Marketing effect analysis method and system based on AI

Abstract

The invention relates to the technical field of business marketing activity management and effect analysis, in particular to an AI-based marketing effect analysis method and system. The method comprises the steps of obtaining user behavior data, marketing activity data and acquisition constraint configuration, performing unified time reference binding, identifier binding and noise segment elimination processing, and generating a model training input set structure; executing evaluation model training and reasoning, and generating a model evaluation result structure; bid adjustment, channel distribution and audience subdivision calculation are carried out, and a strategy update package is generated in combination with strategy conflict prompt information; and executing state registration and feedback processing to dynamically generate acquisition constraint configuration, and forming an acquisition constraint configuration write-back structure after multi-dimensional analysis and aggregation to realize an adaptive closed loop of acquisition, training, reasoning and strategy backtracking. According to the method, automation and traceable optimization of marketing data analysis are realized, and the accuracy and the intelligent level of a putting strategy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of commercial marketing activity management and effectiveness analysis, and particularly relates to a marketing effectiveness analysis method and system based on AI. BACKGROUND

[0002] In the technical field of commercial marketing activity management and effectiveness analysis, the existing scheme of marketing activity data processing and delivery management usually relies on independent data collection processes, offline index statistics processes and manually configured delivery strategy issuing processes, and has limitations such as difficulty in binding user behavior data and marketing activity data under a unified time reference, difficulty in keeping sampling frequency fields and filtering threshold fields consistent for a long time, and lack of dynamic feedback of the fields. The existing method collects user behavior data and marketing activity data through fixed monitoring rules, and then pushes delivery instructions by manual or semi-automatic methods, and relies on post-reporting to review the delivery status. In the scene where the sampling constraint configuration changes rapidly, the delivery strategy conflicts, and the target audience overlaps, it is easy to have situations such as un-identified strategy conflict prompt information, delayed bid adjustment calculation and channel distribution adjustment calculation, and out-of-sync audience segmentation adjustment calculation and actual delivery instruction issuing, which makes it difficult to meet the stable execution of bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation of marketing activities. For the joint processing of user behavior data, marketing activity data and sampling constraint configuration, the existing technology generally lacks continuous cooperation from unified time reference binding, user identifier and activity identifier binding, missing fragment only marking and noise segment removal processing, to generating model training input set structure and driving evaluation model training, reasoning and result normalization processing, lacks stable association of historical version reference registration and group label solidification combined with unique version number, and forming a model evaluation result structure containing an activity performance index set, a user group score set and a model version record, and also lacks automatic connection of directly converting the model evaluation result structure into a strategy update package generation input set and outputting a strategy update package with a reduction mark and a limitation identifier, which makes it difficult to form a closed-loop process of delivery instruction issuing, execution state registration, return channel binding processing, state abnormal field recording, sampling constraint configuration generation, multi-dimensional analysis index aggregation, strategy change record aggregation, execution state aggregation, visual report input set generation, chart configuration generation, natural language summary generation, time series comparison item generation to sampling constraint configuration feedback structure in the marketing activity delivery site, resulting in update lag of the sampling frequency field, the filtering threshold field and the pull channel priority field in the actual commercial delivery resource allocation scene, and the activity performance index set, the user group score set and the model version record of the marketing activity cannot timely support the coordinated adjustment of the strategy update package and the sampling constraint configuration. SUMMARY

[0003] In order to solve the above technical problems, the application provides an AI-based marketing effect analysis method, comprising: Obtain user behavior data, marketing activity data and collection constraint configuration, perform unified time reference binding, user identification and activity identification binding, execute missing segment only marking and noise segment elimination processing containing sampling frequency field and filtering threshold field in collection constraint configuration, and perform field dictionary unified mapping processing to generate model training input set structure; Perform evaluation model training, reasoning and result normalization processing, combine unique version number to perform historical version reference registration and group label solidification, generate model evaluation result structure containing activity performance indicator set, user group score set and model version record, and build strategy update package generation input set; Perform bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation, generate strategy update package with reduction mark and limit mark combined with strategy conflict prompt information, perform delivery instruction issuing, execution state registration and return channel binding processing, and generate collection constraint configuration containing state exception field; Perform multi-dimensional analysis indicator aggregation, strategy change record aggregation and execution state aggregation, generate visualization report input set, and build collection constraint configuration backwrite structure through chart configuration generation, natural language summary generation and time sequence comparison item generation, for backwriting sampling frequency field, filtering threshold field and pull channel priority field.

