Advertising and Marketing System Based on Data Update and Sorting

Abstract

The present invention relates to the technical field of advertising data marketing. Specifically, it relates to an advertising marketing system based on data update and collation, which includes a data collection and transmission unit, a data processing and analysis unit, an advertising strategy generation and optimization unit, and an advertising placement monitoring unit. The data collection and transmission unit of the present invention collects multi-source data, cleans and processes it, and provides a comprehensive data set. The data processing and analysis unit uses in-memory computing, parallel processing architecture, and various algorithms to calculate user behavior and market trend indicators, mine graph associations, and predict user behavior. The advertising strategy generation and optimization unit constructs a strategy generation engine to generate an initial strategy, dynamically optimizes it using deep reinforcement learning, and evaluates the return on advertising investment in real time. The advertising placement monitoring unit uses the optimized strategy for actual placement, monitors the conversion rate in real time, detects anomalies through an autoencoder model, and adjusts the strategy, improving the accuracy and efficiency of advertising marketing, increasing the return on advertising investment, and enhancing the advertising effect.

Description

Technical Field

[0001] The present invention relates to the technical field of advertising data marketing, and more specifically, to an advertising marketing system based on data update and collation. Background Art

[0002] Advertising data marketing is an important technology. With the rapid development of the Internet, great changes have taken place in the field of advertising marketing, and data-driven advertising marketing strategies have gradually become the mainstream.

[0003] Traditional systems do not comprehensively collect user behavior data, market competition data, and industry trend data, and the data cleaning and processing methods are relatively simple, making it difficult to cope with complex and changeable data environments, resulting in low data quality, unable to accurately reflect the real market situation, and lacking an efficient processing architecture and advanced algorithms, unable to dynamically adjust the analysis window according to real-time data traffic, difficult to quickly and accurately calculate user behavior indicators and market trend indicators, and also difficult to deeply explore potential associations in user behavior maps and market relationship maps, unable to accurately predict future user behavior trends, resulting in the initial advertising placement strategy lacking scientificity and accuracy, mostly based on experience or simple data analysis, difficult to adapt to complex market environments. Moreover, strategy adjustments are often lagging and cannot be optimized in real time according to advertising placement effects, resulting in an unsatisfactory input-output ratio of advertising placement. To solve this technical problem, we provide an advertising marketing system based on data update and collation. Summary of the Invention

[0004] The purpose of the present invention is to provide an advertising marketing system based on data update and collation to solve the problems raised in the above background art.

[0005] To achieve the above object, an advertising marketing system based on data update and collation is provided, including a data collection and transmission unit, a data processing and analysis unit, an advertising strategy generation and optimization unit, and an advertising placement monitoring unit;

[0006] The data collection and transmission unit obtains user behavior data, market competition data, and industry trend data through a data collector, and performs data cleaning and processing on the collected data to obtain a comprehensive data set;

[0007] The data processing and analysis unit processes the data in the comprehensive dataset using a data processing module based on in-memory computing and parallel processing architecture, adjusts the window size according to the real-time traffic of the data using an adaptive window algorithm to obtain the real-time change characteristics of the data. At the same time, it combines asynchronous I / O technology and vectorized computing methods to perform real-time aggregation analysis on the data, calculates user behavior metrics and market trend metrics, stores the user behavior metrics and market trend metrics in the graph database, then uses the data analysis module to extract the user behavior graph and market relationship graph from the graph database, and performs association mining on the user behavior graph and market relationship graph through a graph neural network algorithm based on the attention mechanism to obtain the user behavior sequence, and then uses a long short-term memory network to model the user behavior sequence to predict future behavior trends;

[0008] The advertising strategy generation and optimization unit constructs a strategy generation engine through the strategy generation module, combines the real-time data analysis results and preset business rules to generate an initial advertising placement strategy, and introduces a dynamic optimization algorithm based on multi-agent deep reinforcement learning through the strategy optimization module, using a method that combines experience replay and prioritized experience replay to adjust the strategy parameters in real time according to the advertising placement effect. At the same time, it establishes a real-time advertising effect evaluation index system and uses the real-time calculation function of the effect evaluation module to evaluate the advertising return on investment in real time;

[0009] The advertising placement monitoring unit applies the optimized advertising strategy to actual placement, and monitors the conversion rate of the advertisement in real time, uses an anomaly detection algorithm based on streaming time series analysis to monitor anomalies in the advertising placement process, and analyzes the monitoring data in real time, and adjusts the advertising placement strategy according to the analysis results.

[0010] As a further improvement of this technical solution, when the data collection and transmission unit performs data cleaning, it uses a density-based local outlier factor algorithm for outlier detection in user behavior data. The specific steps are as follows:

[0011] For each data point in the user behavior dataset , calculate its distance to nearest neighbor points, where is a pre-set parameter, denoted as , where , represents the nearest neighbor point, represents the th nearest neighbor point, and take the distance of the th nearest neighbor point as the th distance distance ;

[0012] Calculate of The average distance from all points within the distance neighborhood to is denoted as the reachability distance of Based on the reachability distance of the local reachability density of is calculated, and then based on the local reachability density of the local outlier factor of is calculated. If the local outlier factor of is greater than 1, then

[0013] As a further improvement of this technical solution, in the data processing and analysis unit, when the data processing module based on the in-memory computing and parallel processing architecture adopts the adaptive window algorithm, the window size adjustment formula is as follows:

[0014] Let the current window size be , the real-time data traffic be , and the traffic change rate be , where is the traffic at the previous moment. Define an adaptive coefficient ;

[0015] When , the window size is adjusted to ; when , the window size is adjusted to .

