Flow casting service profit prediction and evaluation method and system

Abstract

The invention aims to provide a flow casting business profit prediction and evaluation method and system, and belongs to the technical field of digital advertisement, and the method comprises the steps: collecting user click behaviors, advertisement attributes, cost consumption and other multi-dimensional data in real time through a multi-platform data interface, and carrying out the time sequence alignment processing of the data; comprising abnormal data detection and correction, feature vector extraction and standardized data set generation. Constructing an LSTM neural network model with a dynamic memory ability, and learning a long-term evolution rule of an advertisement effect through a memory unit state retention module and a forgetting gating mechanism; value evaluation of dynamic attenuation is executed based on a prediction result, a linear attenuation mode is adopted for brand advertisements, an exponential attenuation mode is adopted for effect advertisements, and a full-life-cycle income index is calculated; and finally generating a visual prediction trend. According to the method, the problems of insufficient data time sequence feature capture, long-term dependence modeling deficiency and static evaluation in a traditional method are solved, and accurate prediction and dynamic optimization of advertisement putting benefits are realized.

Description

Technical Field

[0001] The present invention relates to the field of digital advertising technology, and in particular to a method and system for predicting and evaluating profitability of a traffic investment business. Background Art

[0002] With the rapid development of the internet advertising industry, advertisers are increasingly demanding accurate evaluation and prediction of advertising effectiveness. Traditional advertising profit forecasting methods rely primarily on simple statistical analysis and empirical judgment, such as estimating revenue through historical data averages or linear regression models. However, advertising data exhibits complex time series characteristics and dynamic patterns, making it difficult for traditional statistical methods to accurately capture these characteristics. The multidimensional data generated during the advertising process is underutilized, and advertising effectiveness is influenced by multiple factors, including ad type, user behavior patterns, and market environment. The interplay of these factors further complicates forecasting.

[0003] Although some machine learning-based methods have been applied to advertising effectiveness prediction, most of these methods focus only on short-term click-through rate or conversion rate predictions and lack a comprehensive profit assessment of the entire life cycle of advertising. These methods typically use simple linear models to predict click-through rates, ignoring the changes in costs during the advertising process and the temporal dynamics of effectiveness. In addition, existing technologies have limited ability to handle abnormal data, and most use simple elimination or average filling methods, which may lead to biased prediction results. In terms of the dynamic memory capacity of the model, existing technologies also have obvious shortcomings and are unable to effectively store and utilize the long-term dependence characteristics of advertising exposure effects, thereby affecting the accuracy of predicting long-term trends in advertising effectiveness. Therefore, how to construct a profit prediction method that can fully utilize multidimensional data, accurately process abnormal data, have dynamic memory capabilities, and can comprehensively evaluate the entire life cycle of advertising has become a technical problem that needs to be solved in the current Internet advertising field. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a profit forecasting and evaluation method and system for investment and traffic business. By collecting multi-platform advertising data in real time and performing time alignment and exception processing, a neural network model with dynamic memory capability is constructed to achieve comprehensive profit evaluation of the entire life cycle of advertising. It can fully utilize multi-dimensional data, effectively process abnormal data, capture the evolution of advertising effects over time, and provide advertisers with accurate profit forecasts and decision support.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for evaluating profit forecasting of a flow investment business, comprising the following steps:

[0007] Step S1: Collecting advertising data in real time through a multi-platform data interface, the data including user click sequences, advertising material attributes, and cost consumption records;

[0008] Step S2: performing time series alignment processing on the data, including: detecting and correcting abnormal data points, extracting multi-dimensional feature vectors, and generating a standardized time series data set;

[0009] Step S3: constructing a neural network model with dynamic memory capability, using the data set generated in step S2 to train the parameters of the neural network model so that the neural network model learns the evolution of advertising effectiveness over time;

[0010] Step S4: Based on the prediction results of the neural network model, a dynamic attenuation value assessment is performed to calculate the comprehensive revenue index of the entire life cycle of the advertising;

[0011] Step S5: Based on the prediction results and the value assessment, a visual interactive interface is generated to display the predicted return trend and risk distribution.