[0004] Further, the process of user behavior data further comprises: Obtain strategy update package generation input set, perform bid adjustment calculation for low performance activities, perform channel distribution adjustment calculation for high performance activities, and perform audience segmentation adjustment calculation for stable group labels combined with strategy conflict prompt information, to generate strategy update package containing reduction mark and limit mark; Extract delivery instruction set, budget allocation instruction set and audience instruction set, perform delivery execution end issuing, execution state registration based on unified activity identification unique code and return channel binding processing, and generate execution back record with state exception field; Perform sampling frequency update calculation, filtering threshold update calculation and collection end pull rule update processing, dynamically adjust sampling frequency field, filtering threshold field and pull channel priority field combined with state exception field, and generate new version of collection constraint configuration.

[0005] Further, automatically adjust the collection frequency field according to the request receiving state, the applied state and the partially applied state in the execution state item, when the execution state item displays the applied state, shorten the interval of the collection frequency field corresponding to the unified activity identification unique code and the channel code field.

[0006] Furthermore, when the execution status entry shows that the budget allocation instruction set is marked as scheduled or partially applied and the status exception field indicates that resources are limited, the interval of the collection frequency field under the corresponding unified activity identifier unique code in the channel code field is extended.

[0007] Furthermore, the rejection reason status is bound to the target population label description and stable subgroup label, recorded as a sampling frequency exception rule, and directly written into the collection constraint configuration in the next round of collection.

[0008] Furthermore, the system reads the status anomaly field, reduction flag, limit identifier, and return path, jointly analyzes the execution status entries, updates the filter threshold field, and writes the results into the set of filter threshold fields to be applied.

[0009] Furthermore, when an execution status entry carries a decrease flag, the system increases the filtering threshold field of the corresponding stable group label in order to remove records with extremely short dwell times, repeated click records, or transaction behaviors that are inconsistent with the target group label description.

[0010] Furthermore, based on the analysis of the return path field of the execution return record, it is determined whether the behavior feedback flows into the browser terminal's data collection program, transaction service logs, or customer service supplementary entry channel, and the priority field of the pull channel is rearranged.

[0011] Furthermore, after completing the calculation of sampling frequency update, filtering threshold update, and acquisition end pull rule update, the updated sampling frequency field, filtering threshold field, pull channel priority field, sampling frequency exception rule, and pull channel masking table are merged together to form a new version of the acquisition constraint configuration.

[0012] Furthermore, an AI-based marketing effectiveness analysis system, applied to any of the methods described above, includes: The data access and alignment module is used to acquire user behavior data, marketing activity data and collection constraint configuration, perform unified time base binding, user identifier binding, activity identifier binding, missing segment marking without evaluation processing, noise segment removal processing, unified field dictionary processing, activity tag binding, version number registration, and distribution channel mapping processing, generate model training input set structure, and provide the model training input set structure to the evaluation model training and inference module; The evaluation model training and inference module is used to perform model loading, historical version reference registration, and inference session initialization processing on the model training input set structure to generate a preliminary inference result set. It is used to extract the activity performance indicator set, user group score set, and model version record from the preliminary inference result set, perform indicator normalization processing, group label solidification, and version disk registration to generate the model evaluation result structure and provide the model evaluation result structure to the strategy generation and execution feedback module. The strategy generation and execution back module is used for performing activity priority grouping, low performance activity list extraction and high performance activity list extraction on the model evaluation result structure, generating a strategy update package generation input set, performing bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation after obtaining the strategy update package generation input set, generating a strategy update package, extracting a delivery instruction set, a budget allocation instruction set and an audience instruction set from the strategy update package, performing execution end delivery, execution state registration and return channel binding processing, generating an execution back record, performing sampling frequency update calculation, filtering threshold update calculation and collection end pull rule update processing according to the execution back record, generating collection constraint configuration, and providing the collection constraint configuration to the visual report generation module and the data access and alignment module respectively; The visual report generation module is used for performing multi-dimensional analysis index aggregation, strategy change record aggregation and execution state aggregation processing after obtaining the collection constraint configuration, generating a visual report input set, extracting an activity performance index set, a strategy change record and an execution state from the visual report input set, performing chart configuration generation, text summary paragraph generation and time sequence comparison item generation, generating a multi-dimensional analysis report structure, generating a collection constraint configuration backwrite structure after performing strategy update tracking record extraction, collection strategy backtracking record extraction and activity configuration version record extraction on the multi-dimensional analysis report structure, and feeding back the collection constraint configuration backwrite structure to the data access and alignment module.