[0016] As a further improvement of this technical solution, in the data processing module, when performing real-time aggregation analysis by combining asynchronous I / O technology and vectorized calculation methods, the calculation of user behavior metrics is as follows:

[0017] Extract the user behavior dataset from the comprehensive dataset of the data collection and transmission unit. The user behavior dataset includes user ID, behavior time, and behavior type. Asynchronously read the data in parallel using asynchronous I / O technology and convert the read data into vector form;

[0018] The behavior types include browsing, clicking, and purchasing. For the behavior types, one-hot encoding is used for vectorization, and the user behavior frequency metric is calculated. For each user, the cumulative sum of different behavior type vectors within a certain time window is statistically calculated, and then divided by the time window length to obtain the behavior frequency vector. The user behavior metric is calculated based on the behavior frequency vector.

[0019] As a further improvement of this technical solution, in the data analysis module, when the graph neural network algorithm based on the attention mechanism performs correlation mining on the user behavior graph and the market relationship graph, the specific steps are as follows:

[0020] Record settings are respectively made for the node feature vectors of the user behavior graph and the market relationship graph. For each node in the user behavior graph, a query vector is obtained through linear transformation. For each node in the market relationship graph, a key vector and a value vector are obtained;

[0021] Then, attention scores are calculated based on the query vector and the key vector. The attention scores are combined with a preset softmax function to calculate attention weights. Then, based on the attention weights and the value vector, associated feature vectors are calculated. After concatenating them with the original node feature vectors and performing a non-linear transformation, updated node feature vectors are obtained, and the association relationships between the two graphs are mined based on the node feature vectors.

[0022] As a further improvement of this technical solution, in the data analysis module, when the long short-term memory network models the user behavior sequence, a gated recurrent unit structure is used for optimization. The specific operations are as follows:

[0023] For the input user behavior sequence and the hidden state at the previous moment , calculate the reset gate , the update gate , where is the sigmoid function, is the weight matrix. The reset gate is used to control the reset of information in the hidden state at the previous moment, and the update gate is used to control the transmission and update of information;

[0024] Calculate the candidate hidden state based on the user behavior sequence, the hidden state at the previous moment, and the reset gate, and calculate the final hidden state based on the candidate hidden state. Finally, based on the final hidden state, the long-term dependencies in the user behavior sequence are obtained and modeled.

[0025] As a further improvement of this technical solution, in the advertisement strategy generation and optimization unit, when the strategy generation module constructs the strategy generation engine, a genetic algorithm is used to optimize the generation of the initial advertisement placement strategy. The specific steps are as follows:

[0026] Encode the advertisement placement time, placement channels, and placement budget into an integer sequence as a chromosome, and randomly generate a certain number of chromosomes as the initial population. Each chromosome represents an initial advertisement placement strategy, and then calculate the fitness based on the expected advertisement conversion rate;

[0027] Select the parent chromosomes according to the fitness values, perform crossover and mutation operations on the parent chromosomes to generate offspring chromosomes, and repeat the above steps until the maximum number of iterations is reached to obtain the optimized initial advertisement placement strategy.

[0028] As a further improvement of this technical solution, the specific operation of the policy optimization module for real-time adjustment of policy parameters according to the advertising delivery effect by using a method combining experience replay and prioritized experience replay is as follows:

[0029] During the advertising delivery process, multiple agents interact with the advertising delivery environment. Each agent experiences a state at each time step, selects an action based on the current policy for advertising delivery operations, and then obtains a reward. The reward is set by the advertising conversion rate. At the same time, it enters the next state. This set of information is stored as a sample in the experience replay pool. The experience replay pool is a first-in-first-out queue. When the queue is full, the earliest entered sample will be removed;

[0030] Calculate a priority for each sample in the experience replay pool. The calculation of the priority is jointly determined by the reward size, timestamp, and prediction error. Each time a batch of samples is sampled from the experience replay pool for training the deep reinforcement learning model, sampling is performed according to the priority of the samples to obtain a batch of samples. These samples are input into the deep reinforcement learning model for training. According to the algorithm of deep reinforcement learning, calculate the loss function and update the parameters of the model through the backpropagation algorithm.

[0031] As a further improvement of this technical solution, in the real-time advertising effect evaluation index system established by the effect evaluation module, when calculating the advertising return on investment index, the delayed effect of advertising delivery is considered. The specific formula is as follows:

[0032] Let the advertising delivery time be , and within the time after advertising delivery, the revenue is statistically counted, and the advertising delivery cost is . Considering that the revenue is generated within a certain period of time after advertising delivery, the weighted average method is used to calculate the delayed revenue, that is:

[0033] ;

[0034] where is the revenue at moment, is the weight determined according to the distance between the time and the advertising delivery time . Then the advertising return on investment .

[0035] As a further improvement of this technical solution, in the advertising delivery monitoring unit, the anomaly detection algorithm based on streaming time series analysis uses an autoencoder model. The specific steps are as follows:

[0036] Normalize the monitoring data in the advertising placement process and use it as the input of the autoencoder. The monitoring data is the advertising conversion rate, and the autoencoder consists of an encoder and a decoder.

