Intelligent marketing decision-making method and system based on big data

Abstract

The present application relates to the technical field of marketing decision, in particular to an intelligent marketing decision method and system based on big data, comprising: obtaining a plurality of historical marketing activities of a target user group, and obtaining user behavior characteristics, positive feedback keywords and market trend correlation data corresponding to a high conversion period of each historical marketing activity; determining a decision sample set based on a user behavior data set with conversion records; the conversion record is used to record conversion data of a user's purchase behavior to a historical marketing activity. The present application defines a negative sample, that is, it is required that a non-converted user must have a purchase behavior to other historical marketing activities of the same type, which avoids mixing completely silent or irrelevant users into the training set, so that the model learns the user behavior pattern of the user who has the ability to purchase but does not choose the current activity, thereby enabling the model to more accurately identify the key factors that truly affect conversion and reduce misjudgment.

Description

Technical Field

[0001] This invention relates to the field of marketing decision-making technology, specifically to an intelligent marketing decision-making method and system based on big data. Background Technology

[0002] Currently, in terms of sample construction, traditional methods often simply treat all users who browsed the activity but did not purchase as negative samples, including a large number of silent users who are completely uninterested or never shop. This makes it difficult for the model to distinguish between people who have the purchasing power but have not been converted and ineffective people who do not pay attention to the category. The model is prone to learning incorrect patterns, such as misjudging user inactivity as disinterest in a specific activity, thereby reducing the accuracy of identifying truly potential conversion users. Moreover, traditional methods often simply judge whether the user's historical purchase category is consistent with the current activity category, obtaining a binary match feature. This approach loses a lot of information, such as the degree of overlap of user interests in multiple related categories and the strength of differences in interests. As a result, the model can only obtain a coarse matching signal and cannot accurately perceive the matching depth between users and activities, thus limiting its predictive ability.

[0003] Furthermore, in terms of feature extraction, traditional methods often rely on human experience to define user preferences and marketing campaign attributes. For example, operations staff manually set which behaviors represent high intent and which keywords represent positive feedback. This approach is highly subjective, with inconsistent standards defined by different personnel, and it is difficult to cover complex and ever-changing user behavior patterns and market trends. At the same time, traditional methods rarely incorporate budget factors and brand tone matching into the feature selection process, which may lead to the extracted features being out of touch with actual business objectives and failing to reflect the constraints of cost and brand style on the campaign. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a big data-based intelligent marketing decision-making method, comprising:

[0005] Obtain multiple historical marketing campaigns for the target user group, and acquire user behavior characteristics, positive feedback keywords, and market trend correlation data for the high-conversion periods corresponding to each historical marketing campaign;

[0006] Based on a set of user behavior data with conversion records, a decision sample set is determined; conversion records are used to record conversion data of users' purchase behavior in response to historical marketing campaigns.

[0007] Determine the user preference characteristics, marketing campaign attribute characteristics, user historical behavior characteristics, marketing campaign conversion characteristics, and user-marketing campaign interaction characteristics corresponding to the decision sample;

[0008] The user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and interaction features corresponding to the decision sample are input into the initial intelligent marketing decision model for prediction processing to obtain the prediction score corresponding to each decision sample; the prediction score is used to characterize the user's conversion tendency for historical marketing activities.

[0009] Based on the predicted scores corresponding to the decision samples and the positive and negative sample labels of the decision samples, the parameters of the intelligent marketing decision model are adjusted until the convergence condition is met, and the trained intelligent marketing decision model is obtained. The current candidate marketing activities are obtained, and the trained intelligent marketing decision model is used to score the candidate marketing activities. Based on the scores, the intelligent marketing decision scheme for the target user group is determined.

[0010] Preferably, a decision sample set is determined based on a set of user behavior data with conversion records, including:

[0011] Based on the conversion logs of the marketing platform, determine the set of historical marketing campaigns with conversion records within a preset time period, and the positive samples corresponding to each historical marketing campaign in the set of historical marketing campaigns; wherein, the conversion logs include all conversion records in the marketing platform, and the number of positive samples corresponding to historical marketing campaigns is equal to the number of users who have converted from historical marketing campaigns;

[0012] Based on the browsing logs, conversion logs, and preset sample sampling ratio of the marketing platform, the negative samples corresponding to each historical marketing activity in the historical marketing activity set are determined. The browsing logs include browsing records of each marketing activity in the marketing platform, and the sample sampling ratio is the ratio between positive and negative samples set for the marketing activity category. The number of negative samples corresponding to the historical marketing activity is determined based on the browsing volume, conversion volume, and sample sampling ratio of the historical marketing activity.

[0013] The decision sample set includes multiple decision samples corresponding to each historical marketing campaign. Among the multiple decision samples, the positive samples are the matching pairs between the historical marketing campaign and the corresponding converted users, and the negative samples are the matching pairs between the historical marketing campaign and the non-converted users. Converted users are users who have converted behavior in the historical marketing campaign, and non-converted users are users who have not converted behavior in the historical marketing campaign but have converted behavior in other historical marketing campaigns of the same category.

[0014] Preferably, based on the marketing platform's browsing logs, conversion logs, and a preset sample sampling ratio, negative samples are determined for each historical marketing campaign in the historical marketing campaign set, including:

[0015] For any historical marketing campaign in the collection of historical marketing campaigns, obtain the page views of that historical marketing campaign and the total page views of all historical marketing campaigns in the category to which that historical marketing campaign belongs from the browsing log;

[0016] Based on conversion records, obtain the conversion volume of the historical marketing campaign and the total conversion volume of all historical marketing campaigns in the category to which the historical marketing campaign belongs; the conversion volume is used to represent the number of users who have converted for the historical marketing campaign.

[0017] Calculate the negative sample coefficient corresponding to the historical marketing campaign based on the sample sampling ratio, the total page views of all historical marketing campaigns, and the total conversion of all historical marketing campaigns;

[0018] Multiply the difference between the page views and conversions of the historical marketing campaign by the negative sample coefficient to obtain the number of negative samples corresponding to the historical marketing campaign.

[0019] Based on the number of negative samples corresponding to the historical marketing campaign, a corresponding number of non-converted users are selected from the negative sample candidate set of the category to which the historical marketing campaign belongs, forming the negative samples corresponding to the historical marketing campaign; wherein, the negative sample candidate set is the set of users who have converted for any historical marketing campaign in the same category.

