Marketing decision analysis method based on artificial intelligence

Abstract

The invention discloses a marketing decision analysis method based on artificial intelligence, and the method comprises the steps: S1, obtaining the marketing data of a plurality of channels, carrying out the preprocessing of the marketing data, and carrying out the classification of the preprocessed marketing data according to the data type; s2, obtaining a marketing scene demand, obtaining a corresponding analysis model according to the marketing scene demand, and generating a marketing decision-making scene through the intelligent decision-making model based on the analysis model; s3, acquiring user behavior data, then collecting behavior feedback of the user on the marketing decision scene, merging user behavior characteristics into user portrait data, and analyzing the user portrait data; and S4, extracting feature information in the marketing decision scene, constructing a marketing effect evaluation index according to the extracted feature information, and predicting a marketing effect suitable for the marketing decision scene from the historical marketing data based on the marketing effect evaluation index. According to the method, the problem of data islands when the multi-source heterogeneous marketing data is processed by a traditional method is solved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent marketing technology, and specifically relates to a marketing decision analysis method based on artificial intelligence. Background Art

[0002] Marketing decision analysis methods use specific technologies, tools, or processes to assist marketers, decision makers, or business users in making more efficient decisions during marketing activities. These methods encompass a wide range of areas, from user behavior analysis and market data processing to marketing effectiveness evaluation and personalized recommendations. They aim to reduce the subjectivity of human judgment and improve the accuracy and efficiency of marketing decisions. These methods can be rule-based, such as keyword matching and statistical analysis, or incorporate artificial intelligence technologies like machine learning, deep learning, and natural language processing. These methods help users analyze market trends, predict marketing effectiveness, and provide decision-making recommendations, thus playing a vital role in business practice and marketing management.

[0003] Existing methods often rely on static historical data to analyze marketing scenarios, lacking dynamically generated intelligent marketing scenarios. This makes it impossible to provide decision makers with real-time market insights. Furthermore, existing marketing effectiveness evaluation methods are overly simplistic and rigid, often based on a single metric, making it difficult to ensure accurate and comprehensive marketing decisions. Traditional methods, when processing heterogeneous marketing data from multiple sources, suffer from data silos and are unable to effectively integrate multi-dimensional information such as user behavior, transaction records, and market feedback. This results in a lack of a holistic perspective and intelligent support for marketing decisions.

[0004] Therefore, a marketing decision analysis method based on artificial intelligence is provided. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a marketing decision analysis method based on artificial intelligence, which solves the data island problem existing in traditional methods when processing multi-source heterogeneous marketing data.

[0006] The technical solution to achieve the above purpose is: A marketing decision analysis method based on artificial intelligence, comprising: Step S1: acquiring marketing data from multiple channels, preprocessing the marketing data, and then classifying the preprocessed marketing data according to data type; Step S2: Obtain marketing scenario requirements, obtain corresponding analysis models based on the marketing scenario requirements, and generate marketing decision scenarios through an intelligent decision model based on the analysis models; Step S3: Obtain user behavior data, then collect user behavioral feedback on marketing decision scenarios, merge user behavior features into user portrait data and analyze them; Step S4: extracting characteristic information from the marketing decision scenario, then constructing a marketing effect evaluation index based on the extracted characteristic information, and predicting the marketing effect applicable to the marketing decision scenario from historical marketing data based on the marketing effect evaluation index; Step S5: Generate personalized marketing suggestions for marketing decision scenarios based on marketing data.

[0007] Preferably, the step S1 includes: Step S11, obtaining user data and transaction data from any marketing data; Step S12: pre-processing the user data in the marketing data to obtain a first data set; Step S13, pre-processing the transaction data in the marketing data to obtain a second data set; Step S14: Process each marketing data set according to steps S11 to S13 to obtain a first data set and a second data set corresponding to each marketing data set, merge the first data sets of all marketing data sets into user basic data, and merge the second data sets of all marketing data sets into transaction behavior data; Step S15, constructing a corresponding feature word library according to the data type; Step S16: number the data types and count the frequency of occurrence of each feature word in the feature word library corresponding to any data type in the user basic data. and importance , calculate the intelligent feature scores of feature lexicons corresponding to different data types in user basic data according to the formula ; In step S17, the calculated smart feature scores are sorted, and the data type corresponding to the highest smart feature score is used as the data type of the marketing data. Then, the smart feature scores of each marketing data are calculated one by one to obtain the data type corresponding to each marketing data, and the corresponding marketing data are classified and saved according to the data type.

[0008] Preferably, in step S12, the user data in the marketing data is standardized, and then redundant information is removed to obtain first text information, which is then merged to obtain a first data set, wherein the first text information is classified according to data stability based on data time attribute analysis, and the classification standard is as follows: data with a change cycle greater than 6 months is long-cycle data, data with a change cycle between 1-6 months is medium-cycle data, and data with a change cycle less than 1 month is short-cycle data; In step S13, invalid transactions in the transaction data corresponding to the marketing data are determined to be redundant information, and then the redundant information in the transaction data is removed to obtain a second data set.

[0009] Preferably, in step S16, the time-degradation intelligent feature scoring algorithm TDIFSA is used to calculate the intelligent feature score , the calculation formula is as follows: ; Where, is the feature word number, and its value range is , is the total number of feature words, indicating the size of the feature word library. Numbers are used to distinguish different types of marketing data. is the frequency of feature words, and its value range is , indicating the The feature word in Normalized frequency of occurrence in class data, is the feature word importance coefficient, and its value range is , reflecting the weight importance of feature words, is the trend value of the feature word in the time series, , the larger the value, the more obvious the upward trend. is the dispersion of feature words, , measures the degree of discreteness of the distribution of feature words in the data, is the time decay coefficient of the feature word, , used to calculate the time decay effect, is the attenuation control parameter, , controlling the speed of aging decay; To dynamically adjust the above parameters: Build a parameter calibration feedback mechanism, optimize parameters using historical marketing effectiveness data, and trigger parameter recalibration when the prediction accuracy of the marketing effectiveness evaluation index falls below 80%. The calibration process uses a grid search method, traversing the parameter range in steps of 0.1 to select the parameter combination with the highest prediction accuracy. The parameter adjustment cycle is set to automatically calibrate every 30 days, or manually trigger calibration when the data distribution changes significantly. A hierarchical scoring strategy is adopted for feature words of different time periods, with long-term feature words accounting for 40% of the weight, medium-term feature words accounting for 35% of the weight, and short-term feature words accounting for 25% of the weight. A time window sliding mechanism is established, with a 12-month sliding window for long-term data, a 6-month sliding window for medium-term data, and a 1-month sliding window for short-term data. Data outside the time window will be gradually downgraded until it is removed. In the TDIFSA algorithm, the logarithmic function Used to realize nonlinear enhancement of trend value, square root function Used to provide normalization to prevent the value from being too large, exponential function Used to implement the time decay mechanism, the summation symbol Σ accumulates all feature words and finally divides them by Achieve averaging processing.