[0012] Beneficial Effects: Through real-time data collection from multiple platforms, comprehensive advertising data is integrated, providing a rich foundation for subsequent processing. Time series alignment and anomaly correction improve data quality and reliability. A dynamic memory neural network model is constructed to accurately learn the dynamic evolution of advertising effectiveness, making predictions more realistic. Dynamic decay value assessment comprehensively considers the entire advertising lifecycle to provide a more scientific profit assessment. A visual interactive interface intuitively displays forecast results and risk distribution, facilitating quick decision-making for advertisers. Overall, this significantly improves the accuracy and scientific nature of profit forecasts for advertising traffic business, providing strong decision-making support for advertisers.

[0013] Furthermore, the abnormal data correction in step S2 includes:

[0014] Conduct discreteness analysis on user click sequences and construct a sliding time window to calculate the normal data distribution range;

[0015] The historical data interpolation method is used to replace abnormal values that are beyond the distribution range.

[0016] Beneficial effects: The dispersion analysis is combined with the sliding time window to calculate the normal data distribution range, and the historical data interpolation method is used to replace the outliers, accurately correcting abnormal data points and avoiding the interference of abnormal data on model training. It effectively improves data quality, enhances the accuracy and stability of model predictions, and makes profit forecast results more reliable.

[0017] Furthermore, the dispersion analysis specifically includes:

[0018] Calculate the data volatility index in each time window;

[0019] When the fluctuation rate of three consecutive sampling points exceeds 200% of the benchmark value, it is determined to be an abnormal fluctuation range;

[0020] The data within the abnormal interval is smoothed using the weighted average of the previous and next periods.

[0021] Beneficial effects: The specific dispersion analysis method further improves the accuracy and flexibility of abnormal data correction by calculating the volatility index and setting clear criteria for determining the abnormal fluctuation range, and smoothing the abnormal data with the weighted average of the previous and subsequent time periods. It better adapts to the data characteristics in different scenarios, ensures the continuity and consistency of the data, provides higher-quality data support for model training, and improves prediction accuracy.

[0022] Furthermore, the dynamic memory capability in step S3 is achieved by:

[0023] A memory unit state retention module is set up in the hidden layer of the neural network to dynamically store the long-term dependency characteristics of advertising exposure effects;

[0024] Set up a forget gating mechanism to adjust the retention ratio of historical memory according to new input data.

[0025] Beneficial Effects: By incorporating a memory cell state retention module and a forget gate mechanism into the neural network's hidden layer, the model is endowed with dynamic memory capabilities, enabling it to dynamically store the long-term dependency characteristics of ad exposure effects and adjust the proportion of historical memory retained based on new input data. This effectively addresses the inability of existing models to effectively utilize long-term dependency characteristics, significantly enhancing the model's ability to learn the dynamic evolution of advertising effects, improving the accuracy and adaptability of profit forecasts, and more accurately capturing the changing patterns of advertising effectiveness over time.

[0026] Furthermore, the workflow of the memory unit state retention module includes:

[0027] Calculate the forgetting factor based on the current input features and historical state;

[0028] Selectively erasing historical memory information based on the forgetting factor;

[0029] Write the new feature information into the updated memory unit according to the weight ratio.

[0030] Beneficial Effects: The workflow of the memory unit state retention module is clarified, including the calculation of the forgetting factor, the selective erasure of historical memory information, and the weighted writing of new feature information, making the model's dynamic memory mechanism clearer and more efficient. This refined dynamic memory management method further optimizes the model's storage and update process for long-term dependent features, improving the model's learning efficiency and prediction accuracy, ensuring that the model can better adapt to the dynamic changes in advertising data and providing a more reliable basis for profit forecasting.

[0031] Furthermore, the dynamic attenuation value assessment in step S4 includes:

[0032] Establishing a time value decay curve, wherein the decay rate of the curve is adaptively adjusted according to the advertisement type;

[0033] Perform weighted cumulative calculation of the predicted returns over time;

[0034] Perform dynamic ratio analysis on weighted cumulative value and real-time cost consumption.

[0035] Beneficial Effects: This method establishes a time-value decay curve and adaptively adjusts the decay rate based on ad type. It also performs a time-weighted cumulative calculation of predicted revenue and dynamic ratio analysis, enabling a dynamic assessment of the comprehensive revenue over the entire advertising lifecycle. This method overcomes the shortcomings of existing technologies, which often rely on one-sided and static approaches to advertising profitability assessment. It more comprehensively and accurately reflects the true profitability of advertising, providing advertisers with a more valuable reference for decision-making and effectively improving the economic benefits of advertising.