[0020] Preferably, the user preference characteristics and marketing activity attribute characteristics corresponding to the decision sample are determined, including:

[0021] Obtain user behavior characteristics, positive feedback keywords, and market trend correlation data during high-conversion periods of historical marketing campaigns;

[0022] Extract user behavior pattern features based on user behavior characteristics during high-conversion periods;

[0023] Based on the timestamps, positive feedback keywords, and market trend correlation data corresponding to user behavior characteristics during high-conversion periods, the correlation characteristics are determined;

[0024] Based on user behavior patterns and relationship characteristics, multiple historical marketing campaigns are clustered to obtain multiple marketing campaign categories.

[0025] Identify the target marketing campaign category from multiple marketing campaign categories;

[0026] User behavior characteristics during high-conversion periods of historical marketing campaigns included in the target marketing campaign category are used as user preference characteristics, and positive feedback keywords and market trend correlation data of historical marketing campaigns included in the target marketing campaign category are used as marketing campaign attribute characteristics.

[0027] Preferably, the target marketing campaign category is determined from multiple marketing campaign categories, including:

[0028] Obtain user feedback density, positive feedback sentiment index, and conversion growth rate index for historical marketing campaigns within a preset time period for each marketing campaign category;

[0029] Based on user feedback density, positive feedback sentiment index, and conversion growth rate index, determine the marketing performance evaluation value corresponding to each marketing campaign category;

[0030] The marketing campaign category corresponding to the highest marketing performance evaluation value will be used as the target marketing campaign category.

[0031] Preferably, determining the target marketing campaign category from multiple marketing campaign categories also includes:

[0032] Obtain the target budget factor corresponding to the target user group, and obtain the budget factor corresponding to the historical marketing activities in each marketing activity category; where the budget factor represents the budget amount of the marketing activity;

[0033] Determine the difference between the budget factor corresponding to the historical marketing campaign and the target budget factor, and use this as the budget factor difference value;

[0034] For each marketing campaign category, calculate the sum of the marketing performance evaluation value corresponding to that category and the budget factor difference value corresponding to that category, and use it as the first updated evaluation value for that category;

[0035] The marketing campaign category corresponding to the highest first update rating value will be used as the target marketing campaign category.

[0036] Preferably, determining the target marketing campaign category from multiple marketing campaign categories also includes:

[0037] Obtain the target brand tone characteristic value corresponding to the target user group, and obtain the brand tone characteristic value corresponding to the historical marketing activities in each marketing activity category; among them, the brand tone characteristic value represents the youthfulness and high-end characteristics of the marketing activities;

[0038] The difference between the brand tone characteristic value corresponding to the historical marketing campaign and the target brand tone characteristic value is determined as the brand tone difference value;

[0039] For each marketing campaign category, calculate the sum of the marketing performance evaluation value, budget factor difference value, and brand tone difference value corresponding to that category, and use it as the second updated evaluation value for that category;

[0040] The marketing campaign category corresponding to the second highest updated evaluation value will be used as the target marketing campaign category.

[0041] Preferably, the interaction characteristics between the users corresponding to the decision sample and the marketing campaign are determined, including:

[0042] Based on user information and marketing activity information in the decision sample, determine the user's corresponding preference tags and the marketing activity's corresponding attribute tags;

[0043] When preference tags and attribute tags are the same, the category crossover feature between users and marketing campaigns is identified as the first category crossover feature.

[0044] When preference tags and attribute tags are different, the category crossover feature between users and marketing campaigns is identified as the second category crossover feature.

[0045] The number of matching tags between multiple preference tags for users and multiple attribute tags for marketing campaigns is determined as a numerical cross feature.

[0046] Preferably, based on the predicted scores corresponding to the decision samples and the positive and negative sample labels of the decision samples, the parameters of the intelligent marketing decision model are adjusted until the convergence condition is met, resulting in a trained intelligent marketing decision model, including:

[0047] Discrete features among user preference features, marketing campaign attribute features, user historical behavior features, marketing campaign conversion features, and user-marketing campaign interaction features are embedded to obtain dense feature vectors, and continuous features are normalized to obtain numerical feature vectors.

[0048] Dense feature vectors and numerical feature vectors are concatenated to generate a unified feature representation vector;

[0049] The unified feature representation vector is input into the factorization machine layer to learn the second-order interaction relationship between features and generate the first interaction feature vector.

[0050] A unified feature representation vector is simultaneously input into a multi-head self-attention layer to capture the deep interaction relationship between user history behavior and marketing campaign attributes, generating a second interaction feature vector.

[0051] The unified feature representation vector, the first interactive feature vector, and the second interactive feature vector are concatenated to obtain the fused feature vector;

[0052] The fused feature vector is input into a deep neural network consisting of multiple fully connected layers, and after activation function and regularization processing, an abstract feature vector is obtained.

[0053] The abstract feature vector is input into the output layer, processed by a fully connected layer and a mapping function to obtain the prediction score corresponding to the decision sample;

[0054] Calculate the binary cross-entropy loss using the predicted score and the positive and negative sample labels of the decision sample;

[0055] Based on the binary cross-entropy loss, the parameters of the factorization machine layer, multi-head self-attention layer, deep neural network, and output layer are updated through gradient backpropagation;

[0056] Repeat the parameter update process until the convergence condition is met, and the trained intelligent marketing decision model is obtained.

[0057] A big data-based intelligent marketing decision-making system, applicable to the aforementioned big data-based intelligent marketing decision-making methods, including:

[0058] The data acquisition unit is configured to acquire multiple historical marketing campaigns for the target user group, and to acquire user behavior characteristics, positive feedback keywords, and market trend correlation data during the high-conversion periods of each historical marketing campaign.

[0059] The sample construction unit is configured to determine the decision sample set based on a set of user behavior data with conversion records; the conversion records are used to record conversion data of users' purchase behavior in historical marketing activities;

[0060] The feature extraction unit is configured to determine the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and user-marketing activity interaction features corresponding to the decision sample;

[0061] The marketing prediction unit is configured to input the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and interaction features corresponding to the decision sample into the initial intelligent marketing decision model for prediction processing, and obtain the prediction score corresponding to each decision sample; the prediction score is used to characterize the user's conversion tendency for historical marketing activities;

[0062] The marketing decision-making unit is configured to adjust the parameters of the intelligent marketing decision-making model based on the predicted score corresponding to the decision sample and the positive and negative sample labels of the decision sample until the convergence condition is met, thereby obtaining the trained intelligent marketing decision-making model; acquire the current candidate marketing activities, use the trained intelligent marketing decision-making model to score the candidate marketing activities, and determine the intelligent marketing decision-making plan for the target user group based on the score.