[0010] Preferably, in step S17, if the smart feature scores are the same, the smart feature scores of the feature lexicons corresponding to different data types in the transaction behavior data are calculated, and the marketing data are classified and saved according to the data type based on the smart feature scores.

[0011] Preferably, the step S2 includes: Step S21: Segment the marketing scenario requirements into words, input the segmented words into a feature word library, and then calculate the intelligent feature score of the marketing scenario requirements to obtain an analysis model of the marketing scenario requirements; Step S22: Building an intelligent decision-making model based on the analysis model corresponding to the marketing scenario requirements; Step S23: Compare the constructed marketing decision scenario with any marketing data. If the marketing decision scenario is identical to any marketing data, remove the corresponding marketing decision scenario and reconstruct the marketing decision scenario. If the marketing decision scenario is different from any marketing data, proceed to the next step. The step S22 includes: Step S221: Based on the analysis model corresponding to the marketing scenario requirements, the corresponding data type is obtained from the classified marketing data, the marketing data of the corresponding data type is segmented, and then the words are classified to obtain user characteristics, product characteristics, channel characteristics, and time characteristics, which are combined into feature data of the marketing data; Step S222, constructing an intelligent decision-making model; Step S223: randomly select characteristic data of the classified marketing data, and then add the characteristic data of the marketing data to the intelligent decision-making model to construct a marketing decision scenario.

[0012] Preferably, step S3 includes: Step S31: construct a user behavior database and obtain user behavior data based on the user behavior database; Step S32, comparing any behavioral feature in the user profile data with the user behavior data one by one; Step S33: If the corresponding features in the user portrait data and the user behavior data are the same, proceed to the next step; if any feature in the user portrait data is different from the corresponding feature in the user behavior data, the user portrait data is determined to be abnormal.

[0013] Preferably, step S4 includes: Step S41, obtaining product features and user features in the marketing decision scenario, and extracting first feature information of the product features and second feature information of the user features; Step S42: Segmenting the product features and user features, then labeling the segmented product features and user features, and extracting the content of specific tags to obtain first feature information of the product features and second feature information of the user features; Step S43: traverse the marketing cases in the historical marketing data one by one and calculate the semantic similarity between the first feature information and the marketing cases. and association strength , get the dynamic evaluation value of the first feature information ; Step S44: Calculate the semantic similarity between the second feature information and the marketing case and association strength , get the dynamic evaluation value of the second feature information ; Step S45: Calculate the marketing effect evaluation index using the multi-dimensional fusion intelligent evaluation algorithm MDFIEA ; Step S46: When the marketing effect evaluation index of any marketing case is greater than or equal to the evaluation index threshold, the marketing case is used as a reference case applicable to the marketing decision scenario, and one or more existing marketing cases applicable to the same marketing decision scenario are combined and recorded as the first marketing data; Step S47: When the marketing effect evaluation index is less than the evaluation index threshold, no operation is performed.

[0014] Preferably, in step S43, the semantic similarity between the first feature information and the marketing case The cosine similarity algorithm is used for calculation, and the formula is: ; Where, is the number of product feature dimensions, The first feature information dimensional feature vector value, The first marketing case The formula measures the similarity by calculating the cosine of the angle between two eigenvectors. The range is [0,1]. The closer to 1, the higher the similarity. The strength of the correlation between the first characteristic information and the marketing case The Pearson correlation coefficient algorithm is used for calculation, and the formula is: Where, is the mean of the first feature information eigenvector, is the mean of the first marketing case feature vector, is the standard deviation of the first eigenvector, is the standard deviation of the first marketing case eigenvector. This formula measures the degree of linear correlation between two eigenvectors. Its value range is [-1,1]. The larger the absolute value, the higher the correlation strength. Dynamic evaluation value of the first characteristic information and marketing case The calculation formula is: ; Where, and are weight coefficients, and ; In step S44, the semantic similarity between the second feature information and the marketing case is calculated. The cosine similarity algorithm is used for calculation, and the formula is: ; Where, is the number of user feature dimensions, The second feature information dimensional feature vector value, For the second marketing case The formula measures the similarity by calculating the cosine of the angle between two eigenvectors. The range is [0,1]. The closer to 1, the higher the similarity. The strength of the association between the second characteristic information and the marketing case The Pearson correlation coefficient algorithm is used for calculation, and the formula is: Where, is the mean of the second feature information eigenvector, is the mean of the feature vector of the second marketing case, is the standard deviation of the second eigenvector, is the standard deviation of the eigenvector of the second marketing case. This formula measures the degree of linear correlation between two eigenvectors. Its value range is [-1,1]. The larger the absolute value, the higher the correlation strength. Dynamic evaluation value The calculation formula is: ; In step S45, the marketing effect evaluation index is calculated by the multi-dimensional fusion intelligent evaluation algorithm MDFIEA. , the calculation formula is as follows: ; Where, is the product feature weight coefficient, and the constraints are , used to balance the importance of product features and user features, is the similarity adjustment parameter, , used to control the saturation speed of the hyperbolic tangent function, The larger the value, the more sensitive the function. is the correlation strength adjustment parameter, , used to adjust the impact of association strength on the evaluation results, is the synergistic effect coefficient, , used to control the intensity of the synergy between product features and user features; Hyperbolic tangent function The range of is (−1,1), when hour, ,when hour, , used to implement nonlinear mapping; and Used to provide stable normalization effect; Quantitative assessment for achieving synergistic effects; Used to calculate the difference between two evaluation values; Used to take the maximum of the two as the normalization factor.