[0036] Furthermore, the adaptive adjustment is achieved by:

[0037] A linear decay mode is used for brand advertising, with the decay coefficient decreasing by a fixed percentage each cycle;

[0038] An exponential decay model is used for performance advertising, and the decay coefficient decreases in a power law manner over time.

[0039] Beneficial effects: The specific method of adaptive adjustment is clarified, and a linear attenuation mode is adopted for brand advertising, and an exponential attenuation mode is adopted for performance advertising. This targeted attenuation mode design accurately matches the profit characteristics of different types of advertising, making the time value attenuation curve more in line with reality, further improving the scientificity and accuracy of dynamic attenuation value evaluation, providing a more accurate evaluation method for profit prediction of different types of advertising, and enhancing the versatility and practicality of the system.

[0040] In a second aspect, the present invention provides a profit forecasting and evaluation system for a traffic investment business, which is used in any of the profit forecasting and evaluation methods for a traffic investment business described in the first aspect, comprising:

[0041] The data collection module obtains user click behavior data and advertising delivery parameters in real time through multiple advertising platform API interfaces and distributed crawler components;

[0042] The preprocessing module is connected to the data acquisition module and includes:

[0043] an anomaly detection unit configured to identify distribution density and anomalies of the data based on a sliding time window;

[0044] a feature extraction unit configured to generate feature values of click-through conversion rate and user retention rate;

[0045] The prediction model module is deployed in the server cluster and includes an LSTM neural network processor and a model training unit. The hidden layer dimension of the LSTM neural network processor is set as an adjustable parameter between 64 and 256.

[0046] A dynamic evaluation module, connected to the prediction model module and configured to calculate a time-decayed profit indicator; including:

[0047] A time decay calculator configured to perform a dynamic discounting calculation for returns;

[0048] Risk assessment matrix generator, outputs risk level distribution data;

[0049] The visualization module is used to output visualization charts including prediction curves and risk indicators.

[0050] Beneficial Effects: The profit forecasting and evaluation system for advertising traffic business encompasses a full range of modules, including data collection, preprocessing, forecasting models, dynamic evaluation, and visualization, providing an integrated solution from data acquisition to forecast result presentation. These modules work together to fully leverage the advantages of the proposed method, improving the efficiency and accuracy of profit forecasting for advertising traffic business. This provides advertisers with a comprehensive and user-friendly profit forecasting and evaluation tool, helping to enhance the efficiency and effectiveness of advertising placement decisions.

[0051] Furthermore, the anomaly detection unit sets a 24-hour sliding time window, and the amount of data in the sliding time window is not less than 1,000 samples.

[0052] Beneficial Effects: The anomaly detection unit sets a 24-hour sliding window and stipulates that the data volume within the window must be no less than 1,000 samples, ensuring the timeliness and data sufficiency of anomaly detection. The longer sliding window better captures short-term dynamic changes in data, while the sufficient sample size provides a data foundation for accurate detection of anomalies. This further improves the accuracy and reliability of anomaly detection, provides higher-quality data for subsequent data processing and model training, and enhances the overall performance of the system.

[0053] Furthermore, the feature extraction unit includes a feature selector configured to preferentially extract user behavior features with a conversion rate higher than 1%.

[0054] Beneficial Effects: A feature selector was added to the feature extraction unit to prioritize user behavior features with conversion rates above 1%, improving the relevance and effectiveness of feature extraction. By selecting more representative and valuable features, the interference of invalid or redundant features on model training was reduced, improving the model's learning efficiency and prediction accuracy. This enabled the system to more accurately assess the profitability of advertising, enhancing its practicality and accuracy.