[0037] The encoder maps the input to a low-dimensional representation, and the decoder reconstructs the low-dimensional representation into a new input of the autoencoder. The reconstruction error is defined by calculating the difference between the new input and the input of the autoencoder. Train the autoencoder to minimize the reconstruction error.

[0038] During actual detection, for new monitoring data, calculate its reconstruction error. If the reconstruction error is greater than a pre-set threshold, determine that the data point is an outlier and conduct real-time monitoring. Then, according to the anomaly detection result, use the anomaly situation as the state input to the reinforcement learning model. The model outputs adjustment actions. By continuously interacting with the environment, the reinforcement learning model learns the optimal adjustment strategies under different abnormal states.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] In the advertising marketing system based on data update and collation, the data processing and analysis unit uses a memory-based computing and parallel processing architecture, combined with an adaptive window algorithm, asynchronous I / O technology, and vectorized computing methods, to accurately calculate various indicators, deeply explore the potential value of data, accurately predict the user behavior trend, and provide data support for the formulation of advertising strategies. The strategy generation module uses a genetic algorithm to optimize the initial strategy and improve the scientificity of the strategy. The strategy optimization module uses multi-agent deep reinforcement learning and experience replay technology to adjust parameters in real time. At the same time, the effect evaluation module calculates the input-output ratio considering the advertising placement delay effect, making the advertising strategy more in line with market demand and improving the input-output ratio of advertising placement. The advertising placement monitoring unit monitors the advertising conversion rate in real time, uses an autoencoder model to detect anomalies, and combines a reinforcement learning model to adjust strategies in a timely manner to ensure that the advertising placement is always in the best state and improve the accuracy and efficiency of advertising marketing. Brief Description of the Drawings

[0041] Figure 1 It is the overall block diagram of the present invention.

[0042] The meanings of the various labels in the figure are as follows:

[0043] 1. Data acquisition and transmission unit; 2. Data processing and analysis unit; 21. Data processing module; 22. Data analysis module; 3. Advertising strategy generation and optimization unit; 31. Strategy generation module; 32. Strategy optimization module; 33. Effect evaluation module; 4. Advertising placement monitoring unit. Detailed Embodiments

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0045] The present invention provides an advertising marketing system based on data update and collation. Please refer to Figure 1 as shown, which includes a data collection and transmission unit 1, a data processing and analysis unit 2, an advertising strategy generation and optimization unit 3, and an advertising placement monitoring unit 4;

[0046] The data collection and transmission unit 1 obtains user behavior data, market competition data, and industry trend data through a data collector, and performs data cleaning and processing on the collected data to obtain a comprehensive data set.

[0047] When the data collection and transmission unit 1 performs data cleaning, the density-based local outlier factor algorithm is used for outlier detection in user behavior data. The specific steps are as follows:

[0048] During the collection process of user behavior data, it will be interfered by various factors, resulting in outliers. These outliers may seriously affect the subsequent data processing and analysis results, leading to inaccurate advertising strategy generation. For each data point in the user behavior data set , calculate its distance to the nearest neighbor points, where is a pre-set parameter, denoted as , and take the distance of the th nearest neighbor point as the distance.

[0049] User behavior data usually has complex distribution characteristics. The behavior patterns of different users may vary greatly, and there may be multiple regions with different densities in the data set. The density-based local outlier factor algorithm does not depend on specific distribution assumptions of the data and can accurately identify outliers in complex data distributions. By considering the local density information of data points, it can distinguish those points with large density differences from surrounding data points in the local area and determine them as outliers.

[0050] Calculate the average distance from all points in the distance neighborhood of to , denoted as the reachable distance of . The specific formula is as follows: ; Calculate based on the reachable distance of The local reachability density, that is ; where is of the distance neighborhood, calculate the local outlier factor of ; If the local outlier factor of is greater than 1, then is determined as an outlier.

[0051] The LOF algorithm is a local density-based method. It considers the local environment around each data point. The LOF algorithm can detect data points that behave abnormally in the local area, even if these points may not be obvious in the global distribution. It can accurately identify those data points that behave abnormally in these small groups without misjudging the entire small group as abnormal, making the subsequent data processing and analysis results more accurately reflect the true behavior patterns of users.

[0052] The data processing and analysis unit 2 uses the data processing module 21 based on the in-memory computing and parallel processing architecture to process the data in the comprehensive dataset, adjusts the window size according to the real-time traffic of the data using the adaptive window algorithm, obtains the real-time change characteristics of the data. At the same time, combines the asynchronous I / O technology and the vectorized computing method to perform real-time aggregation analysis on the data, calculates the user behavior indicators and market trend indicators, and stores the user behavior indicators and market trend indicators in the graph database. Then, uses the data analysis module 22 to extract the user behavior graph and market relationship graph from the graph database, and performs association mining on the user behavior graph and market relationship graph through the graph neural network algorithm based on the attention mechanism to obtain the user behavior sequence, and then uses the long short-term memory network to model the user behavior sequence to predict the future behavior trend.