[0015] Preferably, step S5 includes: Step S51: If all behavioral features in the user profile data match the first marketing data, then the optimal marketing recommendation is generated; if the number of features in the user profile data is less than the number of features in the first marketing data, or any user profile data does not match the first marketing data, or a new user and new product market are identified, then proceed to the next step; Step S52: Use Gaussian kernel enhanced intelligent matching algorithm GKEIMA to calculate the intelligent matching degree of marketing suggestions ; Step S53: When the smart matching degree is less than or equal to the matching degree threshold, a basic marketing suggestion is generated; when the smart matching degree is greater than the matching degree threshold, a personalized marketing suggestion is generated; In step S51, whether the user is a new user is determined by: the number of historical behavior data records of the user is less than a preset threshold or the user registration time is less than a preset time window; When a new user is identified, a demographic-based initialization strategy is adopted, using historical data from user groups with similar demographic characteristics to infer the user profile. A new user feature vector is constructed, including basic attributes and inferred attributes. The inferred attributes are estimated using the average feature value of similar user groups. The criteria for determining whether a product is new is: the product's shelf life is less than the preset time window or the product's historical sales record is less than the preset threshold; When a new product is identified, a recommendation strategy based on product attribute similarity is adopted, calculating the similarity between the product feature vector and historical products. A new product feature matrix is ​​constructed, using the average market performance data of products in the same category as the initial prediction value. An uncertainty adjustment factor is set for the new product, and real market feedback data is collected through A / B testing and progressive promotion strategies. For the new user-new product combination scenario, a hybrid strategy combining content-based recommendations and collaborative filtering is adopted. Initial matching is performed using product content features and user basic features. Collaborative filtering is supplemented by using similar users' preference data for similar products. A conservative recommendation strategy is set, prioritizing product categories with high market validation before proceeding to subsequent steps. In step S52, the Gaussian kernel enhanced intelligent matching algorithm GKEIMA is used to calculate the intelligent matching degree of the marketing suggestion , the calculation formula is as follows: ; Where, is the feature number, and its value range is A positive integer used to traverse all matching features. is the number of matching features, , used to indicate the total number of features that match between user portrait data and marketing data, is the number of mismatched features, , used to indicate the total number of unmatched features, is the key feature number, , used to represent the total number of key features in marketing data, satisfying , For the The weight coefficient of the feature, , used to reflect the importance of features, is the Gaussian kernel parameter, , used to control the width of the Gaussian kernel function, The smaller the value, the sharper the kernel function. is the matching enhancement index, , used to control the strength of the logarithmic enhancement effect, For the The distance metric of each feature is calculated using the Euclidean distance, and the formula is: ; Where, is the feature vector dimension, User portrait data Feature No. The value of the dimension, For marketing data Feature No. The value of the dimension, , the smaller the distance, the higher the similarity; For the The penalty factor for mismatched features is calculated as follows: ; Where, is the penalty adjustment parameter, , The value range is [0, 1); when hour No punishment, when When it increases, Trend 1 punishment enhancement; In the GKEIMA algorithm, The range of is (0, 1], when When the function value is 1, it means a perfect match. When , the function value approaches 0, indicating a complete mismatch; Square root function Used to calculate the L2 norm of the weight vector and achieve vector normalization; Penalty In, when When the penalty term is 1, there is no effect. When it increases, the penalty effect increases; Logarithmic enhancement term middle, The ratio of response matching features to key features, index Achieve nonlinear enhancement effect.

[0016] Compared with the prior art, the beneficial effects of the present invention are: the present invention first obtains marketing data from multiple channels, and then pre-processes the marketing data, classifies the pre-processed marketing data according to the data type, and simultaneously obtains the marketing scenario requirements, obtains the corresponding analysis model according to the marketing scenario requirements, and generates a marketing decision scenario based on the analysis model through an intelligent decision-making model. In actual application, user behavior data is obtained, and user behavioral feedback on the marketing decision scenario is collected, and user behavior characteristics are merged into user portrait data and analyzed. If the analysis is correct, the feature information in the marketing decision scenario is extracted, and then a marketing effect evaluation index is constructed based on the extracted feature information. The marketing effect applicable to the marketing decision scenario is predicted from historical marketing data based on the marketing effect evaluation index. If there are one or more marketing cases applicable to the marketing decision scenario, they are merged and recorded as the first marketing data. Finally, personalized marketing suggestions are generated for the marketing decision scenario based on the first marketing data. The present invention combines the dynamically generated intelligent marketing decision scenario to realize artificial intelligence assistance for marketing decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 It is a flow chart of a marketing decision analysis method based on artificial intelligence of the present invention; Figure 2 This is a flow chart of classifying pre-processed marketing data according to data type in the present invention; Figure 3 It is a flowchart of generating a marketing decision scenario through an intelligent decision model based on an analysis model in the present invention; Figure 4 This is a flowchart of the present invention for constructing an intelligent decision-making model based on the analysis model corresponding to the marketing scenario requirements; Figure 5 This is a flow chart of combining user behavior features into user portrait data and performing analysis in the present invention; Figure 6 This is a flow chart of predicting marketing effects applicable to marketing decision-making scenarios from historical marketing data based on a marketing effect evaluation index in the present invention; Figure 7 This is a flowchart of the present invention for generating personalized marketing suggestions for marketing decision scenarios based on marketing data. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, a marketing decision analysis method based on artificial intelligence includes: Step S1: acquiring marketing data from multiple channels, preprocessing the marketing data, and then classifying the preprocessed marketing data according to data types.

[0020] like Figure 2 As shown, step S1 includes: Step S11, obtaining user data and transaction data from any marketing data.

[0021] Step S12: pre-process the user data in the marketing data to obtain a first data set.

[0022] In an embodiment, user data in marketing data is standardized, and then redundant information is removed to obtain first text information, which is then merged to obtain a first data set; based on data time attribute analysis, the first text information is classified according to data stability, and the classification criteria are as follows: long-cycle data (data change cycle is greater than 6 months) includes basic demographic characteristics, such as age, gender, educational background, occupation type, etc.; medium-cycle data (data change cycle is between 1-6 months) includes income level, residential address, interest preferences, etc.; short-cycle data (data change cycle is less than 1 month) includes recent browsing history, recent purchase behavior, current search keywords, etc.; a time decay weight matrix is ​​constructed, and corresponding time decay coefficients are set for data of different cycles: the decay control parameter of long-cycle data is set to 0.1-0.3, indicating lower time sensitivity; the decay control parameter of medium-cycle data is set to 0.4-0.6; the decay control parameter of short-cycle data is set to 0.7-1.0, indicating higher time sensitivity; and the first text information is reweighted based on the time decay weight matrix.