[0055] In summary, compared with the existing technology, the present invention significantly improves the data processing capability through multi-platform real-time data collection and advanced pre-processing technology, can efficiently integrate advertising data and accurately correct abnormal data, and provide a high-quality data foundation for subsequent predictions. At the same time, the dynamic memory neural network model constructed by the present invention can learn the evolution law of advertising effects over time, and combined with the dynamic attenuation value assessment method, it greatly improves the prediction accuracy and overcomes the one-sided and static limitations of the prediction method in the existing technology. In addition, the present invention adaptively adjusts the attenuation rate according to the type of advertisement, introduces a variety of dynamic mechanisms, enhances the adaptability and flexibility of the system, and can better adapt to the characteristics of advertising data in different scenarios. The generated visual interactive interface can intuitively display the predicted revenue trend and risk distribution, providing advertisers with a more scientific and convenient decision support tool, and solving the problem of non-intuitive display of prediction results in the existing technology. The present invention also provides a one-stop full-process solution, covering all links from data collection to prediction result display, achieving efficient collaboration of each link, and improving the overall efficiency of profit forecasting of advertising flow business. In summary, the present invention is superior to existing technologies in terms of data processing, prediction accuracy, adaptability, decision support, and system integration. It can effectively improve the profit prediction level of advertising flow business and bring significant economic benefits to advertisers. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present invention and, together with the description, serve to explain the principles of the present invention.

[0057] The present invention can be more clearly understood from the following detailed description with reference to the accompanying drawings, in which:

[0058] Figure 1 This is an overall flow chart of the profit forecasting and evaluation method for investment flow business provided by an embodiment of the present invention;

[0059] Figure 2 This is a sub-flowchart of data timing alignment processing in the profit forecasting and evaluation method for investment flow business provided by an embodiment of the present invention;

[0060] Figure 3 This is a sub-flowchart for constructing a dynamic memory neural network model in the profit forecasting and evaluation method for investment flow business provided by an embodiment of the present invention;

[0061] Figure 4 This is a sub-flow chart of dynamic attenuation value assessment in the profit forecasting and assessment method for investment flow business provided by an embodiment of the present invention;

[0062] Figure 5 This is an architecture diagram of a profit forecasting and evaluation system for investment and flow business provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations of the technical solution of the present application. In the absence of conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0064] The term "and / or" in this document simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " in this document generally indicates an "or" relationship between the related objects.

[0065] Example 1

[0066] like Figures 1-4 As shown, the embodiment of the present invention provides a method for predicting and evaluating the profitability of a flow investment business, comprising the following steps:

[0067] Step S1: Collect advertising delivery data in real time through a multi-platform data interface, the data including user click sequences, advertising material attributes and cost consumption records; in some embodiments, the user click sequence includes timestamp, hourly / daily clicks, page dwell time, number of pages viewed, behavior chain from click to final conversion, and user attribute labels (such as age group, gender, interests and hobbies); advertising material attributes include ad location (such as header banner, information flow, sidebar), form, size, creative elements; cost consumption records: including consumption amount by time dimension (can be hourly, daily, weekly), cost data by billing model, budget execution progress (can be the proportion of consumed amount to the total budget), bid adjustment records (including bid increase / decrease time point and amplitude), cost distribution and consumption proportion of different delivery channels.

[0068] Step S2: performing time series alignment processing on the data, including: detecting and correcting abnormal data points, extracting multi-dimensional feature vectors, and generating a standardized time series data set;

[0069] Step S3: Construct a neural network model with dynamic memory capability, and use the data set generated in step S2 to train the parameters of the neural network model so that the neural network model learns the evolution of advertising effects over time;

[0070] Step S4: Based on the prediction results of the neural network model, perform dynamic attenuation value assessment and calculate the comprehensive benefit indicators of the entire life cycle of advertising. The prediction results refer to the predicted values of the neural network model for the effect indicators that may be produced by advertising in a future period of time, specifically including: time series prediction data, effect prediction indicators, and investment return predictions. The time series prediction data can be the expected daily click volume in the next 7 days, the expected conversion volume in the future period, and the new user acquisition volume in the future period. The effect prediction indicators can be the expected click-through rate (CTR) change trend, the expected conversion rate (CVR) change trend, the expected average order value and repurchase rate, and the user retention change trend. The investment return prediction can be: the expected cost per click (CPC) change trend, the expected cost per acquisition (CAC) estimation, the return on investment (ROI) prediction of advertising, and the prediction of the contribution of each channel effect. These prediction results together constitute a comprehensive prediction of the future effect of advertising and provide basic data for subsequent dynamic attenuation value assessment.

[0071] Step S5: Based on the prediction results and the value assessment, a visual interactive interface is generated to display the predicted return trend and risk distribution.