[0053] In the data processing and analysis unit 2, when the data processing module 21 based on the in-memory computing and parallel processing architecture uses the adaptive window algorithm, the window size adjustment formula is as follows:

[0054] Before starting to apply the adaptive window algorithm, some necessary initial parameters need to be determined. These parameters are the basis for subsequent calculation and adjustment of the window size. Set the initial window size , at the same time, start real-time monitoring of the data traffic, record the data traffic at the current moment , in addition, define the adaptive coefficient , The value range of is .

[0055] At each time step , obtain the current real-time data traffic as ​, and based on the traffic at the previous moment Calculate the traffic change rate ; The traffic change rate is a key indicator to measure the real-time traffic change of data. By calculating the traffic change rate, it is possible to understand whether the data traffic is increasing, decreasing or remaining stable, thus providing a basis for adjusting the window size.

[0056] Adjust the window size according to the positive or negative situation of the calculated traffic change rate. When it indicates that the data traffic is increasing. At this time, the window size is adjusted to ; When it indicates that the data traffic is decreasing or remaining unchanged. The window size is adjusted to ; The adaptive window size adjustment method can dynamically optimize the window size according to the actual change of data traffic, avoiding both the waste of computing resources caused by too large a window and the problem that too small a window cannot capture important data features.

[0057] In the data processing module 21, when performing real-time aggregation analysis by combining asynchronous I / O technology and vectorized calculation methods, the calculation of user behavior indicators is as follows:

[0058] Screen out the records containing user ID, behavior time and behavior type from the comprehensive data set of the data acquisition and transmission unit 1 to form a user behavior data set. The user behavior data set may be very large. If the synchronous I / O technology is used to sequentially read the data, a large amount of time will be spent waiting for the I / O operation to complete, resulting in low data processing efficiency. Use a programming language library that supports asynchronous I / O to parallelly read the user behavior data set. After reading the data, convert each record into a vector form.

[0059] Asynchronous I / O technology and parallel reading greatly shorten the data reading time and improve the overall efficiency of data processing. The vectorized data representation form is convenient for batch calculation and can accelerate the subsequent data analysis process. The behavior type is discrete categorical data, and the computer cannot directly process this text information. Using one-hot encoding to convert the behavior type into a vector form can convert the categorical data into numerical data.

[0060] Define the set of behavior types , for each behavior type, create a vector with a length equal to the number of behavior types, where only the position corresponding to the behavior type is 1 and the rest are 0. Converting the behavior type into a computable vector form provides a basis for subsequent calculation of user behavior frequency indicators and behavior indicators.

[0061] The frequency of user behavior can reflect the user's activity level and behavior preferences within a certain period of time. By calculating the frequency indicators of different behavior types, we can understand the user's inclination towards different behaviors, providing an important reference basis for formulating advertising strategies. For each user, within a certain time window we calculate the cumulative sum of the vectors of different behavior types, and then divide the cumulative sum by the length of the time window to obtain the behavior frequency vector. The behavior frequency indicator can quantify the user's behavior characteristics. By comparing the behavior frequency vectors of different users, we can discover the behavior differences between users, providing a basis for personalized advertising recommendations.

[0062] The behavior frequency vector is only a preliminary quantification of user behavior. By further calculating the user behavior indicators, we can comprehensively consider the importance of different behavior types and more comprehensively evaluate the user's behavior characteristics, assigning different weights to different behavior types , let the set of behavior types correspond to the weight vector , then the behavior indicator of user is ; where is the i-th element of the behavior frequency vector .

[0063] In the data analysis module 22, when the graph neural network algorithm based on the attention mechanism conducts correlation mining on the user behavior graph and the market relationship graph, the specific steps are as follows:

[0064] Before conducting correlation mining, it is necessary to clarify the feature representations of each node in the user behavior graph and the market relationship graph. The node feature vector is the basis for subsequent calculations. Let the user behavior graph be , where is the node set of the user behavior graph. For each node , its feature vector is denoted as , and the market relationship graph is . For each node , its feature vector is denoted as . These feature vectors ensure the accuracy and consistency of subsequent correlation mining calculations, avoiding calculation errors or result deviations caused by unclear feature representations.

[0065] By linearly transforming the node feature vectors to obtain query vectors, key vectors, and value vectors, we can map the node features to different subspaces, enabling us to measure the correlation between nodes from different perspectives when calculating the attention scores. For each node in the user behavior graph, we obtain the query vector through linear transformation, where ​is a learnable weight matrix for each node in the market relationship graph , and the key vector is obtained through linear transformation and the value vector , where and are learnable weight matrices. Introducing learnable weight matrices for linear transformation increases the flexibility and expressive power of the model, and can automatically learn the importance of different node features in association mining.

[0066] The attention score is used to measure the correlation between the nodes in the user behavior graph and the nodes in the market relationship graph. For the node in the user behavior graph and the node in the market relationship graph, calculate the attention score, that is ; where is the dimension of the query vector and the key vector, and dividing by is to prevent the dot product result from being too large and avoid the problem of gradient disappearance when using the softmax function in the subsequent process.

[0067] Input the attention score into the preset softmax function to calculate the attention weight . The softmax function can normalize the attention score, making the weights comparable and highlighting the importance of nodes with higher correlation, which is convenient for more reasonable weight allocation in the subsequent calculation.