[0023] For example, the user data content in the marketing data is: "User Zhang, 28 years old, lives in Beijing, is a software engineer, has a monthly income of 15,000 yuan, prefers to buy electronic products, and recently browsed mobile phones, laptops, and other products"; the user data content is standardized and redundant information is removed to obtain the first text information, which is: "Zhang", "28 years old", "Beijing", "software engineer", "15,000", "electronic products", "mobile phone", "laptop computer", and the above information is merged to obtain the first data set; the information that has been classified and weighted by time period is merged to obtain the first data set; in this example, "28 years old", "Beijing", and "software engineer" are classified as long-cycle data, "15,000", and "electronic products" are classified as medium-cycle data, and "mobile phone" and "laptop computer" are classified as short-cycle data; the marketing data can be e-commerce platform data, social media data, or CRM system data.

[0024] Step S13: pre-process the transaction data in the marketing data to obtain a second data set.

[0025] In an embodiment, invalid transactions in the transaction data corresponding to the marketing data are determined to be redundant information, and then the redundant information in the transaction data is removed to obtain a second data set; wherein, since the transaction data usually contains a large number of identifiers and status codes automatically generated by the system, the invalid transaction records can be removed as redundant information; the invalid transactions in the marketing data can be canceled orders, refund records or system test data in the transaction data, etc.

[0026] Step S14: Operate each marketing data one by one according to steps S11 to S13 to obtain the first data set and the second data set corresponding to any marketing data, merge the first data set of all marketing data into user basic data, and merge the second data set of all marketing data into transaction behavior data.

[0027] Step S15: construct a corresponding feature word library according to the data type.

[0028] In an embodiment, the data types include user portrait data, product sales data, and marketing activity data, among which the feature words covered by the user portrait data include age, income, preference, region, occupation, and behavior, and the feature words covered by the above user portrait data are merged into a user portrait feature vocabulary, and then a product sales feature vocabulary corresponding to the data type of product sales data and a marketing activity feature vocabulary corresponding to the data type of marketing activity data are obtained.

[0029] Step S16: number the data types and count the frequency of occurrence of each feature word in the feature word library corresponding to any data type in the user basic data. and importance , calculate the intelligent feature scores of feature lexicons corresponding to different data types in user basic data according to the formula .

[0030] In the embodiment, the time-degradation intelligent feature scoring algorithm TDIFSA is used to calculate the intelligent feature score , the calculation formula is as follows: ; Where, is the feature word number, and its value range is , is the total number of feature words, indicating the size of the feature word library. Numbers are used to distinguish different types of marketing data. is the frequency of feature words, and its value range is , indicating the The feature word in Normalized frequency of occurrence in class data, is the feature word importance coefficient, and its value range is , reflecting the weight importance of feature words, is the trend value of the feature word in the time series, , the larger the value, the more obvious the upward trend. is the dispersion of feature words, , measures the degree of discreteness of the distribution of feature words in the data, is the time decay coefficient of the feature word, , used to calculate the time decay effect, time decay coefficient The calculation formula is: , where the data type standard period is set to 180 days, 90 days, and 30 days according to the long, medium, and short periods respectively; is the attenuation control parameter, , for long period data settings =0.1-0.3, for mid-cycle data setting =0.4-0.6, for short period data setting =0.7-1.0, controls the speed of aging decay; In the embodiment, for the dynamic adjustment of the above parameters, a parameter calibration feedback mechanism is constructed to optimize parameters using historical marketing effect data. When the prediction accuracy of the marketing effect evaluation index is lower than 80%, parameter recalibration is triggered. The calibration process uses a grid search method, traversing the parameter range with a step size of 0.1 to select the parameter combination with the highest prediction accuracy. The parameter adjustment cycle is set to automatically calibrate every 30 days, or manually trigger calibration when the data distribution changes significantly. In this embodiment, a hierarchical scoring strategy is adopted for feature words of different time periods, with the weight of long-term feature words accounting for 40%, the weight of medium-term feature words accounting for 35%, and the weight of short-term feature words accounting for 25%. A time window sliding mechanism is established, with a 12-month sliding window for long-term data, a 6-month sliding window for medium-term data, and a 1-month sliding window for short-term data. Data outside the time window will be gradually downgraded until it is removed. In the TDIFSA algorithm, the logarithmic function Used to realize nonlinear enhancement of trend value, square root function Used to provide normalization to prevent the value from being too large, exponential function Used to implement the time decay mechanism, the summation symbol Σ accumulates all feature words and finally divides them by Achieve averaging processing.

[0031] Assume that the user portrait data type number is 1, and its feature word library contains feature words such as "age", "income", and "preference". Count the frequency of the feature word "age" in the user basic data. , importance coefficient , time series trend value , dispersion , time-dependent attenuation coefficient , attenuation control parameter , then the contribution value of the feature word to the intelligent feature score is: .

[0032] In step S17, the calculated smart feature scores are sorted, and the data type corresponding to the highest smart feature score is used as the data type of the marketing data. Then, the smart feature scores of each marketing data are calculated one by one to obtain the data type corresponding to each marketing data, and the corresponding marketing data are classified and saved according to the data type.

[0033] In an embodiment, if the smart feature scores are the same, the smart feature scores of the feature lexicons corresponding to different data types in the transaction behavior data are calculated, and the marketing data is classified and stored according to the data type based on the smart feature scores.

[0034] Step S2: Obtain marketing scenario requirements, obtain corresponding analysis models based on the marketing scenario requirements, and generate marketing decision scenarios through intelligent decision models based on the analysis models.

[0035] like Figure 3 As shown, step S2 includes: In step S21, the marketing scenario requirements are segmented and input into a feature word library after segmentation, and then the intelligent feature score of the marketing scenario requirements is calculated to obtain an analysis model of the marketing scenario requirements.