[0063] Compared with the prior art, the beneficial effects of the present invention are:

[0064] (1) By defining negative samples, this invention requires that users who have not converted must have made purchases in other historical marketing activities of the same category, thus avoiding mixing completely silent or irrelevant users into the training set. This allows the model to learn the behavior patterns of users who have the ability to purchase but have not chosen the current activity, thereby enabling the model to more accurately identify the key factors that truly affect conversion and reduce misjudgments. Moreover, by clustering behavioral characteristics, positive feedback keywords and market trend-related data during high-conversion periods, and by comprehensively considering multiple criteria such as marketing performance evaluation value, budget factor matching degree and brand tone matching degree to screen target marketing activity categories, the system can automatically summarize the most representative and best-performing activity patterns. This avoids the subjectivity and limitations of manually defined features, making feature extraction more objective, comprehensive and aligned with business objectives.

[0065] (2) This invention distinguishes whether user preference tags and activity attribute tags are the same to obtain two different categories of cross features. At the same time, it also counts the number of matches between multiple tags as numerical cross features. This design enables the model to capture both qualitative information on whether categories are aligned and quantitative information on the degree of tag overlap, thereby enhancing the model's ability to understand the matching degree between users and activities. Moreover, it can automatically determine which users are suitable to push the current candidate marketing activities to, at what time to push them, and what strategies to use, thereby reducing the cost of manual trial and error, shortening the marketing decision-making cycle, and improving the marketing input-output ratio. Attached Figure Description

[0066] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0068] In the diagram: 1. Data acquisition unit; 2. Sample construction unit; 3. Feature extraction unit; 4. Marketing prediction unit; 5. Marketing decision-making unit. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1, please refer to Figure 1 This invention provides a technical solution: an intelligent marketing decision-making method based on big data, comprising:

[0071] S1. Obtain multiple historical marketing campaigns for the target user group, and obtain user behavior characteristics, positive feedback keywords, and market trend correlation data for the high-conversion periods corresponding to each historical marketing campaign;

[0072] S2. Determine the decision sample set based on the user behavior data set with conversion records; conversion records are used to record conversion data of users' purchase behavior in historical marketing activities;

[0073] S3. Determine the user preference characteristics, marketing activity attribute characteristics, user historical behavior characteristics, marketing activity conversion characteristics, and user-marketing activity interaction characteristics corresponding to the decision sample;

[0074] S4. Input the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and interaction features corresponding to the decision sample into the initial intelligent marketing decision model for prediction processing to obtain the prediction score corresponding to each decision sample; the prediction score is used to characterize the user's conversion tendency for historical marketing activities.

[0075] S5. Based on the predicted scores corresponding to the decision samples and the positive and negative sample labels of the decision samples, adjust the parameters of the intelligent marketing decision model until the convergence condition is met, and obtain the trained intelligent marketing decision model; obtain the current candidate marketing activities, use the trained intelligent marketing decision model to score the candidate marketing activities, and determine the intelligent marketing decision scheme for the target user group based on the scores.

[0076] It should be noted that the system obtains information on multiple marketing campaigns that the target user group has participated in throughout history. For each historical marketing campaign, the system also needs to obtain three additional types of key information: first, the specific behavioral characteristics of users during the period with the highest conversion rate in that campaign; second, the keywords in the positive feedback given by users during that period; and third, the market trend correlation data corresponding to that period. This information provides raw materials for subsequent feature extraction and model training.

[0077] The decision sample set is constructed based on the data set that actually records user purchase behavior; the conversion record here clearly indicates that the user completed a purchase for a certain historical marketing campaign; the system filters positive samples through conversion logs, that is, the matching pairs of users who actually made a purchase and the campaign; at the same time, combined with browsing logs and a preset sample sampling ratio, negative samples are constructed, that is, the matching pairs of users who browsed the campaign but did not make a purchase, but had made purchase behavior for other campaigns in the same category and the campaign; the positive and negative samples together form the decision sample set;

[0078] Five types of features were identified for each decision sample: user preference features were derived from user behavior patterns summarized from similar high-conversion activities; marketing activity attribute features were derived from positive feedback keywords and market trend correlation data from similar activities; user historical behavior features recorded the frequency of users' past participation in activities, click habits, etc.; marketing activity conversion features reflected the overall conversion indicators of the activity, such as click-through rate and add-to-cart rate; and user-marketing activity interaction features included the category matching indicator and the number of multi-tag matching. These five types of features comprehensively depicted the multi-dimensional information of the sample.

[0079] The five categories of features corresponding to each decision sample are input into an untrained initial intelligent marketing decision model. The model performs a series of processing on these features, including embedding discrete features, normalizing continuous features, learning second-order interactions using factorization machines, learning deep interactions using multi-head self-attention, and extracting abstract representations through deep neural networks. Finally, the model outputs a prediction score for each sample, which represents the degree to which the model judges that the user is likely to convert to the historical marketing campaign.

[0080] The predicted score of each sample is compared with its true positive or negative sample label, the loss is calculated, and the model parameters are continuously adjusted through gradient backpropagation. This process is repeated until the model's performance reaches the convergence condition, at which point a trained intelligent marketing decision model is obtained. Then, when faced with a new current candidate marketing campaign, the system uses this trained model to score the campaign or the combination of users and campaigns. Based on the scoring results, the system finally determines what kind of intelligent marketing decision-making scheme should be adopted for the target user group, such as which users to push to, when to push, and what kind of discount to use.

[0081] In an optional embodiment, determining a decision sample set based on a set of user behavior data with conversion records includes:

[0082] Based on the conversion logs of the marketing platform, determine the set of historical marketing campaigns with conversion records within a preset time period, and the positive samples corresponding to each historical marketing campaign in the set of historical marketing campaigns; wherein, the conversion logs include all conversion records in the marketing platform, and the number of positive samples corresponding to historical marketing campaigns is equal to the number of users who have converted from historical marketing campaigns;

[0083] Based on the browsing logs, conversion logs, and preset sample sampling ratio of the marketing platform, the negative samples corresponding to each historical marketing activity in the historical marketing activity set are determined. The browsing logs include browsing records of each marketing activity in the marketing platform, and the sample sampling ratio is the ratio between positive and negative samples set for the marketing activity category. The number of negative samples corresponding to the historical marketing activity is determined based on the browsing volume, conversion volume, and sample sampling ratio of the historical marketing activity.