[0068] The associated feature vector synthesizes the information of each node in the market relationship graph. By weighted summing the value vectors with the attention weights, it can integrate the information of the market relationship graph nodes with higher correlation with the user behavior graph nodes into a vector, thereby mining the association information between the two graphs.

[0069] For the node in the user behavior graph, calculate the associated feature vector. The associated feature vector can effectively fuse the information related to the user behavior graph nodes in the market relationship graph, concentrate the association information scattered in multiple nodes into a vector, which is convenient for subsequent processing and analysis. After splicing the associated feature vector with the original node feature vector and passing through a non-linear transformation, the mined association information can be incorporated into the features of the user behavior graph nodes. The updated node feature vector contains more information about the association between the two graphs.

[0070] In the data analysis module 22, when the long short-term memory network models the user behavior sequence, it adopts the gated recurrent unit structure for optimization. The specific operations are as follows:

[0071] When processing user behavior sequences, traditional neural networks have difficulty capturing long-term dependencies in the sequences. Gated Recurrent Units (GRUs) solve this problem by introducing reset gates and update gates. The reset gate can control how much information from the previous hidden state needs to be forgotten or reset, while the update gate determines how much new information should be added to the current hidden state.

[0072] Let the vector of the input user behavior sequence at time be , and the previous hidden state . Then the calculation formula for the reset gate is , and the update gate , where is the sigmoid function, is the weight matrix. The introduction of the reset gate and update gate enables the model to dynamically control the flow and update of information according to the input, avoiding the problems of gradient vanishing or gradient explosion in traditional recurrent neural networks and enhancing the model's ability to capture long-term dependencies.

[0073] The calculation formula for the candidate hidden state is , where is the hyperbolic tangent function, which maps the input to the interval (-1, 1); is the learnable weight matrix, denotes element-wise multiplication. The use of the reset gate enables the model to flexibly determine which parts of the previous hidden state need to be retained or forgotten according to the input, thereby generating a candidate hidden state that better conforms to the current input and improving the flexibility and adaptability of the model.

[0074] The final hidden state is calculated based on the update gate and the candidate hidden state. The update gate determines the proportion of the previous hidden state and the candidate hidden state in the final hidden state. The calculation formula for the final hidden state is ; where represents the retention proportion of the previous hidden state, represents the incorporation proportion of the candidate hidden state. The introduction of the update gate enables the model to dynamically balance the weights of historical information and current input information, avoiding the problem of over-reliance on historical information or current input information and improving the stability and generalization ability of the model.

[0075] The final hidden state contains the comprehensive information from the past to the current moment in the user behavior sequence. By analyzing and utilizing the final hidden state, the long-term dependencies in the user behavior sequence can be mined. Taking the final hidden state at each moment as features and inputting them into the subsequent analysis module, with the final hidden state as the input and the actual user behavior or other relevant metrics as the output, a model of the user behavior sequence is established to achieve more accurate prediction and analysis of user behavior.

[0076] The advertisement strategy generation and optimization unit 3 constructs a strategy generation engine through the strategy generation module 31, combines the real-time data analysis results and the preset business rules to generate an initial advertisement placement strategy, and introduces a dynamic optimization algorithm based on multi-agent deep reinforcement learning through the strategy optimization module 32. Using the method of combining experience replay and prioritized experience replay, the strategy parameters are adjusted in real time according to the advertisement placement effect. Meanwhile, a real-time advertisement effect evaluation index system is established, and the real-time calculation function of the effect evaluation module 33 is used to evaluate the advertising return on investment in real time.

[0077] In the advertisement strategy generation and optimization unit 3, when the strategy generation module 31 constructs the strategy generation engine, a genetic algorithm is used to optimize the generation of the initial advertisement placement strategy. The specific steps are as follows:

[0078] Assume that the advertisement placement time can be divided into time periods, the placement channels are In, the placement budget is different value levels, then the advertisement placement time can be represented by an integer from 1 to , the placement channel can be represented by an integer from 1 to , and the placement budget can be represented by an integer from 1 to . Arrange these three integers in order to form an integer sequence with a length of 3 as a chromosome.

[0079] Randomly generate such chromosomes to form an initial population , where represents the th chromosome, and each chromosome represents an initial advertisement placement strategy. The diverse initial population increases the search space of the algorithm and helps to avoid falling into local optimal solutions.

[0080] For each chromosome in the initial population, according to the corresponding advertisement placement time, placement channel and placement budget, calculate the expected advertisement conversion rate under this strategy by combining historical data. The fitness function , that is, the chromosome The fitness value is equal to the corresponding expected conversion rate of the advertisement. Taking the expected conversion rate of the advertisement as the fitness is directly related to the core goal of advertising marketing. It can guide the algorithm to evolve in the direction that is most conducive to improving the conversion rate, making the final obtained advertising placement strategy more practically valuable.

[0081] Calculate the selection probability of each chromosome ; where is the size of the population, and then generate a random number within the interval , and accumulate the selection probabilities of each chromosome in turn . When , select the -th chromosome as the parental chromosome. Repeat this process to select a sufficient number of parental chromosomes for subsequent crossover and mutation operations. Through the selection operation, the chromosomes with high fitness are retained and participate in reproduction, gradually improving the overall fitness of the population, which is conducive to the algorithm finding a better advertising placement strategy.