[0072] Furthermore, the abnormal data correction in step S2 includes:

[0073] Conduct discreteness analysis on user click sequences and construct a sliding time window to calculate the normal data distribution range;

[0074] Use historical data interpolation method to replace abnormal values outside the distribution range;

[0075] Preferably, a 24-hour sliding window statistical method is used to calculate data distribution characteristics:

[0076]

[0077] Where W = 24 represents the time window length (hours); μ is the window mean; σ represents the standard deviation; x i is the original data sequence; when a data point is detected When the historical interpolation correction is performed, specifically, Where Δ=24 represents the historical time offset.

[0078] Furthermore, the dispersion analysis includes:

[0079] Calculate the volatility indicator of the data within each time window;

[0080] When the fluctuation rate of three consecutive sampling points exceeds 200% of the benchmark value, it is determined to be an abnormal fluctuation range;

[0081] The data in the abnormal interval are smoothed by using the weighted average of the previous and next periods;

[0082] Preferably, the volatility indicator is calculated by the following formula:

[0083]

[0084] Where: V t represents the volatility during period t; x t Indicates the data value of the current period;

[0085] When three consecutive sampling points meet V t When the value is greater than 200%, weighted smoothing is performed on the abnormal interval data, as shown in the following formula:

[0086]

[0087] Among them, 0.7 is the weight of the forward period and 0.3 is the weight of the backward period.

[0088] Furthermore, the dynamic memory capability in step S3 is achieved by:

[0089] A memory unit state retention module is set up in the hidden layer of the neural network to dynamically store the long-term dependency characteristics of advertising exposure effects;

[0090] Set up a forget gating mechanism to adjust the retention ratio of historical memory according to new input data.

[0091] Specifically, the workflow of the memory cell state retention module includes:

[0092] Calculate the forgetting factor based on the current input features and historical state;

[0093] Selectively erase historical memory information based on the forgetting factor;

[0094] Write the new feature information into the updated memory unit according to the weight ratio;

[0095] Specifically, the workflow of the memory cell state retention module is implemented through the LSTM architecture:

[0096] f t =σ(W f [h t-1 ,xt ]+b f )

[0097] C t =f t ·C t-1 +i t tanh(W c [h t-1 ,x t ]+b c )

[0098] Where, f t ∈(0,1) represents the output value of the forget gate; C t is the updated memory unit state; h t-1 Represents the hidden layer output at the previous moment; x t Represents the current input feature vector; W f ,W c Both are trainable weight matrices; b f ,b c All are bias terms.

[0099] Specifically, the dynamic attenuation value assessment in step S4 includes:

[0100] Establish a time value decay curve, and the decay rate of the curve is adaptively adjusted according to the type of advertisement;

[0101] Perform weighted cumulative calculation of the predicted returns over time;

[0102] Perform dynamic ratio analysis on weighted cumulative value and real-time cost consumption.

[0103] Furthermore, adaptive adjustment is achieved by using different attenuation modes for two different types of advertisements:

[0104] A linear decay mode is used for brand advertising, with the decay coefficient decreasing by a fixed percentage each cycle;

[0105] An exponential decay model is used for performance advertising, and the decay coefficient decreases in a power law manner over time;

[0106] Brand advertising refers to advertising primarily aimed at increasing brand awareness, reputation, and loyalty. Its characteristics include a focus on shaping and disseminating brand image; a long lifespan, typically requiring sustained exposure to build brand recognition; a lengthy conversion path, potentially involving multiple encounters from exposure to the final purchase decision; and effectiveness evaluation metrics including brand mentions, brand recognition, and brand favorability. The impact of brand advertising typically decays evenly over time. Once brand recognition is established, it typically weakens at a relatively steady rate. Linear decay more accurately reflects the natural fading of brand memory. For example, the decay coefficient is 0.8 in the first cycle, decreases by 5% in the second cycle to 0.76, and then decreases by another 5% in the third cycle to 0.722.

[0107] Performance advertising refers to a form of advertising that aims to directly drive user behavior conversion. Its characteristics include: focusing on immediate and quantifiable conversion results, such as clicks, registrations, purchases, etc.; a short effect cycle, usually expecting direct conversions in a short period of time; a relatively short conversion path, emphasizing immediate action; and performance evaluation indicators including click-through rate, conversion rate, and return on investment. Conversions of performance advertising mostly occur within a short period of time after exposure, after which the conversion rate drops sharply. As time goes by, the user's willingness to buy will decrease at an accelerating rate. Exponential decay can better capture the "short-term burst, weak long tail" characteristics of performance advertising. For example: the attenuation coefficient is initially 0.8, and then decreases as time t increases. The regular decay shows the characteristics of rapid decline in the early stage and slow decline in the later stage.