[0084] The decision sample set includes multiple decision samples corresponding to each historical marketing campaign. Among the multiple decision samples, the positive samples are the matching pairs between the historical marketing campaign and the corresponding converted users, and the negative samples are the matching pairs between the historical marketing campaign and the non-converted users. Converted users are users who have converted behavior in the historical marketing campaign, and non-converted users are users who have not converted behavior in the historical marketing campaign but have converted behavior in other historical marketing campaigns of the same category.

[0085] It should be noted that, relying on the conversion logs of the marketing platform, all historical marketing activities that have had real purchase conversions within a preset time period are selected to form an activity set; for each historical marketing activity in the set, each user who actually purchased the product of the activity is matched with the activity, and these matching pairs are the positive samples; therefore, the number of positive samples of an activity is exactly equal to the number of actual purchasing users of the activity.

[0086] Negative samples also need to be constructed for each historical marketing campaign. This requires combining browsing logs and conversion logs, and following a pre-defined sampling ratio. This ratio specifies the quantitative relationship between positive and negative samples for each type of marketing campaign. Specifically, the number of negative samples for a historical marketing campaign is not arbitrary, but calculated based on the total number of visitors, the actual number of conversions, and this sampling ratio. From users who viewed the campaign but did not make a purchase, a corresponding number of users are extracted according to the calculation results. Each such user forms a matching pair with the campaign, which is the negative sample.

[0087] The decision sample set consists of all these positive and negative samples. Each positive sample is a pair of historical marketing campaigns and a converted user, which is someone who made a purchase during the campaign. Each negative sample is a pair of historical marketing campaigns and a non-converted user, but the non-converted user has special requirements: not only did they not make a purchase during the campaign, but they must have also made a purchase during another historical marketing campaign in the same product category. In other words, negative sample users are not completely ineffective users who do not shop, but rather users who have the ability to purchase similar products but choose not to choose the current campaign.

[0088] In an optional embodiment, the negative samples corresponding to each historical marketing activity in the historical marketing activity set are determined based on the marketing platform's browsing logs, conversion logs, and a preset sample sampling ratio, including:

[0089] For any historical marketing campaign in the collection of historical marketing campaigns, obtain the page views of that historical marketing campaign and the total page views of all historical marketing campaigns in the category to which that historical marketing campaign belongs from the browsing log;

[0090] Based on conversion records, obtain the conversion volume of the historical marketing campaign and the total conversion volume of all historical marketing campaigns in the category to which the historical marketing campaign belongs; the conversion volume is used to represent the number of users who have converted for the historical marketing campaign.

[0091] Calculate the negative sample coefficient corresponding to the historical marketing campaign based on the sample sampling ratio, the total page views of all historical marketing campaigns, and the total conversion of all historical marketing campaigns;

[0092] Multiply the difference between the page views and conversions of the historical marketing campaign by the negative sample coefficient to obtain the number of negative samples corresponding to the historical marketing campaign.

[0093] Based on the number of negative samples corresponding to the historical marketing campaign, a corresponding number of non-converted users are selected from the negative sample candidate set of the category to which the historical marketing campaign belongs, forming the negative samples corresponding to the historical marketing campaign; wherein, the negative sample candidate set is the set of users who have converted for any historical marketing campaign in the same category.

[0094] It should be noted that for any activity in the historical marketing campaign set, firstly, two values ​​are obtained from the browsing log: the number of views of the activity itself, and the total number of views of all historical marketing campaigns in the category to which the activity belongs; then, two values ​​are also obtained from the conversion records: the conversion rate of the activity itself, and the total conversion rate of all historical marketing campaigns in the category; here, the conversion rate refers to the number of users who made a purchase behavior for the activity.

[0095] The negative sample coefficient for this campaign is calculated based on three factors: the preset sample sampling ratio, the total page views of all historical marketing campaigns in this category, and the total conversion rate of all historical marketing campaigns in this category. The result of the negative sample coefficient calculation is directly used to adjust the number of negative samples.

[0096] After obtaining the negative sample coefficient, subtract the conversion rate of the activity from the number of views of the activity. This difference represents the number of users who viewed the activity but did not make a purchase. Then, multiply this difference by the negative sample coefficient to obtain the number of negative samples that should be drawn for the activity.

[0097] Based on this negative sample count, a corresponding number of non-converted users are extracted from the negative sample candidate set of the category to which the activity belongs. The negative sample candidate set has a clear definition: it is a set of all users who have made purchases in any historical marketing activity under this category. That is to say, although these users did not purchase in the current specific activity, they have made conversions in other activities under this category. Users are selected from this candidate set and matched with the activity to form the negative sample corresponding to the activity.

[0098] In an optional embodiment, determining the user preference features and marketing activity attribute features corresponding to the decision sample includes:

[0099] Obtain user behavior characteristics, positive feedback keywords, and market trend correlation data during high-conversion periods of historical marketing campaigns;

[0100] Extract user behavior pattern features based on user behavior characteristics during high-conversion periods;

[0101] Based on the timestamps, positive feedback keywords, and market trend correlation data corresponding to user behavior characteristics during high-conversion periods, the correlation characteristics are determined;

[0102] Based on user behavior patterns and relationship characteristics, multiple historical marketing campaigns are clustered to obtain multiple marketing campaign categories.

[0103] Identify the target marketing campaign category from multiple marketing campaign categories;

[0104] User behavior characteristics during high-conversion periods of historical marketing campaigns included in the target marketing campaign category are used as user preference characteristics, and positive feedback keywords and market trend correlation data of historical marketing campaigns included in the target marketing campaign category are used as marketing campaign attribute characteristics.

[0105] It should be noted that three types of data are obtained for each historical marketing campaign during high-conversion periods: user behavior characteristics, positive feedback keywords, and market trend correlation data; these three types of data form the basis for subsequent feature extraction.

[0106] Based on user behavior characteristics during high-conversion periods, user behavior pattern characteristics are extracted. This process essentially transforms the original behavioral trajectory into more generalized behavioral patterns, such as frequently occurring operation sequences or typical behavioral rhythms.