[0013] The key innovations of the present application include: (1) After obtaining user behavior data, marketing activity data and collection constraint configuration, the results are subjected to unified time reference binding, user identifier and activity identifier binding, missing segment only marking and not valuation and noise segment elimination processing, field dictionary unified mapping processing, a model training input set structure is generated, and the model training input set structure is used as a unified entrance to pass in the marketing activity analysis link, forming a structured input end control for user behavior data and marketing activity data.

[0014] (2) In the evaluation model training, reasoning and result normalization processing process, the historical version reference registration and group label solidification are performed in combination with the unique version number, the model evaluation result structure containing the activity performance index set, the user group score set and the model version record is generated, and the strategy update package generation input set is directly constructed, and then the bid adjustment calculation, the channel distribution adjustment calculation and the audience segmentation adjustment calculation are performed, the strategy update package with the reduction mark and the limitation mark is generated in combination with the strategy conflict prompt information, and the automatic connection link from the model evaluation result structure to the strategy update package is formed.

[0015] (3) Based on the delivery instruction issuing, execution state registration and return channel binding processing, the collection constraint configuration containing the state exception field is obtained, multi-dimensional analysis index aggregation, policy change record aggregation and execution state aggregation are performed, visual report input set is generated, and through chart configuration generation, natural language summary generation and time sequence comparison item generation, the collection constraint configuration write-back structure is constructed, which is used to write back the sampling frequency field, the filtering threshold field and the pull channel priority field, so that the collection constraint configuration completes generation, use, write-back and re-activation in the same process, forming a closed-loop update control link covering the sampling frequency field, the filtering threshold field and the pull channel priority field.

[0016] The following are its main beneficial effects: (1) Through the model training input set structure, the user behavior data, the marketing activity data and the collection constraint configuration are uniformly bound in the generation stage, the missing fragments are only marked and not valued, the noise segments are removed, and the field dictionary is uniformly mapped, so that the input end forms a single and traceable data entry, avoiding the analysis link fragmentation caused by inconsistent time sequence, inconsistent field scope and noise segment interference in the prior art, supporting the direct connection between the data collection link and the evaluation link in the business marketing activity management and effectiveness analysis scene.

[0017] (2) Through the unique version number, historical version reference registration and group label solidification, the activity performance index set, the user group evaluation score set and the model version record are combined into the model evaluation result structure, and the strategy update package generation input set is directly generated from the model evaluation result structure, and then through bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation, the strategy update package with the reduction mark and the limitation mark is output, realizing the continuous control process from evaluation to delivery instruction issuing, avoiding the problem that the evaluation link and the delivery link are independent of each other and the strategy conflict prompt information is difficult to fall into the delivery instruction in time in the prior art, supporting the targeted scheduling of bid, channel and audience in the business marketing activity management and effectiveness analysis scene.

[0018] (3) The acquisition constraint configuration containing the state exception field obtained by performing the state registration and return channel binding processing is aggregated after multi-dimensional analysis index aggregation, policy change record aggregation, and execution state aggregation to form a visualization report input set, and then the acquisition constraint configuration is written back to the structure through chart configuration generation, natural language summary generation, and time series comparison item generation. The sampling frequency field, the filtering threshold field, and the pull channel priority field are written back, so that the acquisition constraint configuration is updated in a closed loop in the same business marketing activity management and effectiveness analysis process, avoiding the long-term fixed state of the acquisition frequency field and the filtering threshold field in the traditional scheme, and the disconnection between the execution state and the subsequent acquisition strategy. The acquisition constraint configuration and the strategy update package form a continuous corresponding relationship in the business delivery scenario. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of an AI-based marketing effectiveness analysis method provided by an embodiment of the present application; Figure 2 A structural block diagram of an AI-based marketing effectiveness analysis system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0020] Embodiment one: refer to Figure 1 A flowchart of an AI-based marketing effectiveness analysis method provided by an embodiment of the present application, which can at least include steps S100-S400: S100, acquire user behavior data, marketing activity data, and acquisition constraint configuration, perform unified time reference binding, user identifier and activity identifier binding, execute missing segment only marking and noise segment rejection processing of the sampling frequency field and the filtering threshold field in the acquisition constraint configuration, and perform field dictionary unified mapping processing to generate a model training input set structure; S200, execute evaluation model training, reasoning, and result normalization processing, perform historical version reference registration and group label solidification combined with a unique version number, generate a model evaluation result structure containing an activity performance index set, a user group score set, and a model version record, and build a strategy update package generation input set; S300, perform bid adjustment calculation, channel distribution adjustment calculation, and audience segmentation adjustment calculation, generate a strategy update package with a reduction mark and a restriction identifier combined with a strategy conflict prompt information, perform delivery instruction issuance, execution state registration, and return channel binding processing, and generate an acquisition constraint configuration containing a state exception field; S400, performing multi-dimension analysis index aggregation, strategy change record aggregation and execution state aggregation, generating a visual report input set, and constructing a collection constraint configuration write-back structure through chart configuration generation, natural language summary generation and time series comparison item generation, for writing back a sampling frequency field, a filtering threshold field and a pull channel priority field.