[0108] By adopting differentiated decay models for different types of advertising, we can more accurately assess the true value of various types of advertising during their life cycle, thereby achieving more scientific investment decisions.

[0109] Specifically, brand advertising uses linear decay: D b (t) = D0(1-k l t); Effective advertising uses exponential decay: The life cycle benefit calculation formula is:

[0110] Among them, D0 is the initial attenuation coefficient, the default value is 0.8; k l is a linear decay rate, 0.05 / cycle for brand advertising; k e is an exponential decay rate, effective advertising 0.12 / cycle; r = 5%: base discount rate; R t Represents the predicted return at time t output by the neural network.

[0111] like Figure 5 As shown, as an embodiment, a profit forecasting and evaluation system for investment and flow business is used to execute any profit forecasting and evaluation method for investment and flow business, including:

[0112] The data collection module obtains user click behavior data and advertising delivery parameters in real time through multiple advertising platform API interfaces and distributed crawler components;

[0113] The preprocessing module is connected to the data acquisition module and includes:

[0114] an anomaly detection unit configured to identify distribution density and anomalies of the data based on a sliding time window;

[0115] a feature extraction unit configured to generate feature values of click-through conversion rate and user retention rate;

[0116] The prediction model module is deployed in the server cluster and includes an LSTM neural network processor and a model training unit. The hidden layer dimension of the LSTM neural network processor is set as an adjustable parameter between 64 and 256.

[0117] The dynamic evaluation module is connected to the prediction model module and is configured to calculate the time-decayed profit indicator; it includes:

[0118] A time decay calculator configured to perform a dynamic discounting calculation for returns;

[0119] Risk assessment matrix generator, outputs risk level distribution data;

[0120] The visualization module is used to output visualization charts including prediction curves and risk indicators.

[0121] Specifically, the anomaly detection unit sets a 24-hour sliding time window, and the amount of data in the sliding time window is not less than 1,000 samples.

[0122] Preferably, the feature extraction unit includes a feature selector configured to preferentially extract user behavior features with a conversion rate higher than 1%.

[0123] The present invention introduces an abnormal data detection and correction mechanism based on discrete degree analysis and sliding time windows, which can accurately identify and correct abnormal data points in user click sequences. It can not only detect statistical outliers, but also determine the abnormal fluctuation range, and use the weighted average of the previous and next time periods for smoothing. A memory unit state retention module is set in the hidden layer of the neural network, which can dynamically store the long-term dependent characteristics of the advertising exposure effect, and a forgetting gating mechanism is set to automatically adjust the retention ratio of historical memory according to the new input data. The implementation of this LSTM architecture enables the model to better learn the evolution of advertising effects over time, and also adaptively adjust the decay rate according to the type of advertisement, using a linear decay mode for brand advertisements and an exponential decay mode for performance advertisements. This differentiated processing is more in line with the actual effect decay characteristics of different types of advertisements. The present invention not only focuses on short-term click-through rate or conversion rate, but also calculates the comprehensive revenue indicators of the entire life cycle of advertising. It uses the time value decay curve to perform weighted cumulative calculation of the predicted revenue in the time dimension, which more comprehensively reflects the actual profitability of advertising. It solves the problems of inaccurate data processing, lack of dynamic memory ability of the prediction model, one-sided static evaluation method and non-intuitive display of prediction results in traditional advertising profit prediction, and significantly improves the accuracy and practicality of profit prediction of advertising flow business.

[0124] Example 2

[0125] This embodiment provides a process for establishing a neural network model with dynamic memory capability, and the steps are as follows:

[0126] (1) Data preparation and preprocessing:

[0127] Before building the neural network model, we collected real-time ad placement data through multi-platform data interfaces and completed time series alignment processing, including abnormal data detection and correction, feature vector extraction, and standardized data set generation, in accordance with steps S1 and S2 in the claims. This processed data served as input to the neural network model.