[0107] By utilizing timestamps, positive feedback keywords, and market trend data corresponding to user behavior characteristics during high-conversion periods, we can construct correlation features. Specifically, we align and combine user behavior at a specific point in time, positive words used by users during that period, and external market dynamics to find the inherent connections between the three.

[0108] User behavior pattern features and association features are combined as the basis for clustering, and all historical marketing activities are grouped. After clustering, multiple marketing activity categories are obtained, and the activities within each category are similar in terms of user behavior patterns and associations.

[0109] From these marketing campaign categories, identify one or more target marketing campaign categories; the criteria for identification can be set according to the needs of subsequent modeling, such as selecting categories with a high proportion or outstanding conversion results.

[0110] The user behavior characteristics during high-conversion periods of all historical marketing activities included in the target marketing campaign category are collectively used as user preference characteristics. This means that user preference characteristics do not come from a single activity, but from typical behavioral performances summarized from similar activities. At the same time, positive feedback keywords and market trend correlation data of all historical marketing activities in this category are collectively used as marketing campaign attribute characteristics. In this way, user preference characteristics reflect the common behavioral patterns of users under this type of activity, while marketing campaign attribute characteristics reflect the common positive feedback points and external market environment of this type of activity.

[0111] In an optional embodiment, determining a target marketing campaign category from a plurality of marketing campaign categories includes:

[0112] Obtain user feedback density, positive feedback sentiment index, and conversion growth rate index for historical marketing campaigns within a preset time period for each marketing campaign category;

[0113] Based on user feedback density, positive feedback sentiment index, and conversion growth rate index, determine the marketing performance evaluation value corresponding to each marketing campaign category;

[0114] The marketing campaign category corresponding to the highest marketing performance evaluation value will be used as the target marketing campaign category.

[0115] It should be noted that three types of data are obtained for each historical marketing campaign during high-conversion periods: user behavior characteristics, positive feedback keywords, and market trend correlation data; these three types of data form the basis for subsequent feature extraction.

[0116] Based on user behavior characteristics during high-conversion periods, user behavior pattern characteristics are extracted. This process essentially transforms the original behavioral trajectory into more generalized behavioral patterns, such as frequently occurring operation sequences or typical behavioral rhythms.

[0117] At the same time, we use the timestamps, positive feedback keywords, and market trend data corresponding to user behavior characteristics during high conversion periods to construct correlation features; specifically, we align and combine user behavior at a specific point in time, positive words in user feedback during that period, and external market dynamics to find the inherent connection between the three.

[0118] User behavior pattern features and association features are combined as the basis for clustering, and all historical marketing activities are grouped. After clustering, multiple marketing activity categories are obtained, and the activities within each category are similar in terms of user behavior patterns and associations.

[0119] From these marketing campaign categories, identify one or more target marketing campaign categories; the criteria for identification can be set according to the needs of subsequent modeling, such as selecting categories with a high proportion or outstanding conversion results.

[0120] The user behavior characteristics during high-conversion periods of all historical marketing activities included in the target marketing campaign category are collectively used as user preference characteristics. This means that user preference characteristics do not come from a single activity, but from typical behavioral performances summarized from similar activities. At the same time, positive feedback keywords and market trend correlation data of all historical marketing activities in this category are collectively used as marketing campaign attribute characteristics. In this way, user preference characteristics reflect the common behavioral patterns of users under this type of activity, while marketing campaign attribute characteristics reflect the common positive feedback points and external market environment of this type of activity.

[0121] In an optional embodiment, determining the target marketing campaign category from a plurality of marketing campaign categories further includes:

[0122] Obtain the target budget factor corresponding to the target user group, and obtain the budget factor corresponding to the historical marketing activities in each marketing activity category; where the budget factor represents the budget amount of the marketing activity;

[0123] Determine the difference between the budget factor corresponding to the historical marketing campaign and the target budget factor, and use this as the budget factor difference value;

[0124] For each marketing campaign category, calculate the sum of the marketing performance evaluation value corresponding to that category and the budget factor difference value corresponding to that category, and use it as the first updated evaluation value for that category;

[0125] The marketing campaign category corresponding to the highest first update rating value will be used as the target marketing campaign category.

[0126] It should be noted that, based on the previous selection of target marketing campaign categories based on marketing performance evaluation values, a new constraint factor has been introduced, namely the budget factor.

[0127] Obtain the target budget factor corresponding to the target user group; this target budget factor represents the planned budget amount when conducting marketing activities for this user group; at the same time, also obtain the budget factors corresponding to all historical marketing activities in each marketing activity category, and the budget factor of each activity represents the actual budget amount invested in the activity;

[0128] For each historical marketing campaign, the difference between the campaign's budget factor and the target budget factor is calculated. This difference is called the budget factor difference value. The smaller the difference value, the closer the budget level of the historical campaign is to the current target budget.

[0129] For each marketing campaign category, the previously obtained marketing performance evaluation value for that category is added to the budget factor difference value corresponding to that category to obtain a new value, called the first updated evaluation value. It should be noted that a marketing campaign category may contain multiple historical marketing campaigns, so the budget factor difference value for that category needs to be aggregated in some way, such as taking the average or representative value of the budget factor difference values ​​of all campaigns within the category.

[0130] The first updated evaluation value of all marketing campaign categories is compared, and the category with the highest value is determined as the target marketing campaign category. This means that not only the historical performance of the campaign itself is considered, but also the degree of matching between the past budget size of the campaign and the current target budget. The category with the best overall performance is selected.

[0131] In an optional embodiment, determining the target marketing campaign category from a plurality of marketing campaign categories further includes:

[0132] Obtain the target brand tone characteristic value corresponding to the target user group, and obtain the brand tone characteristic value corresponding to the historical marketing activities in each marketing activity category; among them, the brand tone characteristic value represents the youthfulness and high-end characteristics of the marketing activities;

[0133] The difference between the brand tone characteristic value corresponding to the historical marketing campaign and the target brand tone characteristic value is determined as the brand tone difference value;

[0134] For each marketing campaign category, the sum of the marketing performance evaluation value, budget factor difference value, and brand tone difference value corresponding to that category is calculated as the second updated evaluation value for that category;

[0135] The marketing campaign category corresponding to the second highest updated evaluation value will be used as the target marketing campaign category.

[0136] It should be noted that, based on the existing combination of marketing performance evaluation values ​​and budget factor differences, brand tone characteristics are further introduced as a new constraint dimension.

