Deep Neural Network (DNN) part uses deep neural networks to learn high-order nonlinear interactions; The fusion layer fuses the outputs of the factorization machine part and the deep neural network part; The output layer first outputs the probability of the user clicking on the advertisement. The probability of the user purchasing the advertised product can be output by sharing the underlying network or by independent training.

[0073] In one embodiment of the present invention, the conversion rate is calculated as follows:

[0074] In the formula, To calculate the conversion rate of the corresponding user, It is a preset weighting coefficient that determines the contribution of click probability and purchase probability to the final conversion rate. is the probability that a user clicks on an ad, is the probability that the user purchases the advertised product.

[0075] In the traditional weighted average algorithm, the user's click probability and purchase probability are linearly weighted. Since the relationship between click and purchase behavior is more complex and the impact on conversion rate is nonlinear, it can be fused through a nonlinear function. This method is more sensitive to higher values ​​of click and purchase probabilities, so that users with higher click and purchase rates can get greater weights in the fusion. Through this nonlinear transformation, users with high conversion behavior can be effectively highlighted.

[0076] In other embodiments of the present invention, the conversion rate is calculated as follows:

[0077] In the formula, Indicates that at a given click probability and purchase probability Under the condition of Likelihood distribution, Indicates that at a given conversion rate Under the condition of, the probability of users clicking on ads is, Indicates that at a given conversion rate Under the condition of, the probability of users purchasing the advertised product is, is the prior probability, indicating the user's conversion rate The distribution of , estimated using historical data or expert knowledge, is the marginal probability of click probability, which represents the probability distribution of all users clicking ads. is the marginal probability of purchase probability, which represents the probability distribution of all users purchasing the advertised product.

[0078] This method can effectively combine prior knowledge (e.g., the distribution of historical data) and current click and purchase behavior probabilities to generate a more reasonable conversion rate estimate. It can effectively handle missing data and noise, and can enhance the reliability of the model through prior knowledge. It is suitable for probabilistic inference when the data is incomplete or the uncertainty is high.

[0079] Allocate the initial advertising budget based on the historical data of the delivery channels. The allocation ratio of the initial budget can be determined by analyzing the historical performance of different channels (such as click-through rate, conversion rate, user activity, cost-effectiveness, etc.).

[0080] Taking channel conversion rate as an example, on each channel, the initial advertising budget is allocated based on its historical conversion rate performance. Channels with higher conversion rates will get more budget, and channels with lower conversion rates will get less budget. Suppose there are three advertising channels: Channel A, Channel B, and Channel C, and their historical data are as follows:

[0081] In this example, assuming a total initial advertising budget of $300,000, budget allocation can be adjusted based on each channel’s historical conversion rate to cost ratio: First, calculate the budget allocation ratio for each channel: Budget ratio = conversion rate and cost ratio of each channel / sum of all channel ratios.

[0082] The sum of all channel ratios = 0.05 + 0.0375 + 0.0583 = 0.1458; Budget ratio of channel A = 0.05 / 0.1458≈34.3%; Budget ratio of channel B = 0.0375 / 0.1458 ≈ 25.7%; The budget ratio of channel C = 0.0583 / 0.1458≈39.9%.

[0083] Then make a preliminary budget allocation: Channel A budget = 300,000 yuan × 34.3% ≈ 102,900 yuan; Channel B budget = 300,000 yuan × 25.7% ≈ 77,100 yuan; Channel C budget = 300,000 yuan × 39.9% ≈ 119,700 yuan.

[0084] In this way, the preliminary advertising budget is reasonably allocated to the three channels based on historical data. In addition, the allocation of the preliminary advertising budget can also be based on historical click volume and cost-effectiveness, based on historical ROI (return on investment), based on user activity or potential.

[0085] The calculation method of the channel conversion rate is:

[0086] In the formula, Indicates Channel conversion rate of each channel, Represents the total number of users in all channels, Indicates The number of users of each channel, Indicates Channel When calculating the conversion rate, this embodiment designs an adjustment coefficient to compare the number of users in each channel with the maximum number of users. The channel with a large number of users has a larger adjustment coefficient, so its conversion rate will be appropriately improved, thereby reflecting the impact of more users on the conversion rate of the channel.