[0128] (2) Dynamic memory capability of building neural network models:

[0129] According to step S3 in the claims, the dynamic memory capability of the neural network model is achieved by:

[0130] (2.1) Set up the memory unit state retention module:

[0131] A cell state retention module is implemented in the hidden layer of the neural network to dynamically store the long-term dependencies of ad exposure. The cell state is the core structure of the LSTM network, storing information and passing it to subsequent time steps, thereby enabling the storage of long-term dependencies.

[0132] (2.2) Set up the forget gate mechanism:

[0133] A forget gate mechanism is set up to adjust the retention ratio of historical memory based on new input data. The forget gate determines which information in the memory unit needs to be forgotten, thereby preventing outdated or irrelevant information from interfering with the current prediction.

[0134] (3) Workflow of the memory unit state retention module

[0135] According to step S4 and step S5 in the claims, the workflow of the memory unit state retention module includes the following steps:

[0136] (3.1) Calculate the forgetting factor: Calculate the forgetting factor based on the current input features and historical state. The forgetting factor determines which information in the memory unit needs to be forgotten. Its calculation formula is:

[0137] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0138] Among them, f t represents the forgetting factor; σ represents the Sigmoid activation function, which is used to limit the output value to between (0,1); W f and b f are the weight matrix and bias term of the forget gate respectively; h t-1 Indicates the hidden layer state at the previous moment; x t Represents the input data at the current moment.

[0139] (3.2) Selectively erase historical memory information: Selectively erase historical memory information based on the forgetting factor. The updated memory unit state is:

[0140] c t =f t c t-1

[0141] Among them, c t Indicates the current state of the memory unit; c t-1 Indicates the state of the memory unit at the previous moment. t , the model can selectively retain or forget historical memory information, thereby dynamically adjusting the content of the memory unit.

[0142] (3.3) Calculating the input gate and candidate memory cells: The input gate determines which new information needs to be written into the memory cell, and the candidate memory cell is the specific content of the new information. The calculation formulas are:

[0143] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0144]

[0145] Among them, i t represents the input gate; Represents a candidate memory unit; tanh represents the hyperbolic tangent activation function, which is used to limit the output value to (-1, 1); W i 、b i 、W c and b c are the weight matrix and bias term of the input gate and candidate memory unit respectively.

[0146] (3.4) The new feature information is written into the updated memory unit according to the weight ratio. The update formula is:

[0147]

[0148] Through the input gate i t and candidate memory cells The model can write new feature information into the memory unit in proportion to the weight, thereby updating the content of the memory unit.

[0149] (3.5) Calculate the output gate and the current hidden layer state

[0150] Calculate the output gate and the current hidden layer state. The output gate determines which information in the memory unit needs to be output to the hidden layer state. The calculation formula is:

[0151] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0152] h t =o t ·tanh(c t )

[0153] Among them, t represents the output gate; h t Represents the hidden layer state at the current moment. Through the output gate o t, the model can selectively output the information in the memory unit to the hidden layer state, thereby providing a basis for subsequent predictions.

[0154] (4) Model training and optimization

[0155] The neural network model is trained using the preprocessed dataset. The training process includes forward propagation, loss function calculation, backpropagation, and iterative optimization. During training, the model uses its dynamic memory to learn the evolution of advertising effectiveness over time and dynamically adjusts the contents of its memory cells based on new input data, thereby accurately predicting the effectiveness of advertising.

[0156] Through the above steps, we built a neural network model with dynamic memory capabilities, which can effectively capture the long-term dependency characteristics of advertising effectiveness and provide accurate prediction results for subsequent dynamic attenuation value evaluation.

[0157] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for profit forecasting and evaluation of investment flow business, characterized in that: The following steps are involved: Step S1: Collecting advertising delivery data in real time through a multi-platform data interface, the data including user click sequences, advertising material attributes, and cost consumption records; wherein the user click sequence includes the timestamp of user interaction, click frequency, browsing depth, and user attribute tags; the advertising material attributes include ad location, format, size, and target audience targeting parameters; the cost consumption records include the time dimension consumption amount, billing model cost, and budget execution progress; Step S2: performing time series alignment processing on the data, including: detecting and correcting abnormal data points, extracting multi-dimensional feature vectors, and generating a standardized time series data set; Step S3: constructing a neural network model with dynamic memory capability, using the data set generated in step S2 to train the parameters of the neural network model so that the neural network model learns the evolution of advertising effectiveness over time; Step S4: Based on the prediction results of the neural network model, a dynamic decay value assessment is performed to calculate the comprehensive revenue indicators of the entire life cycle of the advertising; wherein the prediction results include the expected click volume, conversion volume, click-through rate change trend, conversion rate change trend and return on investment forecast in the future period; Step S5: Generate a visual interactive interface based on the prediction results and the value assessment to display the predicted return trend and risk distribution.