[0137] Obtain the target brand tone characteristic value corresponding to the target user group; this characteristic value is used to characterize the brand style preferred by the target user group, specifically including two dimensions: youthfulness and premiumization; at the same time, also obtain the brand tone characteristic value corresponding to all historical marketing activities in each marketing campaign category, and the brand tone characteristic value of each activity also measures the performance of the activity in terms of youthfulness and premiumization.

[0138] For each historical marketing campaign, the difference between the campaign's brand tone characteristic value and the target brand tone characteristic value is calculated. This difference is called the brand tone difference value. The smaller the difference value, the more consistent the brand style of the historical campaign is with the preferences of the target user group.

[0139] For each marketing campaign category, three values ​​are added together: the first is the marketing performance evaluation value previously obtained for that category, the second is the budget factor difference value for that category, and the third is the brand tone difference value for that category; the new value obtained after adding them together is called the second updated evaluation value.

[0140] The second updated evaluation value of all marketing campaign categories is compared, and the category with the highest value is determined as the target marketing campaign category. This means that the category that best combines the historical marketing performance of the campaign, the degree of matching of budget size, and the degree of matching of brand style with target user preferences is finally selected.

[0141] In an optional embodiment, determining the interaction characteristics between the user corresponding to the decision sample and the marketing campaign includes:

[0142] Based on user information and marketing activity information in the decision sample, determine the user's corresponding preference tags and the marketing activity's corresponding attribute tags;

[0143] When preference tags and attribute tags are the same, the category crossover feature between users and marketing campaigns is identified as the first category crossover feature.

[0144] When preference tags and attribute tags are different, the category crossover feature between users and marketing campaigns is identified as the second category crossover feature.

[0145] The number of matching tags between multiple preference tags for users and multiple attribute tags for marketing campaigns is determined as a numerical cross feature.

[0146] It should be noted that, based on the user information corresponding to the user in the decision sample, the user's preference tags are extracted. The preference tags represent the user's interest tendencies in different product categories. At the same time, based on the activity information of the marketing campaign, the attribute tags of the campaign are extracted. The attribute tags represent the product category or activity type to which the campaign belongs.

[0147] Compare the user's preference tags with the marketing campaign's attribute tags; if they are the same, meaning the category the user is interested in happens to be the category to which this campaign belongs, then the category cross feature is determined as the first category cross feature, which is a specific value or identifier; if they are different, meaning the category the user is interested in is inconsistent with the category to which this campaign belongs, then the category cross feature is determined as the second category cross feature, which is another different value or identifier.

[0148] Further processing is done for cases where both users and marketing campaigns contain multiple tags; the number of tags that match between multiple preference tags of users and multiple attribute tags of marketing campaigns is counted; that is, each category that the user is interested in is checked one by one to see if it appears in the attribute tags of the campaign, and then the total number of successful matches is counted; this number of matched tags is directly used as a numerical cross feature; this feature is a numerical value that reflects the degree of overlap between users and campaigns across multiple tag dimensions.

[0149] In an optional embodiment, the intelligent marketing decision-making model is parameter-adjusted based on the predicted score corresponding to the decision sample and the positive or negative sample label of the decision sample until the convergence condition is met, thereby obtaining the trained intelligent marketing decision-making model, including:

[0150] Discrete features among user preference features, marketing campaign attribute features, user historical behavior features, marketing campaign conversion features, and user-marketing campaign interaction features are embedded to obtain dense feature vectors, and continuous features are normalized to obtain numerical feature vectors.

[0151] Dense feature vectors and numerical feature vectors are concatenated to generate a unified feature representation vector;

[0152] The unified feature representation vector is input into the factorization machine layer to learn the second-order interaction relationship between features and generate the first interaction feature vector.

[0153] A unified feature representation vector is simultaneously input into a multi-head self-attention layer to capture the deep interaction relationship between user history behavior and marketing campaign attributes, generating a second interaction feature vector.

[0154] The unified feature representation vector, the first interactive feature vector, and the second interactive feature vector are concatenated to obtain the fused feature vector;

[0155] The fused feature vector is input into a deep neural network consisting of multiple fully connected layers, and after activation function and regularization processing, an abstract feature vector is obtained.

[0156] The abstract feature vector is input into the output layer, processed by a fully connected layer and a mapping function to obtain the prediction score corresponding to the decision sample;

[0157] Calculate the binary cross-entropy loss using the predicted score and the positive and negative sample labels of the decision sample;

[0158] Based on the binary cross-entropy loss, the parameters of the factorization machine layer, multi-head self-attention layer, deep neural network, and output layer are updated through gradient backpropagation;

[0159] Repeat the parameter update process until the convergence condition is met, and the trained intelligent marketing decision model is obtained.

[0160] It should be noted that the input features are preprocessed; all discrete features, such as category information in user preference tags or type information in marketing campaign attributes, will be converted into dense feature vectors through embedding, that is, each discrete value will be represented by a dense numerical vector; all continuous features, such as the number of clicks in user history or the conversion rate of marketing campaigns, will be converted into numerical feature vectors through normalization, and unified to the same unit range.

[0161] The dense feature vector and the numerical feature vector are concatenated in a dimensional way to generate a unified feature representation vector; this vector contains all the information of the original input.

[0162] A unified feature representation vector is fed into two different network structures simultaneously. The first is a factorization machine layer, which is specifically designed to learn the second-order interaction relationships between features, such as the combined effect between user preferences and marketing campaign attributes, and ultimately generates the first interaction feature vector. The second is a multi-head self-attention layer, which is used to capture the deep interaction relationships between user historical behavior and marketing campaign attributes. It can pay attention to the association weights between different behavior sequence positions and campaign attributes, and ultimately generate the second interaction feature vector.

[0163] The three vectors are concatenated: the initial unified feature representation vector, the first interaction feature vector, and the second interaction feature vector; the concatenation results in a fused feature vector that simultaneously contains the original features, second-order interaction information, and deep interaction information.

[0164] The fused feature vector is input into a deep neural network consisting of multiple fully connected layers. In this network, the data undergoes linear transformation, activation function processing, and regularization processing at each layer in turn, gradually extracting higher-level and more abstract representations, and finally obtaining abstract feature vectors.

[0165] The abstract feature vector is finally fed into the output layer, which uses a fully connected transformation and a mapping function to convert the abstract feature vector into a scalar value. This value is the prediction score corresponding to the decision sample, which represents the degree to which the model judges that the user is inclined to convert to the marketing campaign.