[0021] Step S100 includes at least steps S110-S130: S110, obtaining user behavior data, marketing activity data and collection constraint configuration, performing unified time reference binding, user identification binding and activity identification binding processing, and obtaining an aligned sampling data set; In step S110, the system acquires user behavior data, marketing activity data, and collection constraint configurations. It then performs unified time base binding, user identifier binding, and activity identifier binding on the aforementioned data, and outputs an aligned sampling dataset. User behavior data refers to the actual interactive behavior records of users in the marketing outreach chain. These records cover individual behavioral events such as page browsing, dwell time, clicks, adding to cart, placing orders, cancellations, returns, and forwarding / sharing. Each behavioral event also carries contextual fields such as event occurrence time, trigger entry point, terminal type, network environment, coarse-grained geographic label, and login status label to describe the background of the behavioral event. Marketing campaign data refers to the description information of marketing campaigns that are in the same period. The marketing campaign description information records the description of the delivery channel, the description of the reach strategy, the description of the target audience tags, the description of the creative version, the description of the budget block, and the description of the delivery time window. Among them, the description of the delivery channel indicates which media, ad placement or delivery network the marketing campaign appears on; the description of the reach strategy indicates the delivery strategy direction; the description of the target audience tags indicates the audience selection logic; the description of the creative version indicates the material number and template structure; the description of the budget block indicates the cost allocation range; and the description of the delivery time window indicates the start and end time of the marketing campaign. The data acquisition constraint configuration refers to the sampling control rules that take effect in the current data acquisition period. The sampling control rules include fields such as sampling frequency, filtering threshold, pull channel priority, maximum allowed segment length, maximum allowed hourly concurrent pulls, and abnormal return reporting strategy. Among them, the sampling frequency field limits the time interval within which the same user is allowed to be recorded once for a behavior event; the filtering threshold field limits the noise intensity below which events are considered invalid segments; the pull channel priority field limits which channel's record is retained when the same behavior event occurs in multiple reporting channels; the maximum allowed segment length field limits the maximum duration of a single pull timing window; the maximum allowed hourly concurrent pulls field limits the maximum number of pulls from the same source channel per hour; and the abnormal return reporting strategy field limits how abnormal events are recorded in subsequent steps. In step S110, the above three types of inputs are simultaneously accessed through the data access bus. The data access bus connects to the browser-side event tracking program, transaction service logs, customer service data entry channel, and marketing activity management ledger. The browser-side event tracking program records front-end interaction actions; the transaction service logs record transaction and after-sales actions; the customer service data entry channel records manually corrected behavioral segments; and the marketing activity management ledger records the master file information for the advertising channel descriptions, outreach strategy descriptions, and budget block descriptions mentioned in the marketing activity data. Through this parallel access action, step S110 obtains complete raw records covering both the behavior and configuration sides.

[0022] Specifically, step S110 first performs a unified time base binding process. This process corrects the timestamp fields of records from different sources, enabling subsequent steps to retrieve user behavior trajectories and marketing campaign placement windows based on the same timeline. Since user behavior data is often generated from multiple sources such as browser tracking, transaction service logs, and customer service data entry channels, differences in local time zones, clock drift, and network transmission latency can lead to inconsistencies in the timestamp fields of the same real-world behavioral event across different sources. The unified time base binding process comprises three actions. The first action involves the system reading the original timestamp field and associating it with the source channel identifier, registering the source channel identifier in a clock difference table. This clock difference table is an internal registration structure used to record the offset between each source channel and the server's master clock. The second action involves the system calculating the offset between the source channel and the server's master clock based on the clock difference table, and then adding this offset to the original timestamp field to form the corrected timestamp field. In the third step, the system registers the corrected timestamp field as a synchronized timestamp field, which is then written into a cache structure. This cache structure maintains the same reference path as the input in subsequent step S120. The synchronized timestamp field is considered the primary key for subsequent aggregated time sequences. Subsequent steps will sort and concatenate multi-source behavioral events for the same user according to the synchronized timestamp field, and compare the behavioral event with the campaign's delivery time window description in the marketing campaign data to determine if the behavioral event is related to a specific marketing campaign. After the unified time base binding process is completed, step S110 continues with user identifier binding and campaign identifier binding processes.