2. A method for profit forecasting and evaluation of investment flow business according to claim 1, characterized in that: The detection and correction of abnormal data points in step S2 includes: Conduct discreteness analysis on user click sequences and construct a sliding time window to calculate the normal data distribution range; The historical data interpolation method is used to replace abnormal values that are beyond the distribution range.

3. A method for profit forecasting and evaluation of investment flow business according to claim 2, characterized in that: The dispersion analysis specifically includes: Calculate the data volatility index in each time window; When the fluctuation rate of three consecutive sampling points exceeds 200% of the benchmark value, it is determined to be an abnormal fluctuation range; The data within the abnormal interval is smoothed using the weighted average of the previous and next periods.

4. The method for profit forecasting and evaluation of investment flow business according to claim 1, characterized in that: The dynamic memory capability of the neural network model in step S3 is achieved by: A memory unit state retention module is set up in the hidden layer of the neural network to dynamically store the long-term dependency characteristics of advertising exposure effects; Set up a forget gating mechanism to adjust the retention ratio of historical memory according to new input data.

5. A method for profit forecasting and evaluation of investment flow business according to claim 4, characterized in that: The workflow of the memory unit state retention module includes: Calculate the forgetting factor based on the current input features and historical state; Selectively erasing historical memory information based on the forgetting factor; Write the new feature information into the updated memory unit according to the weight ratio.

6. The method for profit forecasting and evaluation of investment flow business according to claim 1, characterized in that: The dynamic attenuation value assessment in step S4 includes: Establishing a time value decay curve, wherein the decay rate of the curve is adaptively adjusted according to the advertisement type; Perform weighted cumulative calculation of the predicted returns over time; Perform dynamic ratio analysis on weighted cumulative value and real-time cost consumption.

7. A method for profit forecasting and evaluation of investment flow business according to claim 6, characterized in that: The adaptive adjustment is achieved by: A linear attenuation model is adopted for brand advertising, with the attenuation coefficient decreasing by a fixed percentage each cycle; wherein brand advertising refers to an advertising form whose primary goal is to enhance brand awareness, reputation, and loyalty; An exponential decay model is adopted for performance advertising, and the decay coefficient decreases in a power law manner over time; the performance advertising mentioned herein refers to an advertising form whose main goal is to directly drive user behavior conversion.

8. A profit forecasting and evaluation system for investment and flow business, used to execute a profit forecasting and evaluation method for investment and flow business according to any one of claims 1 to 7, characterized in that: include: The data collection module obtains user click behavior data and advertising delivery parameters in real time through multiple advertising platform API interfaces and distributed crawler components; The preprocessing module is connected to the data acquisition module and includes: an anomaly detection unit configured to identify distribution density and anomalies of the data based on a sliding time window; a feature extraction unit configured to generate feature values of click-through conversion rate and user retention rate; The prediction model module is deployed in the server cluster and includes an LSTM neural network processor and a model training unit. The hidden layer dimension of the LSTM neural network processor is set as an adjustable parameter between 64 and 256. A dynamic evaluation module, connected to the prediction model module, configured to calculate a time-decayed revenue indicator, including: a time-decay calculator, configured to perform a dynamic revenue discount operation; Risk assessment matrix generator, outputs risk level distribution data; The visualization module is used to output visualization charts including prediction curves and risk indicators.

9. The profit forecasting and evaluation system for investment and flow business according to claim 8, characterized in that: The anomaly detection unit sets a 24-hour sliding time window, and the amount of data in the sliding time window is not less than 1,000 samples.

10. The profit forecasting and evaluation system for investment and flow business according to claim 8, characterized in that: The feature extraction unit includes a feature selector configured to preferentially extract user behavior features with a conversion rate higher than 1%.

Images