[0166] After calculating the predicted score, it is compared with the true positive and negative sample labels of the sample. The binary cross-entropy loss function is used to measure the degree of difference between the predicted score and the true label; the larger the loss value, the less accurate the model prediction is.

[0167] Based on the calculated binary cross-entropy loss, the gradient of the loss is backpropagated layer by layer through the gradient backpropagation algorithm, and the model parameters of the factorization machine layer, the multi-head self-attention layer, all fully connected layers in the deep neural network, and the output layer are updated; each parameter update aims to reduce the loss value.

[0168] Repeat all the above steps, including feature concatenation, forward computation of each layer, loss calculation, and parameter update of gradient backpropagation; this process continues until the preset convergence condition is met, such as the loss value falling below a certain threshold or the parameter update magnitude stabilizing; when the convergence condition is met, the training is complete, and the resulting intelligent marketing decision model can be used to score and predict new candidate marketing activities.

[0169] Example 2, please refer to Figure 2 This invention provides a technical solution: a big data-based intelligent marketing decision-making system, applicable to the aforementioned big data-based intelligent marketing decision-making method, comprising:

[0170] Data acquisition unit 1 is configured to acquire multiple historical marketing campaigns for the target user group, and to acquire user behavior characteristics, positive feedback keywords, and market trend correlation data during the high-conversion periods of each historical marketing campaign.

[0171] Sample construction unit 2 is configured to determine the decision sample set based on a set of user behavior data with conversion records; the conversion records are used to record conversion data of users' purchase behavior in historical marketing activities;

[0172] Feature extraction unit 3 is configured to determine the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and user-marketing activity interaction features corresponding to the decision sample;

[0173] Marketing prediction unit 4 is configured to input the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and interaction features corresponding to the decision sample into the initial intelligent marketing decision model for prediction processing, and obtain the prediction score corresponding to each decision sample; the prediction score is used to characterize the user's conversion tendency for historical marketing activities;

[0174] Marketing decision unit 5 is configured to adjust the parameters of the intelligent marketing decision model based on the predicted score corresponding to the decision sample and the positive and negative sample labels of the decision sample until the convergence condition is met, thereby obtaining the trained intelligent marketing decision model; acquire the current candidate marketing activities, use the trained intelligent marketing decision model to score the candidate marketing activities, and determine the intelligent marketing decision scheme for the target user group based on the score.

[0175] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A big data-based intelligent marketing decision-making method, characterized in that, include: Obtain multiple historical marketing campaigns for the target user group, and acquire user behavior characteristics, positive feedback keywords, and market trend correlation data for the high-conversion periods corresponding to each historical marketing campaign; Based on a set of user behavior data with conversion records, a decision sample set is determined; conversion records are used to record conversion data of users' purchase behavior in response to historical marketing campaigns. Determine the user preference characteristics, marketing campaign attribute characteristics, user historical behavior characteristics, marketing campaign conversion characteristics, and user-marketing campaign interaction characteristics corresponding to the decision sample; The user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and interaction features corresponding to the decision sample are input into the initial intelligent marketing decision model for prediction processing to obtain the prediction score corresponding to each decision sample; the prediction score is used to characterize the user's conversion tendency for historical marketing activities. Based on the predicted scores corresponding to the decision samples and the positive and negative sample labels of the decision samples, the parameters of the intelligent marketing decision model are adjusted until the convergence condition is met, and the trained intelligent marketing decision model is obtained; the current candidate marketing activities are obtained, the trained intelligent marketing decision model is used to score the candidate marketing activities, and the intelligent marketing decision scheme for the target user group is determined based on the score. Among them, a decision sample set is determined based on a set of user behavior data with conversion records, including: Based on the conversion logs of the marketing platform, determine the set of historical marketing campaigns with conversion records within a preset time period, and the positive samples corresponding to each historical marketing campaign in the set of historical marketing campaigns; wherein, the conversion logs include all conversion records in the marketing platform, and the number of positive samples corresponding to historical marketing campaigns is equal to the number of users who have converted from historical marketing campaigns; Based on the browsing logs, conversion logs, and preset sample sampling ratio of the marketing platform, the negative samples corresponding to each historical marketing activity in the historical marketing activity set are determined. The browsing logs include browsing records of each marketing activity in the marketing platform, and the sample sampling ratio is the ratio between positive and negative samples set for the marketing activity category. The number of negative samples corresponding to the historical marketing activity is determined based on the browsing volume, conversion volume, and sample sampling ratio of the historical marketing activity. The decision sample set includes multiple decision samples corresponding to each historical marketing campaign. Among the multiple decision samples, the positive samples are the matching pairs between the historical marketing campaign and the corresponding converted users, and the negative samples are the matching pairs between the historical marketing campaign and the non-converted users. Converted users are users who have converted behavior in the historical marketing campaign, and non-converted users are users who have not converted behavior in the historical marketing campaign but have converted behavior in other historical marketing campaigns of the same category. Specifically, based on the marketing platform's browsing logs, conversion logs, and a preset sample sampling ratio, negative samples are determined for each historical marketing campaign in the historical marketing campaign set, including: For any historical marketing campaign in the collection of historical marketing campaigns, obtain the page views of that historical marketing campaign and the total page views of all historical marketing campaigns in the category to which that historical marketing campaign belongs from the browsing log; Based on conversion records, obtain the conversion volume of the historical marketing campaign and the total conversion volume of all historical marketing campaigns in the category to which the historical marketing campaign belongs; the conversion volume is used to represent the number of users who have converted for the historical marketing campaign. Calculate the negative sample coefficient corresponding to the historical marketing campaign based on the sample sampling ratio, the total page views of all historical marketing campaigns, and the total conversion of all historical marketing campaigns; Multiply the difference between the page views and conversions of the historical marketing campaign by the negative sample coefficient to obtain the number of negative samples corresponding to the historical marketing campaign. Based on the number of negative samples corresponding to the historical marketing campaign, a corresponding number of non-converted users are selected from the negative sample candidate set of the category to which the historical marketing campaign belongs, forming the negative samples corresponding to the historical marketing campaign; wherein, the negative sample candidate set is the set of users who have converted for any historical marketing campaign in the same category.