[0023] The user identification binding processing refers to merging multiple identity sources existing for the same natural person or the same terminal device, and generating a unified user identification unique code. The unified user identification unique code is aggregated from browser fingerprints, login accounts, delivery numbers, payment accounts, device serial numbers and other candidate identifiers. The system sequentially checks the credibility and completeness of the candidate identifiers according to the pull channel priority field in the collection constraint configuration, selects the candidate combination with the highest credibility and the completeness meeting the registration standard, maps the candidate combination into the unified user identification unique code, and supplements the unified user identification unique code in the user behavior data entries related to the natural person or the terminal device. The activity identification binding processing refers to establishing a unified activity identification unique code for multiple presentation forms of the same marketing activity appearing in different delivery channels, different creative versions, different budget blocks and different delivery time windows. Specifically, the system reads the delivery channel description, creative version description, budget block description and delivery time window description in the marketing activity data, retrieves the activity master record in the marketing activity data, and saves the global activity number, channel sub-number and material sub-number of the marketing activity in the activity master record. The system generates a unified activity identification unique code according to the activity master record, and supplements the unified activity identification unique code into all user behavior data entries related to the marketing activity. After completing the user identification binding processing and the activity identification binding processing, the system performs multi-key alignment and de-duplication aggregation on the multi-source original records based on the synchronization timestamp field, the unified user identification unique code and the unified activity identification unique code, removes the behavior fragments repeatedly reported by different sources within the same synchronization timestamp field window, and splits the long continuous record segments according to the allowed maximum segment length field in the collection constraint configuration, while recording the source channel traceback pointer, which is used to check the original behavior data or the original marketing activity data in subsequent steps. The aggregation result obtained through the above processing is called the aligned sampling data set in the present implementation process. The aligned sampling data set is directly passed to step S120 as the output of step S110, and is read by step S120 as an input source, used to extract user behavior feature fields, activity configuration fields and timestamp fields and generate a training input packaging structure. The aligned sampling data set is also used in the visualization report input set construction process of subsequent step S410 to play back the original behavior sequence and the delivery window relationship, thereby forming a traceable link. However, this action occurs in subsequent main steps rather than in the current step.

[0024] S120, extracting user behavior feature fields, activity configuration fields and timestamp fields from the aligned sampling data set, performing missing segment only labeling without estimation, noise segment removal and field dictionary unification processing, and generating a training input packaging structure; In step S120, the system extracts the user behavior feature field, the activity configuration field and the timestamp field from the aligned sampling dataset output from step S110, performs the missing segment only label marking and no estimation processing, the noise segment elimination processing and the field dictionary unification processing on the above-mentioned fields, and finally generates a training input packaging structure. The aligned sampling dataset is used as an input data container here, has been aggregated according to the synchronized timestamp field, the unique coding of the unified user identification and the unique coding of the unified activity identification, and the behavior segments of the same user in the same delivery window have been removed. The user behavior feature field refers to the behavior pattern features extracted from the aligned sampling dataset for a single unified user identification unique code, and the behavior pattern features include click frequency distribution, stay time stage distribution, order action sequence, cancellation action sequence, return action sequence, interval timing structure from adding to ordering, browsing path jump order, etc., which are used to describe the interaction process of the user and a certain marketing activity. The activity configuration field refers to the marketing activity running attributes extracted from the aligned sampling dataset for a unified activity identification unique code, and the marketing activity running attributes include the description of the delivery channel to which the marketing activity belongs, the description of the reach strategy, the description of the target audience label, the description of the creative version, the description of the budget block and the description of the delivery time window, which are used to describe the channel in which the marketing activity is presented, the target audience, the type of creative content and the budget state. The timestamp field refers to the synchronized timestamp field itself and the adjacent time range field appearing in groups with the synchronized timestamp field, which is used to depict the sequence and duration of the behavior segments in the time dimension. The above-mentioned three types of fields are extracted one by one and put into a transition cache structure in step S120, and the transition cache structure is a time sequence segment level container for a single unified user identification unique code and a single unified activity identification unique code in the current embodiment.