[0036] Step S22: Build an intelligent decision-making model based on the analysis model corresponding to the marketing scenario requirements.

[0037] like Figure 4 As shown, step S22 includes: In step S221, based on the analysis model corresponding to the marketing scenario requirements, the corresponding data type is obtained from the classified marketing data, the marketing data of the corresponding data type is segmented, and then the words are classified to obtain user characteristics, product characteristics, channel characteristics and time characteristics, and merged into the feature data of the marketing data.

[0038] Step S222, constructing an intelligent decision-making model; the intelligent decision-making model is specifically: promoting (product characteristics) through (channel characteristics) in (time characteristics) for (user characteristics), and expected to achieve (effect target).

[0039] In the embodiment, the features in the intelligent decision-making model are specifically: time features are the time arrangements of marketing activities, user features are the attributes of the target user group, product features are the attribute information of the promoted products, channel features are the marketing channel selection, and effect targets are the expected marketing effects.

[0040] Step S223: randomly select characteristic data of the classified marketing data, and then add the characteristic data of the marketing data to the intelligent decision-making model to construct a marketing decision scenario.

[0041] In step S23, the constructed marketing decision scenario is compared with any marketing data. If the marketing decision scenario is the same as any marketing data, the corresponding marketing decision scenario is removed and the marketing decision scenario is reconstructed; if the marketing decision scenario is different from all marketing data, proceed to the next step.

[0042] Step S3: Obtain user behavior data, then collect user behavioral feedback on marketing decision scenarios, merge user behavior features into user portrait data and analyze them.

[0043] like Figure 5 As shown, step S3 includes: Step S31: construct a user behavior database, and obtain user behavior data based on the user behavior database.

[0044] In the embodiment, a user behavior database is first constructed, which contains behavior features corresponding to various user behavior types, and all behavior features constitute user behavior data.

[0045] Step S32: Compare any behavior feature in the user profile data with the user behavior data one by one.

[0046] Step S33: If the corresponding features in the user portrait data and the user behavior data are the same, proceed to the next step; if any feature in the user portrait data is different from the corresponding feature in the user behavior data, the user portrait data is determined to be abnormal.

[0047] Step S4: extract characteristic information from the marketing decision scenario, and then construct a marketing effect evaluation index based on the extracted characteristic information, and predict the marketing effect applicable to the marketing decision scenario from historical marketing data based on the marketing effect evaluation index.

[0048] like Figure 6 As shown, step S4 includes: Step S41: Acquire product features and user features in a marketing decision scenario, and extract first feature information of the product features and second feature information of the user features.

[0049] Step S42, by segmenting the product features and user features, and then labeling the segmented product features and user features, after labeling is completed, extracting the content of the specific tag to obtain the first feature information of the product features and the second feature information of the user features.

[0050] Step S43: traverse the marketing cases in the historical marketing data one by one and calculate the semantic similarity between the first feature information and the marketing cases. and association strength , get the dynamic evaluation value of the first feature information .

[0051] In this embodiment, the semantic similarity between the first feature information and the marketing case The cosine similarity algorithm is used for calculation, and the formula is: ; Where, is the number of product feature dimensions, The first feature information dimensional feature vector value, The first marketing case The formula measures the similarity by calculating the cosine of the angle between two eigenvectors. The range is [0,1]. The closer to 1, the higher the similarity. The strength of the correlation between the first characteristic information and the marketing case The Pearson correlation coefficient algorithm is used for calculation, and the formula is: Where, is the mean of the first feature information eigenvector, is the mean of the first marketing case feature vector, is the standard deviation of the first eigenvector, is the standard deviation of the first marketing case eigenvector. This formula measures the degree of linear correlation between two eigenvectors. Its value range is [-1,1]. The larger the absolute value, the higher the correlation strength. Dynamic evaluation value of the first characteristic information and marketing case The calculation formula is: ; Where, and are weight coefficients, and .

[0052] Step S44: Calculate the semantic similarity between the second feature information and the marketing case and association strength , get the dynamic evaluation value of the second feature information .

[0053] In the embodiment, the semantic similarity between the second feature information and the marketing case is calculated The cosine similarity algorithm is used for calculation, and the formula is: ; Where, is the number of user feature dimensions, The second feature information dimensional feature vector value, For the second marketing case The formula measures the similarity by calculating the cosine of the angle between two eigenvectors. The range is [0,1]. The closer to 1, the higher the similarity. The strength of the association between the second characteristic information and the marketing case The Pearson correlation coefficient algorithm is used for calculation, and the formula is: Where, is the mean of the second feature information eigenvector, is the mean of the feature vector of the second marketing case, is the standard deviation of the second eigenvector, is the standard deviation of the eigenvector of the second marketing case. This formula measures the degree of linear correlation between two eigenvectors. Its value range is [-1,1]. The larger the absolute value, the higher the correlation strength. Dynamic evaluation value The calculation formula is: .

[0054] Step S45: Calculate the marketing effect evaluation index using the multi-dimensional fusion intelligent evaluation algorithm MDFIEA .

[0055] In the embodiment, the marketing effect evaluation index is calculated by the multi-dimensional fusion intelligent evaluation algorithm MDFIEA , the calculation formula is as follows: ; Where, is the product feature weight coefficient, and the constraints are , used to balance the importance of product features and user features, is the similarity adjustment parameter, , used to control the saturation speed of the hyperbolic tangent function, The larger the value, the more sensitive the function. is the correlation strength adjustment parameter, , used to adjust the impact of association strength on the evaluation results, is the synergistic effect coefficient, , used to control the intensity of the synergy between product features and user features; Hyperbolic tangent function The range of is (−1,1), when hour, ,when hour, , used to implement nonlinear mapping; and Used to provide stable normalization effect; Quantitative assessment for achieving synergistic effects; Used to calculate the difference between two evaluation values; Used to take the maximum of the two as the normalization factor.

[0056] Assume that the dynamic evaluation value of the first feature information , the semantic similarity between the first feature information and the marketing case , the correlation strength between the first feature information and the marketing case , the dynamic evaluation value of the second feature information , the semantic similarity between the second feature information and the marketing case , the correlation strength between the second feature information and the marketing case , product feature weight coefficient , similarity adjustment parameter , association strength adjustment parameter , synergistic effect coefficient , then bring each parameter into the formula to calculate the marketing effect evaluation index .