2. The intelligent marketing decision-making method based on big data according to claim 1, characterized in that, Determine the user preference characteristics and marketing activity attribute characteristics corresponding to the decision sample, including: Obtain user behavior characteristics, positive feedback keywords, and market trend correlation data during high-conversion periods of historical marketing campaigns; Extract user behavior pattern features based on user behavior characteristics during high-conversion periods; Based on the timestamps, positive feedback keywords, and market trend correlation data corresponding to user behavior characteristics during high-conversion periods, the correlation characteristics are determined; Based on user behavior patterns and relationship characteristics, multiple historical marketing campaigns are clustered to obtain multiple marketing campaign categories. Identify the target marketing campaign category from multiple marketing campaign categories; User behavior characteristics during high-conversion periods of historical marketing campaigns included in the target marketing campaign category are used as user preference characteristics, and positive feedback keywords and market trend correlation data of historical marketing campaigns included in the target marketing campaign category are used as marketing campaign attribute characteristics.

3. The intelligent marketing decision-making method based on big data according to claim 2, characterized in that, Identify the target marketing campaign category from multiple marketing campaign categories, including: Obtain user feedback density, positive feedback sentiment index, and conversion growth rate index for historical marketing campaigns within a preset time period for each marketing campaign category; Based on user feedback density, positive feedback sentiment index, and conversion growth rate index, determine the marketing performance evaluation value corresponding to each marketing campaign category; The marketing campaign category corresponding to the highest marketing performance evaluation value will be used as the target marketing campaign category.

4. The intelligent marketing decision-making method based on big data according to claim 3, characterized in that, Determining the target marketing campaign category from multiple marketing campaign categories also includes: Obtain the target budget factor corresponding to the target user group, and obtain the budget factor corresponding to the historical marketing activities in each marketing activity category; where the budget factor represents the budget amount of the marketing activity; Determine the difference between the budget factor corresponding to the historical marketing campaign and the target budget factor, and use this as the budget factor difference value; For each marketing campaign category, calculate the sum of the marketing performance evaluation value corresponding to that category and the budget factor difference value corresponding to that category, and use it as the first updated evaluation value for that category; The marketing campaign category corresponding to the highest first update rating value will be used as the target marketing campaign category.

5. The intelligent marketing decision-making method based on big data according to claim 4, characterized in that, Determining the target marketing campaign category from multiple marketing campaign categories also includes: Obtain the target brand tone characteristic value corresponding to the target user group, and obtain the brand tone characteristic value corresponding to the historical marketing activities in each marketing activity category; among them, the brand tone characteristic value represents the youthfulness and high-end characteristics of the marketing activities; The difference between the brand tone characteristic value corresponding to the historical marketing campaign and the target brand tone characteristic value is determined as the brand tone difference value; For each marketing campaign category, the sum of the marketing performance evaluation value, budget factor difference value, and brand tone difference value corresponding to that category is calculated as the second updated evaluation value for that category; The marketing campaign category corresponding to the second highest updated evaluation value will be used as the target marketing campaign category.

6. The intelligent marketing decision-making method based on big data according to claim 5, characterized in that, Determine the interaction characteristics between users and marketing campaigns corresponding to the decision sample, including: Based on user information and marketing activity information in the decision sample, determine the user's corresponding preference tags and the marketing activity's corresponding attribute tags; When preference tags and attribute tags are the same, the category crossover feature between users and marketing campaigns is identified as the first category crossover feature. When preference tags and attribute tags are different, the category crossover feature between users and marketing campaigns is identified as the second category crossover feature. The number of matching tags between multiple preference tags for users and multiple attribute tags for marketing campaigns is determined as a numerical cross feature.

7. The intelligent marketing decision-making method based on big data according to claim 6, characterized in that, Based on the predicted scores and positive / negative labels of the decision samples, the parameters of the intelligent marketing decision model are adjusted until convergence is achieved, resulting in a fully trained intelligent marketing decision model, including: Discrete features among user preference features, marketing campaign attribute features, user historical behavior features, marketing campaign conversion features, and user-marketing campaign interaction features are embedded to obtain dense feature vectors, and continuous features are normalized to obtain numerical feature vectors. Dense feature vectors and numerical feature vectors are concatenated to generate a unified feature representation vector; The unified feature representation vector is input into the factorization machine layer to learn the second-order interaction relationship between features and generate the first interaction feature vector. A unified feature representation vector is simultaneously input into a multi-head self-attention layer to capture the deep interaction relationship between user history behavior and marketing campaign attributes, and generate a second interaction feature vector. The unified feature representation vector, the first interactive feature vector, and the second interactive feature vector are concatenated to obtain the fused feature vector; The fused feature vector is input into a deep neural network consisting of multiple fully connected layers, and after activation function and regularization processing, an abstract feature vector is obtained. The abstract feature vector is input into the output layer, processed by a fully connected layer and a mapping function to obtain the prediction score corresponding to the decision sample; Calculate the binary cross-entropy loss using the predicted score and the positive and negative sample labels of the decision sample; Based on the binary cross-entropy loss, the parameters of the factorization machine layer, multi-head self-attention layer, deep neural network, and output layer are updated through gradient backpropagation; Repeat the parameter update process until the convergence condition is met, and the trained intelligent marketing decision model is obtained.

8. A big data-based intelligent marketing decision-making system, applicable to the big data-based intelligent marketing decision-making method described in any one of claims 1-7, characterized in that, include: The data acquisition unit is configured to acquire multiple historical marketing campaigns for the target user group, and to acquire user behavior characteristics, positive feedback keywords, and market trend correlation data during the high-conversion periods of each historical marketing campaign. The sample construction unit is configured to determine the decision sample set based on a set of user behavior data with conversion records; the conversion records are used to record conversion data of users' purchase behavior in historical marketing activities; The feature extraction unit is configured to determine the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and user-marketing activity interaction features corresponding to the decision sample; The marketing prediction unit is configured to input the user preference features, marketing activity attribute features, user historical behavior features, marketing activity conversion features, and interaction features corresponding to the decision sample into the initial intelligent marketing decision model for prediction processing, and obtain the prediction score corresponding to each decision sample; the prediction score is used to characterize the user's conversion tendency for historical marketing activities; The marketing decision-making unit is configured to adjust the parameters of the intelligent marketing decision-making model based on the predicted score corresponding to the decision sample and the positive and negative sample labels of the decision sample until the convergence condition is met, thereby obtaining the trained intelligent marketing decision-making model; acquire the current candidate marketing activities, use the trained intelligent marketing decision-making model to score the candidate marketing activities, and determine the intelligent marketing decision-making plan for the target user group based on the score.

Images