[0025] Subsequently, step S120 performs a missing segment only marking and non-valuation processing on the transition cache structure. The missing segment only marking and non-valuation processing refers to, when a continuous gap in the time stamp field is found in the user behavior feature field or the activity configuration field, not performing data calculation and not performing empirical extrapolation, but writing a missing mark in the corresponding position. The missing mark is a state record field, which indicates that there is a record gap in this time period, and at the same time records a source channel trace pointer, which is written by step S110 in the aligned sampling data set, used to locate the original reporting channel when needed in the subsequent backtracking. The running process of the missing segment only marking and non-valuation processing is as follows: the system scans all the synchronous time stamp fields in the transition cache structure, detects the breakpoints whose time interval exceeds the threshold of the sampling frequency field in the collection constraint configuration, registers the breakpoint interval as a missing mark section and adds a source channel trace pointer, and does not derive the behavior pattern feature or the marketing activity running attribute for the breakpoint interval, so as to prevent the subsequent steps from mistakenly regarding the blank section as a low interaction section and affecting the training bias. Next, step S120 performs a noise segment elimination processing on the same transition cache structure. The noise segment elimination processing refers to eliminating low credibility segments and abnormal segments based on the filtering threshold field in the collection constraint configuration. The low credibility segments can include extremely short stay time records, repeated click action records in the same synchronous time stamp field window, order placement action records appearing in a population not pointed by an open touch strategy description, etc. The abnormal segments can include behavior events that are not within the description range of the delivery time window but are marked with a uniform activity identification unique code. The system directly discards the records judged as noise segments in the noise segment elimination processing, and retains a removal reason registration field, which will be referenced in the visualization report input set in the subsequent step S410, for generating a strategy change record entry in the visualization report structure.

[0026] Finally, step S120 performs field dictionary uniformization processing on the transition cache structure after the missing segment only label marking and non-estimated value processing and the noise segment elimination processing. Field dictionary uniformization processing refers to remapping field names of different sources, different naming specifications, and different diameters to a unified field dictionary. The unified field dictionary is a standard field set within the system, recording the Chinese name of the field, the meaning of the field, the value range, the source of the acquisition channel, the corresponding unique code of the unified user identifier, and the corresponding unique code of the unified activity identifier. Field dictionary uniformization processing not only unifies the field name, but also establishes a homologous mapping record when there are synonymous fields to prevent the same semantics from being split into multiple training input slots in subsequent steps. After completing the field dictionary uniformization processing, the system re-integrates the user behavior feature fields, activity configuration fields, and timestamp fields into an input container for model training, which is referred to as a training input packaging structure in this implementation process. The training input packaging structure has a synchronous timestamp field sorting result, a unified user identifier unique code, a unified activity identifier unique code, and missing label segment information, elimination reason registration field information, and standard field names after field dictionary uniformization. The training input packaging structure is output at the end of step S120 and is directly read by step S130. The training input packaging structure will be used as input for step S130 to perform activity label binding, version number registration, and delivery channel mapping processing, and to generate a model training input set structure. The model training input set structure will also be passed to step S210 in subsequent main steps as a source of the model training input set structure required for model loading and reasoning session initialization of the evaluation model training and reasoning module.

[0027] S130, activity label binding, version number registration, and delivery channel mapping processing are performed on the training input packaging structure to generate a model training input set structure; In step S130, the system carries out activity label binding, version number registration and delivery channel mapping processing on the training input packaging structure output in step S120, and finally generates a model training input set structure. The training input packaging structure is used as the basic input container in this step, and has been processed by missing segment only label marking non-estimation processing, noise segment elimination processing and field dictionary unification processing in step S120. Therefore, it has consistent formats in terms of timing segmentation, field meaning, field name mapping, elimination reason registration field, etc., and is suitable for directly entering the label binding and version management link. Activity label binding refers to binding the user behavior feature field and activity configuration field in the training input packaging structure to the corresponding activity label. The activity label is a unified representation of the delivery object label, target audience label and creative version label registered in the marketing activity data, and is used to indicate which type of marketing activity instance the input sample belongs to. Specifically, the system reads the unique code of the unified activity identification recorded in the training input packaging structure, retrieves the activity main file record of the marketing activity data, and registers the delivery channel description, reach strategy description, target audience label description and creative version description corresponding to the unique code of the unified activity identification in the activity main file record. The system writes these descriptions into a set of activity labels, and forms a binding relationship between the set of activity labels and the user behavior feature field and timestamp field in the training input packaging structure. The binding relationship is internally registered as a label binding table, which synchronously records the unique code of the unified user identification and the synchronous timestamp field. The label binding table supports subsequent restoration of the behavior trajectory and activity configuration context of the user under a specific marketing activity in the evaluation model training and reasoning module.