[0082] Randomly select two chromosomes from the selected parental chromosomes and , randomly select a crossover point , and combine the first genes of with the last 3 - genes of to generate a sub-chromosome . Combine the first genes of with the last 3 - genes of to generate another sub-chromosome .

[0083] For the generated sub-chromosomes, mutate a certain gene in the chromosome with a mutation probability. Then randomly select a gene in the sub-chromosome and replace it with another random integer within the value range of the gene. The combination of crossover and mutation operations can not only make full use of the excellent genes of the parental chromosomes but also introduce new gene combinations, increasing the diversity of the population and the search space, helping the algorithm jump out of the local optimal solution and find the global optimal solution.

[0084] Replace the generated offspring chromosomes with those having the lowest fitness to form a new population. Repeat the above-mentioned steps of selection, crossover, and mutation, calculate the fitness of each chromosome in the new population, select the parent chromosomes, perform crossover and mutation operations to generate new offspring chromosomes, and repeat this process until the preset maximum number of iterations is reached. After multiple iterations of optimization, the chromosome with the highest fitness in the final population is the optimized initial advertising placement strategy.

[0085] The strategy optimization module 32 uses a method that combines experience replay and prioritized experience replay to adjust the strategy parameters in real time according to the advertising placement effect. The specific operations are as follows:

[0086] During the advertising placement process, assume there are agents. At each time step , the -th agent will be in a state . Based on the current strategy , select an action to perform the advertising placement operation. According to the actual effect of the advertising placement, set the reward with the advertising conversion rate . Subsequently, the agent enters the next state . Store this set of information as a sample in the experience replay pool . The experience replay pool is a first-in-first-out queue. When the queue reaches the maximum capacity , the earliest entered sample will be removed to better understand the advertising placement environment and thus more effectively optimize the advertising placement strategy.

[0087] Calculate a priority for each sample in the experience replay pool. The calculation of the priority is jointly determined by the reward magnitude, the timestamp, and the prediction error. Let the weight of the reward be , the weight of the timestamp be , and the weight of the prediction error be . The reward magnitude is directly taken as . The timestamp factor can be represented by a decay function, that is, , where is the current time, is the time when the sample enters the experience replay pool, is the decay coefficient. The prediction error can be represented by the absolute value of the difference between the predicted reward of this sample by the deep reinforcement learning model and the actual reward . Then the priority calculation formula is ; By comprehensively considering the reward magnitude, timestamp, and prediction error, it can more accurately evaluate the importance of each sample for model training, enabling the model to learn more targeted.

[0088] At each time of sampling a batch of samples of size from the experience replay pool , first calculate the total priority of all samples in the experience replay pool , then for each position to be sampled, generate a random number within the interval , traverse the samples in the experience replay pool, calculate the cumulative priority , when , select the th sample as the sampling result, repeat this process times to obtain a batch of samples: ; enabling the model to learn key policy information faster. By inputting the sampled samples into the deep reinforcement learning model for training, calculating the loss function, and updating the model's parameters, the model can continuously learn and optimize the advertising placement strategy.

[0089] In the real-time advertising effect evaluation index system established by the effect evaluation module 33, when calculating the advertising return on investment index, the delay effect of advertising placement is considered. The specific formula is as follows:

[0090] Record the specific time point when the advertisement starts to be placed , and determine a time interval for statistically analyzing the revenue situation after the advertisement is placed. At the same time, clarify the total cost generated during the advertisement placement process, which includes all expenditures related to advertisement placement such as advertisement production costs and placement platform costs.

[0091] Within the time interval after the advertisement is placed, statistically analyze the revenue at each time point , determine the weight according to the distance between the time and the advertisement placement time, and then calculate the delayed revenue in a weighted average manner ; making the calculation of the advertising return on investment more in line with the actual situation and providing a more accurate reference basis for the adjustment of advertising strategies.

[0092] According to the calculated delayed revenue and the advertising placement cost , use the formula Calculate the advertising return on investment (ROI). The higher this ratio is, the better the advertising effectiveness. Conversely, the advertising strategy needs to be adjusted.

[0093] The advertising placement monitoring unit 4 applies the optimized advertising strategy to actual placements and monitors the conversion rate of the advertisement in real time. It uses an anomaly detection algorithm based on streaming time series analysis to monitor anomalies during the advertising placement process, analyzes the monitoring data in real time, and adjusts the advertising placement strategy according to the analysis results.

[0094] In the advertising placement monitoring unit 4, the anomaly detection algorithm based on streaming time series analysis uses an autoencoder model. The specific steps are as follows:

[0095] Assume that the advertising conversion rate monitoring data is a time series. Using the min-max normalization method, the data is normalized to the interval. The normalized data is used as the input of the autoencoder, which improves the training efficiency and accuracy of the autoencoder, enabling the model to better learn the characteristics of the advertising conversion rate data. The autoencoder consists of an encoder and a decoder.

[0096] The encoder maps the input to a low-dimensional representation. The decoder reconstructs the low-dimensional representation into a new input of the autoencoder, and defines the reconstruction error by calculating the difference between the new input and the input of the autoencoder. The autoencoder is trained to minimize the reconstruction error. By minimizing the reconstruction error, the autoencoder can learn the main features and distribution of the data. It can achieve good reconstruction for normal data, while generating a large reconstruction error for abnormal data, thus providing an effective basis for anomaly detection.