[0087] In order to solve the impact of the number of users on the conversion rate, we can design an adjustment coefficient based on the ratio of the number of users in each channel to the total number, and apply this coefficient to the channel conversion rate calculation. According to this coefficient, we can adjust the conversion rate of each channel. Generally speaking, the more users a channel has, the larger the adjustment coefficient is, and the conversion rate is appropriately improved.

[0088] The initial advertising budget of the dynamically adjusted delivery channel is adjusted based on the conversion rate difference, and the adjustment method is: For each channel, calculate the difference between the channel conversion rate of the corresponding channel and the preset channel conversion rate when allocating the initial advertising budget:

[0089] If the channel conversion rate of a channel is higher than the preset conversion rate, the advertising budget of the group is increased; if it is lower than the preset conversion rate, the advertising budget of the channel is reduced:

[0090] In the formula, It is an adjustment coefficient used to control the sensitivity of budget adjustment. The adjustment coefficient can be set by manual intervention, and different adjustment coefficient values ​​can be set according to the sensitivity of budget control.

[0091] In some other embodiments of the present invention, the adjustment method also includes a multi-armed bandit algorithm (Multi-ArmedBandit, MAB), a weighted linear regression (Weighted Linear Regression), a Bayesian optimization algorithm, a linear programming optimization (Linear Programming) and a gradient descent method (Gradient Descent), etc.

[0092] Embodiment 2 This embodiment is an advertising budget allocation system based on machine learning and user behavior. The system is used to implement the method described in the first embodiment, including: Data collection module, used to collect historical behavior data and real-time behavior data of users; The data processing module processes the data submitted by the data collection module. The processing methods include cleaning, deduplication and missing value filling. A probability prediction model receives the processed real-time behavior data and predicts the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product, wherein the probability prediction model is trained using the processed historical behavior data; The fusion model receives the probability of users clicking on ads and the probability of users purchasing advertised products output by probability prediction, calculates the user conversion rate, and calculates the channel conversion rate based on the delivery channel; The budget allocation module is used to allocate preliminary advertising budgets based on the historical data of the delivery channels; and dynamically adjust the preliminary advertising budgets of the delivery channels based on the channel conversion rate.

[0093] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. An advertising budget allocation method based on machine learning and user behavior, characterized in that: The following steps are involved: Collecting historical behavior data of users, and extracting feature data of the historical behavior data through feature engineering; Constructing a user probability prediction model, and using the feature data to train the user probability prediction model; During the advertising process, real-time user behavior data is collected and input into a trained user probability prediction model to predict the probability of a user clicking on an ad and the probability of a user purchasing an advertised product. Constructing a fusion model, wherein the fusion model calculates the user conversion rate based on the probability of the user clicking on the advertisement and the probability of the user purchasing the advertised product, and calculates the channel conversion rate based on the delivery channel; Allocate preliminary advertising budget based on historical data of delivery channels; Dynamically adjust the initial advertising budget of the delivery channel based on the channel conversion rate; The calculation method of the channel conversion rate is as follows: In the formula, Indicates Channel conversion rate of each channel, Represents the total number of users in all channels, Indicates The number of users of each channel, Indicates Channel The conversion rate of each user.

2. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The user probability prediction model is a gated recurrent unit model. The output layer of the gated recurrent unit model includes two independent neurons, which are used to predict the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product. The structure of the output layer is expressed as follows: In the formula, is the probability of a user clicking on an ad predicted by the gated recurrent unit model, is the probability of a user purchasing an advertised product predicted by the gated recurrent unit model, is the weight matrix corresponding to the user clicking the ad, is the weight matrix corresponding to the user purchasing the advertised product, is the bias term for the user click ad prediction task, is the bias item for the task of predicting whether users purchase advertised products. is the final hidden state after the update gate and reset gate output in the gated recurrent unit model.

3. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The user probability prediction model is two lightweight gradient boosting machine models, which are used to predict the probability of a user clicking on an ad and the probability of a user purchasing an advertised product. The two probabilities are expressed as follows: In the formula, is the probability of a user clicking on an ad predicted by the lightweight gradient boosting model, is the probability of a user purchasing an advertised product predicted by the lightweight gradient boosting model, is the logarithmic probability output by the lightweight gradient boosting machine model in the user click ad prediction task, It is the logarithmic probability output by the lightweight gradient boosting machine model in the task of predicting whether a user will purchase an advertised product.

4. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The user probability prediction model is a random forest model, and the output of the random forest model is expressed as: In the formula, is the probability of a user clicking on an ad predicted by the random forest model, is the probability of a user purchasing an advertised product predicted by the random forest model, is the number of trees in the random forest model, The random forest model The probability of users clicking on ads predicted by Keshu, The random forest model The probability of a user purchasing an advertised product predicted by Keshu.

5. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The user probability prediction model is a hybrid expert model based on multiple gates, including an input layer, an expert network layer, a gated network layer, a task tower layer and an output layer. The structure and function of the model are as follows: In the input layer, the sequence representation is extracted from the user's real-time behavior data, and the user's static features and sequences are concatenated into a complete model input; In the expert network layer, multiple expert networks are defined. Each expert is responsible for extracting a set of common features. Each expert is a fully connected network or a complex network, and outputs a common high-dimensional feature representation. In the gated network layer, two gated networks are constructed to correspond to the probability prediction of users clicking on ads and the probability prediction of users purchasing advertised products. One of them corresponds to the two prediction tasks, and the contribution weight of each expert is dynamically allocated. In the task tower layer, each task has an independent tower. The task tower for predicting the probability of users clicking on ads is used to optimize the cross entropy loss of click-through rate prediction, and the task tower for predicting the probability of users purchasing advertised products is used to optimize the cross entropy loss of purchase probability prediction. The output layer outputs two probability values: the probability that a user clicks on an ad and the probability that a user purchases the advertised product.

6. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The conversion rate is calculated as follows: In the formula, To calculate the conversion rate of the corresponding user, It is a preset weighting coefficient that determines the contribution of click probability and purchase probability to the final conversion rate. is the probability that a user clicks on an ad, is the probability that the user purchases the advertised product.

7. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The conversion rate is calculated as follows: In the formula, Indicates that at a given click probability and purchase probability Under the condition of Likelihood distribution, Indicates that at a given conversion rate Under the condition of, the probability of users clicking on ads is, Indicates that at a given conversion rate Under the condition of, the probability of users purchasing the advertised product is, is the prior probability, indicating the user's conversion rate The distribution of , estimated using historical data or expert knowledge, is the marginal probability of click probability, which represents the probability distribution of all users clicking ads. is the marginal probability of purchase probability, which represents the probability distribution of all users purchasing the advertised product.

8. The method for allocating advertising budget based on machine learning and user behavior according to claim 1, characterized in that: The initial advertising budget of the dynamically adjusted delivery channel is adjusted based on the conversion rate difference, and the adjustment method is: For each channel, calculate the difference between the channel conversion rate of the corresponding channel and the preset channel conversion rate when allocating the initial advertising budget: If the channel conversion rate of a channel is higher than the preset conversion rate, the advertising budget of the group is increased; if it is lower than the preset conversion rate, the advertising budget of the channel is reduced: In the formula, It is the adjustment coefficient used to control the sensitivity of budget adjustment.

9. An advertising budget allocation system based on machine learning and user behavior, characterized in that: The system is used to implement the method according to any one of claims 1 to 8, comprising: Data collection module, used to collect historical behavior data and real-time behavior data of users; The data processing module processes the data submitted by the data collection module. The processing methods include cleaning, deduplication and missing value filling. A probability prediction model receives the processed real-time behavior data and predicts the probability of a user clicking on an advertisement and the probability of a user purchasing an advertised product, wherein the probability prediction model is trained using the processed historical behavior data; The fusion model receives the probability of users clicking on ads and the probability of users purchasing advertised products output by probability prediction, calculates the user conversion rate, and calculates the channel conversion rate based on the delivery channel; The budget allocation module is used to allocate preliminary advertising budgets based on the historical data of the delivery channels; and dynamically adjust the preliminary advertising budgets of the delivery channels based on the channel conversion rate.

Images