[0028] The version number registration refers to generating a version number for the current state of the training input packaging structure and registering the version number. The version number registration is used to record the acquisition constraint configuration, field dictionary uniform processing result, missing segment only labeling and no valuation processing strategy, and noise segment elimination processing strategy relied on by the training input packaging structure at the generation time. Specifically, the system reads the metadata area of the training input packaging structure, and the metadata area includes the sampling frequency field, the filtering threshold field, the pull channel priority field, the allowed maximum segment length field, and the allowed maximum time concurrent pull number field. The system combines these metadata fields with the generated active label in the label binding table to form a set of version information records, and assigns a unique version number to the set of version information records. The unique version number is written back to the training input packaging structure, and the unique version number is used in subsequent steps to track the sample definition and strategy caliber used in the training process, preventing caliber drift from occurring in the training process. At the same time, the system carries out the delivery channel mapping processing. The delivery channel mapping processing refers to mapping the delivery channel description in the training input packaging structure to the standardized channel enumeration structure, and the standardized channel enumeration structure is maintained by the marketing activity management account book, and records all managed delivery channel descriptions and their internal channel codes in the commercial delivery system. The system queries the corresponding delivery channel description through the uniform activity identification unique code, generates a channel code field by comparing the standardized channel enumeration structure, and adds the channel code field to the training input packaging structure, so that the training input packaging structure can be directly read by the downstream training process and inference process in subsequent steps without the need to access the marketing activity management account book.

[0029] After the activity label binding, version number registration and delivery channel mapping processing are completed, the system arranges the training input package structure into a model training input set structure. The model training input set structure is a direct input set for the evaluation model training and inference module, which contains user behavior feature fields, activity configuration fields, timestamp fields, label binding tables, unique version numbers, channel encoding fields, missing marker section information, and rejection reason registration fields, and maintains the same order as the synchronization timestamp field. The model training input set structure is registered at the end of this step as the output product of step S130 and is directly passed to step S210, in which the evaluation model training and inference module performs model loading, historical version reference registration and inference session initialization processing according to the model training input set structure, generating a preliminary inference result set. The model training input set structure also indirectly affects the strategy update package generation input set, the strategy update package and the visualization report input set in the subsequent main steps, because the version number registration and delivery channel mapping processing are later traced back to the strategy update tracking record and the execution state aggregation record for restoring the strategy source. In summary, step S130 converts the training input package structure into a model training input set structure that can be directly used by the evaluation model training and inference module through activity label binding, version number registration and delivery channel mapping processing, and delivers the model training input set structure to step S210. Understandably, the technical effect of this step is summarized as follows: step S130 fixes the behavior side information, activity side information and collection strategy aperture in the time dimension and version dimension into the model training input set structure, ensuring that the subsequent evaluation model training and inference module has consistent data aperture, complete activity context and traceable version record when reading the model training input set structure in step S210.

[0030] Step S200 at least contains steps S210-S230: S210, obtain the model training input set structure, perform model loading, historical version reference registration and inference session initialization processing of the evaluation model training and inference module, and obtain a preliminary inference result set; In step S210, the system obtains the model training input set structure and performs model loading, historical version reference registration and inference session initialization processing of the evaluation model training and inference module on the model training input set structure in the business marketing activity analysis scene, and outputs a preliminary inference result set. The model training input set structure is output and passed by step S130. The model training input set structure includes user behavior feature fields, activity configuration fields, timestamp fields, label binding tables, unique version numbers, channel coding fields, missing marker segment information and rejection reason registration fields. The user behavior feature fields depict the click frequency distribution, dwell time stage distribution, order action sequence, cancellation action sequence and other behavior patterns of the unique user identification unique code under the specified marketing activity. The activity configuration fields depict the corresponding delivery channel description, reach strategy description, target audience label description, creative version description, budget block description and delivery time window description of the unique activity identification unique code. The timestamp field corresponds to the synchronous timestamp field and its adjacent time range field. The label binding table records the association between activity labels and behavior segments. The unique version number records the version information of the training input packaging structure when step S130 is completed. The channel coding field is the channel coding registered in the standardized channel enumeration structure. The missing marker segment information marks the position of the time interval breakpoint in the sampling process. The rejection reason registration field records the behavior segments and their rejection reasons that are rejected by the noise segment rejection processing. Specifically, the system first enters the model loading stage. The model loading stage refers to the evaluation model training and inference module calling the current effective evaluation model parameters and evaluation model structure, and establishing a callable state in the running memory. The evaluation model training and inference module is a model set trained for marketing activity performance prediction, user group determination and delivery strategy effectiveness determination. The model set can establish the correspondence between behavior patterns and marketing activities based on the user behavior feature fields and the activity configuration fields. The model loading stage reads the unique version number, verifies the unique version number, and the verification content includes the existence of the unique version number in the version registration library, whether the sampling frequency field and the filtering threshold field corresponding to the unique version number are consistent with the sampling frequency field and the filtering threshold field recorded in the model training input set structure, and whether the channel coding field corresponding to the unique version number is in an effective state in the standardized channel enumeration structure. If the above consistency is established, the evaluation model training and inference module will register the unique version number to the running state context to form the version locking information of this inference session. If the consistency is not established, the system will write the inconsistency into the exception registration queue, record the inconsistency reason, the corresponding unique user identification unique code and the corresponding unique activity identification unique code, and not perform inference on this model training input set structure, wait for subsequent manual review or offline review, do not count this record into the inference session initialization processing, and avoid the appearance of unknown input in the generation of the preliminary inference result set.