[0097] For new monitoring data, input it into the trained autoencoder to obtain the reconstructed input, and calculate the reconstruction error. Preset a threshold. If the reconstruction error is greater than the threshold, then determine that this data point is an anomaly point. Continuously monitor the advertising conversion rate data in real time. Once an anomaly point is found, immediately mark and record the relevant information.

[0098] Take the anomaly detection result as the state and input it into the reinforcement learning model. The reinforcement learning model selects an adjustment action according to the current state. After executing the adjustment action, the environment will feedback a reward. The reinforcement learning model continuously interacts with the environment, updates the policy according to the reward signal, learns the optimal adjustment strategies under different anomaly states, and continuously optimizes the policy to adapt to different anomaly scenarios, improving the flexibility and effectiveness of advertising placement.

[0099] In the present invention, the multi-source data is collected and cleaned by the data acquisition and transmission unit 1 to provide a comprehensive data set. The data processing and analysis unit 2 uses in-memory computing, parallel processing architecture and various algorithms to calculate user behavior and market trend indicators, mine graph associations, and predict user behavior. The advertising strategy generation and optimization unit 3 constructs a strategy generation engine to generate an initial strategy, dynamically optimizes it using deep reinforcement learning, and evaluates the return on advertising investment in real time. The advertising placement monitoring unit 4 uses the optimized strategy for actual placement, monitors the conversion rate in real time, detects anomalies through an autoencoder model and adjusts the strategy, thereby improving the accuracy and efficiency of advertising marketing, increasing the return on advertising investment, and enhancing the advertising effect.

[0100] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An advertising and marketing system based on data update and collation, characterized in that It includes a data acquisition and transmission unit (1), a data processing and analysis unit (2), an advertising strategy generation and optimization unit (3), and an advertising placement monitoring unit (4); The data acquisition and transmission unit (1) obtains user behavior data, market competition data, and industry trend data through a data collector, and performs data cleaning and processing on the collected data to obtain a comprehensive data set; The data processing and analysis unit (2) uses a data processing module (21) based on in-memory computing and parallel processing architecture to process the data in the comprehensive data set, adjusts the window size according to the real-time traffic of the data using an adaptive window algorithm to obtain the real-time change characteristics of the data. At the same time, combined with asynchronous I / O technology and vectorized computing methods, it performs real-time aggregation analysis on the data, calculates user behavior indicators and market trend indicators, stores the user behavior indicators and market trend indicators in a graph database, and then uses a data analysis module (22) to extract a user behavior graph and a market relationship graph from the graph database, and performs association mining on the user behavior graph and the market relationship graph through a graph neural network algorithm based on the attention mechanism to obtain a user behavior sequence, and then uses a long short-term memory network to model the user behavior sequence to predict future behavior trends; The advertising strategy generation and optimization unit (3) constructs a strategy generation engine through a strategy generation module (31), combines the real-time data analysis results and preset business rules to generate an initial advertising placement strategy, and introduces a dynamic optimization algorithm based on multi-agent deep reinforcement learning through a strategy optimization module (32), uses a method combining experience replay and prioritized experience replay to adjust the strategy parameters in real time according to the advertising placement effect. At the same time, a real-time advertising effect evaluation index system is established, and the real-time calculation function of an effect evaluation module (33) is used to evaluate the advertising return on investment in real time; The advertising placement monitoring unit (4) applies the optimized advertising strategy to actual placement, and monitors the conversion rate of the advertisement in real time, uses an anomaly detection algorithm based on streaming time series analysis to monitor anomalies in the advertising placement process, and performs real-time analysis on the monitoring data, and adjusts the advertising placement strategy according to the analysis results.

2. The advertising and marketing system based on data update and collation according to claim 1, wherein: When the data acquisition and transmission unit (1) performs data cleaning, the density-based local outlier factor algorithm is used for anomaly detection in user behavior data. The specific steps are as follows: For each data point p in the user behavior dataset, calculate its distance to the k nearest neighbor points, where k is a pre-set parameter, denoted as d(p, o i ), where i = (1, 2, …, k), o represents the nearest neighbor point, o i represents the i-th nearest neighbor point, and take the distance of the k-th nearest neighbor point among them as the k-distance d k (p); Calculate the average distance from all points within the k-distance neighborhood of p to p, which is denoted as the reachable distance of p. Calculate the local reachable density of p according to the reachable distance of p, and then calculate the local outlier factor of p according to the local reachable density of p. If the local outlier factor of p is greater than 1, then p is determined as an outlier.

3. The advertising and marketing system based on data update and collation according to claim 1, wherein: In the data processing module (21), when performing real-time aggregation analysis by combining asynchronous I / O technology and vectorized computing methods, the calculation of user behavior indicators is as follows: Extract a user behavior data set from the comprehensive data set of the data acquisition and transmission unit (1). The user behavior data set includes user ID, behavior time, and behavior type. Use asynchronous I / O technology to read the data in parallel and convert the read data into a vector form; The described behavior types include browsing, clicking, and purchasing. For the behavior types, one-hot encoding is used for vectorization, and the user behavior frequency index is calculated. For each user, the cumulative sum of different behavior type vectors within a certain time window is statistically calculated, and then divided by the length of the time window to obtain the behavior frequency vector. The user behavior index is calculated based on the behavior frequency vector.