[0057] Step S46: When the marketing effect evaluation index of any marketing case is greater than or equal to the evaluation index threshold, the marketing case is used as a reference case applicable to the marketing decision scenario, and one or more existing marketing cases applicable to the same marketing decision scenario are merged and recorded as the first marketing data.

[0058] Step S47: When the marketing effect evaluation index is less than the evaluation index threshold, no operation is performed.

[0059] Step S5: Generate personalized marketing suggestions for marketing decision scenarios based on marketing data.

[0060] like Figure 7 As shown, step S5 includes: Step S51: If all behavioral features in the user portrait data match the first marketing data, the optimal marketing recommendation is generated; if the number of features in the user portrait data is less than the number of features in the first marketing data, or any user portrait data does not match the first marketing data, or new users and new product scenarios are identified, proceed to the next step.

[0061] In the embodiment, it is necessary to determine whether the user is a new user; The criteria for determining whether a user is a new user are: the number of historical user behavior data records is less than a preset threshold (set to 3 records) or the user registration time is less than a preset time window (set to 7 days); When a new user is identified, a demographic-based initialization strategy is adopted, using historical data from user groups with similar demographic characteristics to infer a user profile. A new user feature vector is constructed, including basic attributes (age, gender, region, occupation) and inferred attributes (interest preferences, spending power). Inferred attributes are estimated using the average feature value of similar user groups. The confidence weight for new users is set to 0.3-0.5, and gradually increased to 1.0 as user behavior data accumulates. It is necessary to determine whether the product is a new product; The criteria for determining whether a product is new is: the product has been on the shelf for less than a preset time window (set to 30 days) or the product's historical sales records are less than a preset threshold (set to 10 records); When a new product is identified, a recommendation strategy based on product attribute similarity is adopted. The product feature vector (category, price, function, brand) is used to calculate similarity with historical products. A new product feature matrix is ​​constructed, using the average market performance data of products in the same category as the initial prediction value. The new product uncertainty adjustment factor is set to 0.2-0.4, and real market feedback data is collected through A / B testing and progressive promotion strategies. For the new user-new product combination scenario, a hybrid strategy combining content-based recommendation and collaborative filtering is adopted; preliminary matching is performed using product content features and user basic features; the matching weight is set to 0.6; collaborative filtering is supplemented using similar users' preference data for similar products, and the collaborative filtering weight is set to 0.4; a conservative recommendation strategy is set to prioritize product categories with high market verification before proceeding to subsequent steps.

[0062] Step S52: Use Gaussian kernel enhanced intelligent matching algorithm GKEIMA to calculate the intelligent matching degree of marketing suggestions .

[0063] In the embodiment, the Gaussian kernel enhanced intelligent matching algorithm GKEIMA is used to calculate the intelligent matching degree of marketing suggestions , the calculation formula is as follows: ; Where, is the feature number, and its value range is A positive integer used to traverse all matching features. is the number of matching features, , used to indicate the total number of features that match between user portrait data and marketing data, is the number of mismatched features, , used to indicate the total number of unmatched features, is the key feature number, , used to represent the total number of key features in marketing data, satisfying , For the The weight coefficient of the feature, , used to reflect the importance of features, is the Gaussian kernel parameter, , used to control the width of the Gaussian kernel function, The smaller the value, the sharper the kernel function. is the matching enhancement index, , used to control the strength of the logarithmic enhancement effect, For the The distance metric of each feature is calculated using the Euclidean distance, and the formula is: ; Where, is the feature vector dimension, User portrait data Feature No. The value of the dimension, For marketing data Feature No. The value of the dimension, , the smaller the distance, the higher the similarity; For the The penalty factor for mismatched features is calculated as follows: ; Where, is the penalty adjustment parameter, , The value range is [0, 1); when hour No punishment, when When it increases, Trend 1 punishment enhancement; In the GKEIMA algorithm, The range of is (0, 1], when When the function value is 1, it means a perfect match. When , the function value approaches 0, indicating a complete mismatch; Square root function Used to calculate the L2 norm of the weight vector and achieve vector normalization; Penalty In, when When the penalty term is 1, there is no effect. When it increases, the penalty effect increases; Logarithmic enhancement term middle, The ratio of response matching features to key features, index Achieve nonlinear enhancement effect.

[0064] Assuming the number of matching features , mismatched feature count , key feature number , feature weight vector , Gaussian kernel parameters , penalty factor , matching enhancement index , distance metric Based on the feature similarity calculation, each parameter is brought into the formula to calculate the intelligent matching degree .

[0065] Step S53: When the smart matching degree is less than or equal to the matching degree threshold, a basic marketing suggestion is generated; when the smart matching degree is greater than the matching degree threshold, a personalized marketing suggestion is generated.

[0066] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features therein. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A marketing decision analysis method based on artificial intelligence, characterized in that: include: Step S1: acquiring marketing data from multiple channels, preprocessing the marketing data, and then classifying the preprocessed marketing data according to data type; Step S2: Obtain marketing scenario requirements, obtain corresponding analysis models based on the marketing scenario requirements, and generate marketing decision scenarios through an intelligent decision model based on the analysis models; Step S3: Obtain user behavior data, then collect user behavioral feedback on marketing decision scenarios, merge user behavior features into user portrait data and analyze them; Step S4: extracting characteristic information from the marketing decision scenario, then constructing a marketing effect evaluation index based on the extracted characteristic information, and predicting the marketing effect applicable to the marketing decision scenario from historical marketing data based on the marketing effect evaluation index; Step S5: Generate personalized marketing suggestions for marketing decision scenarios based on marketing data.

2. The marketing decision analysis method based on artificial intelligence according to claim 1, characterized in that: The step S1 comprises: Step S11, obtaining user data and transaction data from any marketing data; Step S12: pre-processing the user data in the marketing data to obtain a first data set; Step S13, pre-processing the transaction data in the marketing data to obtain a second data set; Step S14: Process each marketing data set according to steps S11 to S13 to obtain a first data set and a second data set corresponding to each marketing data set, merge the first data sets of all marketing data sets into user basic data, and merge the second data sets of all marketing data sets into transaction behavior data; Step S15, constructing a corresponding feature word library according to the data type; Step S16: number the data types and count the frequency of occurrence of each feature word in the feature word library corresponding to any data type in the user basic data. and importance , calculate the intelligent feature scores of feature lexicons corresponding to different data types in user basic data according to the formula ; In step S17, the calculated smart feature scores are sorted, and the data type corresponding to the highest smart feature score is used as the data type of the marketing data. Then, the smart feature scores of each marketing data are calculated one by one to obtain the data type corresponding to each marketing data, and the corresponding marketing data are classified and saved according to the data type.