4. The advertising and marketing system based on data update and collation according to claim 3, wherein: In the described data analysis module (22), when the graph neural network algorithm based on the attention mechanism performs associated mining on the user behavior graph and the market relationship graph, the specific steps are as follows: Record settings are respectively performed on the node feature vectors of the user behavior graph and the market relationship graph. For each node in the user behavior graph, a query vector is obtained through linear transformation. For each node in the market relationship graph, a key vector and a value vector are obtained; And the attention score is calculated based on the query vector and the key vector. The attention score is combined with the preset softmax function to calculate the attention weight. Then, the associated feature vector is calculated based on the attention weight and the value vector. After splicing it with the original node feature vector and passing through a non-linear transformation, the updated node feature vector is obtained, and the associated relationship between the two graphs is mined based on the node feature vector.

5. The advertising and marketing system based on data update and collation according to claim 4, characterized in that: In the described data analysis module (22), when the long short-term memory network models the user behavior sequence, the gated recurrent unit structure is used for optimization, and the specific operation is as follows: For the input user behavior sequence X t and the hidden state h at the previous moment t-1 , calculate the reset gate r t = σ(W r X t + U r h t-1 ), update the gate z t = σ(W z X t + U z h t-1 ), where σ is the sigmoid function, and W r , U r , W z , U z are weight matrices. The reset gate is used to control the reset of information in the hidden state at the previous moment, and the update gate is used to control the transmission and update of information; The candidate hidden state is calculated according to the user behavior sequence, the hidden state at the previous moment, and the reset gate, and the final hidden state is calculated based on the candidate hidden state. Finally, the long-term dependence relationship in the user behavior sequence is obtained based on the final hidden state and modeled.

6. The advertising and marketing system based on data update and collation according to claim 5, wherein: In the described advertisement strategy generation and optimization unit (3), when the strategy generation module (31) constructs the strategy generation engine, the genetic algorithm is used to optimize the generation of the initial advertisement placement strategy, and the specific steps are as follows: The advertisement placement time, placement channel, and placement budget are encoded as an integer sequence as the chromosome, and a certain number of chromosomes are randomly generated as the initial population. Each chromosome represents an initial advertisement placement strategy, and then the fitness is calculated based on the expected advertisement conversion rate; The parent chromosomes are selected according to the fitness value, and crossover and mutation operations are performed on the parent chromosomes to generate offspring chromosomes. The above steps are repeated until the maximum number of iterations is reached, and the optimized initial advertisement placement strategy is obtained.

7. The advertising and marketing system based on data update and collation according to claim 6, wherein: The described strategy optimization module (32) uses a method combining experience replay and prioritized experience replay to adjust the strategy parameters in real time according to the advertisement placement effect, and the specific operation is as follows: During the advertisement placement process, multiple agents interact with the advertisement placement environment. Each agent will experience a state at each time step and select an action based on the current strategy for advertisement placement operations. Subsequently, a reward will be obtained. The reward is set by the advertisement conversion rate, and at the same time, it enters the next state. This set of information is stored as a sample in the experience replay pool. The experience replay pool is a first-in, first-out queue. When the queue is full, the earliest entered sample will be removed; Calculate a priority for each sample in the experience replay pool. The calculation of the priority is jointly determined by the reward magnitude, timestamp, and prediction error. Each time a batch of samples is sampled from the experience replay pool for training the deep reinforcement learning model, samples are sampled according to the priorities of the samples to obtain a batch of samples. These samples are input into the deep reinforcement learning model for training. According to the algorithm of deep reinforcement learning, calculate the loss function and update the parameters of the model through the backpropagation algorithm.

8. The advertising and marketing system based on data update and collation according to claim 7, characterized in that: In the real-time advertising effect evaluation index system established by the effect evaluation module (33), when calculating the advertising return on investment index, the delay effect of advertising placement is considered. The specific formula is as follows: Let the advertising placement time be T. Statistically calculate the revenue R within the time of T + ΔT after the advertising placement. The advertising placement cost is C. Considering that the revenue is generated within a period of time after the advertising placement, the weighted average method is used to calculate the delayed revenue, that is: where R t is the revenue at time t, ω t is the weight determined according to the distance between time t and the advertising release time T, then the advertising return on investment 9. The advertising and marketing system based on data update and collation according to claim 8, wherein: In the advertising placement monitoring unit (4), the anomaly detection algorithm based on streaming time series analysis uses an autoencoder model. The specific steps are as follows: Perform normalization processing on the monitoring data during the advertising placement process and use it as the input of the autoencoder. The monitoring data is the advertising conversion rate. The autoencoder consists of an encoder and a decoder; The encoder maps the input to a low-dimensional representation. The decoder reconstructs the low-dimensional representation into a new input of the autoencoder, and defines the reconstruction error by calculating the difference between the new input of the autoencoder and the input. Train the autoencoder to minimize the reconstruction error; During actual detection, for new monitoring data, calculate its reconstruction error. If the reconstruction error is greater than a pre-set threshold, determine that the data is an anomaly point and perform real-time monitoring. Then, according to the anomaly detection result, use the anomaly situation as the state input to the reinforcement learning model. The model outputs an adjustment action. By continuously interacting with the environment, the reinforcement learning model learns the optimal adjustment strategy under different anomaly states.

Images