3. The marketing decision analysis method based on artificial intelligence according to claim 2, characterized in that: In step S12, the user data in the marketing data is standardized, and then redundant information is removed to obtain first text information, which is then merged to obtain a first data set. Based on the data time attribute analysis, the first text information is classified according to data stability. The classification standard is as follows: data with a change cycle greater than 6 months is long-cycle data, data with a change cycle between 1-6 months is medium-cycle data, and data with a change cycle less than 1 month is short-cycle data; In step S13, invalid transactions in the transaction data corresponding to the marketing data are determined to be redundant information, and then the redundant information in the transaction data is removed to obtain a second data set.

4. The marketing decision analysis method based on artificial intelligence according to claim 2, characterized in that: In step S16, the time-degradation intelligent feature scoring algorithm TDIFSA is used to calculate the intelligent feature score. , the calculation formula is as follows: ； Where, is the feature word number, and its value range is , is the total number of feature words, indicating the size of the feature word library. Numbers are used to distinguish different types of marketing data. is the frequency of feature words, and its value range is , indicating the The feature word in Normalized frequency of occurrence in class data, is the feature word importance coefficient, and its value range is , reflecting the weight importance of feature words, is the trend value of the feature word in the time series, , the larger the value, the more obvious the upward trend. is the dispersion of feature words, , measures the degree of discreteness of the distribution of feature words in the data, is the time decay coefficient of the feature word, , used to calculate the time decay effect, is the attenuation control parameter, , controlling the speed of aging decay; To dynamically adjust the above parameters: Build a parameter calibration feedback mechanism, optimize parameters using historical marketing effectiveness data, and trigger parameter recalibration when the prediction accuracy of the marketing effectiveness evaluation index falls below 80%. The calibration process uses a grid search method, traversing the parameter range in steps of 0.1 to select the parameter combination with the highest prediction accuracy. The parameter adjustment cycle is set to automatically calibrate every 30 days, or manually trigger calibration when the data distribution changes significantly. A hierarchical scoring strategy is adopted for feature words of different time periods, with long-term feature words accounting for 40% of the weight, medium-term feature words accounting for 35% of the weight, and short-term feature words accounting for 25% of the weight. A time window sliding mechanism is established, with a 12-month sliding window for long-term data, a 6-month sliding window for medium-term data, and a 1-month sliding window for short-term data. Data outside the time window will be gradually downgraded until it is removed. In the TDIFSA algorithm, the logarithmic function Used to realize nonlinear enhancement of trend value, square root function Used to provide normalization to prevent the value from being too large, exponential function Used to implement the time decay mechanism, the summation symbol Σ accumulates all feature words and finally divides them by Achieve averaging processing.

5. The marketing decision analysis method based on artificial intelligence according to claim 2, characterized in that: In step S17, if the smart feature scores are the same, the smart feature scores of the feature lexicons corresponding to different data types in the transaction behavior data are calculated, and the marketing data are classified and stored according to the data type based on the smart feature scores.

6. The marketing decision analysis method based on artificial intelligence according to claim 1, characterized in that: The step S2 comprises: Step S21: Segment the marketing scenario requirements into words, input the segmented words into a feature word library, and then calculate the intelligent feature score of the marketing scenario requirements to obtain an analysis model of the marketing scenario requirements; Step S22: Building an intelligent decision-making model based on the analysis model corresponding to the marketing scenario requirements; Step S23: Compare the constructed marketing decision scenario with any marketing data. If the marketing decision scenario is identical to any marketing data, remove the corresponding marketing decision scenario and reconstruct the marketing decision scenario. If the marketing decision scenario is different from any marketing data, proceed to the next step. The step S22 includes: Step S221: Based on the analysis model corresponding to the marketing scenario requirements, the corresponding data type is obtained from the classified marketing data, the marketing data of the corresponding data type is segmented, and then the words are classified to obtain user characteristics, product characteristics, channel characteristics, and time characteristics, which are combined into feature data of the marketing data; Step S222, constructing an intelligent decision-making model; Step S223: randomly select characteristic data of the classified marketing data, and then add the characteristic data of the marketing data to the intelligent decision-making model to construct a marketing decision scenario.

7. The marketing decision analysis method based on artificial intelligence according to claim 1, characterized in that: The step S3 comprises: Step S31: construct a user behavior database and obtain user behavior data based on the user behavior database; Step S32, comparing any behavioral feature in the user profile data with the user behavior data one by one; Step S33: If the corresponding features in the user portrait data and the user behavior data are the same, proceed to the next step; if any feature in the user portrait data is different from the corresponding feature in the user behavior data, the user portrait data is determined to be abnormal.

8. The marketing decision analysis method based on artificial intelligence according to claim 1, characterized in that: The step S4 comprises: Step S41, obtaining product features and user features in the marketing decision scenario, and extracting first feature information of the product features and second feature information of the user features; Step S42: Segmenting the product features and user features, then labeling the segmented product features and user features, and extracting the content of specific tags to obtain first feature information of the product features and second feature information of the user features; Step S43: traverse the marketing cases in the historical marketing data one by one and calculate the semantic similarity between the first feature information and the marketing cases. and association strength , get the dynamic evaluation value of the first feature information ; Step S44: Calculate the semantic similarity between the second feature information and the marketing case and association strength , get the dynamic evaluation value of the second feature information ; Step S45: Calculate the marketing effect evaluation index using the multi-dimensional fusion intelligent evaluation algorithm MDFIEA ; Step S46: When the marketing effect evaluation index of any marketing case is greater than or equal to the evaluation index threshold, the marketing case is used as a reference case applicable to the marketing decision scenario, and one or more existing marketing cases applicable to the same marketing decision scenario are combined and recorded as the first marketing data; Step S47: When the marketing effect evaluation index is less than the evaluation index threshold, no operation is performed.

9. The marketing decision analysis method based on artificial intelligence according to claim 8, characterized in that: In step S43, the semantic similarity between the first feature information and the marketing case The cosine similarity algorithm is used for calculation, and the formula is: ； Where, is the number of product feature dimensions, The first feature information dimensional feature vector value, The first marketing case The formula measures the similarity by calculating the cosine of the angle between two eigenvectors. The range is [0,1]. The closer to 1, the higher the similarity. The strength of the correlation between the first characteristic information and the marketing case The Pearson correlation coefficient algorithm is used for calculation, and the formula is: ； ； ； ； ； Where, is the mean of the first feature information eigenvector, is the mean of the first marketing case feature vector, is the standard deviation of the first eigenvector, is the standard deviation of the first marketing case eigenvector. This formula measures the degree of linear correlation between two eigenvectors. Its value range is [-1,1]. The larger the absolute value, the higher the correlation strength. Dynamic evaluation value of the first characteristic information and marketing case The calculation formula is: ； Where, and are weight coefficients, and ; In step S44, the semantic similarity between the second feature information and the marketing case is calculated. The cosine similarity algorithm is used for calculation, and the formula is: ； Where, is the number of user feature dimensions, The second feature information dimensional feature vector value, For the second marketing case The formula measures the similarity by calculating the cosine of the angle between two eigenvectors. The range is [0,1]. The closer to 1, the higher the similarity. The strength of the association between the second characteristic information and the marketing case The Pearson correlation coefficient algorithm is used for calculation, and the formula is: Where, is the mean of the second feature information eigenvector, is the mean of the feature vector of the second marketing case, is the standard deviation of the second eigenvector, is the standard deviation of the eigenvector of the second marketing case. This formula measures the degree of linear correlation between two eigenvectors. Its value range is [-1,1]. The larger the absolute value, the higher the correlation strength. Dynamic evaluation value The calculation formula is: ； In step S45, the marketing effect evaluation index is calculated by the multi-dimensional fusion intelligent evaluation algorithm MDFIEA. , the calculation formula is as follows: ； Where, is the product feature weight coefficient, and the constraints are , used to balance the importance of product features and user features, is the similarity adjustment parameter, , used to control the saturation speed of the hyperbolic tangent function, The larger the value, the more sensitive the function. is the correlation strength adjustment parameter, , used to adjust the impact of association strength on the evaluation results, is the synergistic effect coefficient, , used to control the intensity of the synergy between product features and user features; Hyperbolic tangent function The range of is (−1,1), when hour, ,when hour, , used to implement nonlinear mapping; and Used to provide stable normalization effect; Quantitative assessment for achieving synergistic effects; Used to calculate the difference between two evaluation values; Used to take the maximum of the two as the normalization factor.

10. The marketing decision analysis method based on artificial intelligence according to claim 1, characterized in that: The step S5 comprises: Step S51: If all behavioral features in the user profile data match the first marketing data, then the optimal marketing recommendation is generated; if the number of features in the user profile data is less than the number of features in the first marketing data, or any user profile data does not match the first marketing data, or a new user and new product market are identified, then proceed to the next step; Step S52: Use Gaussian kernel enhanced intelligent matching algorithm GKEIMA to calculate the intelligent matching degree of marketing suggestions ; Step S53: When the smart matching degree is less than or equal to the matching degree threshold, a basic marketing suggestion is generated; when the smart matching degree is greater than the matching degree threshold, a personalized marketing suggestion is generated; In step S51, whether the user is a new user is determined by: the number of historical behavior data records of the user is less than a preset threshold or the user registration time is less than a preset time window; When a new user is identified, a demographic-based initialization strategy is adopted, using historical data from user groups with similar demographic characteristics to infer the user profile. A new user feature vector is constructed, including basic attributes and inferred attributes. The inferred attributes are estimated using the average feature value of similar user groups. The criteria for determining whether a product is new is: the product's shelf life is less than the preset time window or the product's historical sales record is less than the preset threshold; When a new product is identified, a recommendation strategy based on product attribute similarity is adopted, calculating the similarity between the product feature vector and historical products. A new product feature matrix is ​​constructed, using the average market performance data of products in the same category as the initial prediction value. An uncertainty adjustment factor is set for the new product, and real market feedback data is collected through A / B testing and progressive promotion strategies. For the new user-new product combination scenario, a hybrid strategy combining content-based recommendations and collaborative filtering is adopted. Initial matching is performed using product content features and user basic features. Collaborative filtering is supplemented by using similar users' preference data for similar products. A conservative recommendation strategy is set, prioritizing product categories with high market validation before proceeding to subsequent steps. In step S52, the Gaussian kernel enhanced intelligent matching algorithm GKEIMA is used to calculate the intelligent matching degree of the marketing suggestion , the calculation formula is as follows: ； Where, is the feature number, and its value range is A positive integer used to traverse all matching features. is the number of matching features, , used to indicate the total number of features that match between user portrait data and marketing data, is the number of mismatched features, , used to indicate the total number of unmatched features, is the key feature number, , used to represent the total number of key features in marketing data, satisfying , For the The weight coefficient of the feature, , used to reflect the importance of features, is the Gaussian kernel parameter, , used to control the width of the Gaussian kernel function, The smaller the value, the sharper the kernel function. is the matching enhancement index, , used to control the strength of the logarithmic enhancement effect, For the The distance metric of each feature is calculated using the Euclidean distance, and the formula is: ； Where, is the feature vector dimension, User portrait data Feature No. The value of the dimension, For marketing data Feature No. The value of the dimension, , the smaller the distance, the higher the similarity; For the The penalty factor for mismatched features is calculated as follows: ； Where, is the penalty adjustment parameter, , The value range is [0, 1); when hour No punishment, when When it increases, Trend 1 punishment enhancement; In the GKEIMA algorithm, The range of is (0, 1], when When the function value is 1, it means a perfect match. When , the function value approaches 0, indicating a complete mismatch; Square root function Used to calculate the L2 norm of the weight vector and achieve vector normalization; Penalty In, when When the penalty term is 1, it has no effect. When it increases, the penalty effect increases; Logarithmic enhancement term middle, The ratio of response matching features to key features, index Achieve nonlinear enhancement effect.

Images