[0031] After the model loading stage is completed, step S210 continues to perform historical version reference registration. The historical version reference registration refers to, before inference begins, the evaluation model training and inference module querying the historical training record and the historical evaluation record associated with the current unique version number and binding these records with the current inference session. The historical training record refers to the archived training round summary information generated for the same activity label, the same channel code field, and the similar target audience label description. The training round summary information can reflect the behavior pattern distribution characteristics of similar marketing activities in the past period. The historical evaluation record refers to the effect archive information generated and stored for the same unified activity identification unique code and the same delivery channel description. The effect archive information records the delivery response form and the audience reach form of similar activities in the commercial delivery process. During the historical version reference registration process, the system binds the above-mentioned historical training record and the historical evaluation record with the current unique version number in both directions and registers them in the current inference session metadata area. The inference session metadata area is maintained by the evaluation model training and inference module and is used for subsequent explanation output source. Through this binding, the inference session under the same unique version number can trace back to the source activity label, the source channel code field, and the source target audience label description. After completing the historical version reference registration, the system starts the inference session initialization process. The inference session initialization process refers to the evaluation model training and inference module sequentially reading the user behavior feature field, the activity configuration field, the timestamp field, and the activity label in the label binding table in the model training input set structure and loading these fields as inference input items. For the time sequence segment with missing marker section information, the system does not perform supplementary derivation on the missing marker section information, but instead writes the missing state identifier in the inference input item to inform the evaluation model training and inference module that the section does not participate in behavior pattern reconstruction. For the segment whose elimination reason registration field is recorded as noise segment, the system no longer includes it in the inference input item. Thus, the inference input item obtained through the inference session initialization process consists of a set of behavior segments that can be directly processed by the evaluation model training and inference module, are arranged in order of the synchronous timestamp field, are free of duplication, and have version locking information. After the inference session initialization process is completed, the evaluation model training and inference module runs the inference logic for each combination of the unified user identification unique code and the unified activity identification unique code, calculates the real-time interaction performance, the subsequent conversion tendency, the churn tendency, the repeated reach sensitivity, and other inference indicators of the combination under the current marketing activity, and packs the inference indicators with the unique version number, the channel code field, the target audience label description, and the activity label to form an output entry for the combination. After all the output entries are combined, the preliminary inference result set is formed.The preliminary inference result set is recorded as the output field name of step S210, and is called in subsequent step S220 to extract the activity performance indicator set, the user cluster score set and the model version record. Meanwhile, the preliminary inference result set is also an important source of the activity performance indicator set for the subsequent step S410 to build the visualization report input set.

[0032] S220, extracting the activity performance indicator set, the user cluster score set and the model version record from the preliminary inference result set, performing indicator normalization processing, cluster label solidification and version disk registration, and generating a model evaluation result structure; In step S220, the system extracts the activity performance indicator set, the user cluster score set and the model version record from the preliminary inference result set output by step S210, performs indicator normalization processing, cluster label solidification and version disk registration around the above extraction results, and finally generates a model evaluation result structure. The preliminary inference result set is a set of inference output entries generated by step S210 after completing model loading, historical version reference registration and inference session initialization processing. In the set of inference output entries, for each combination of a unique user identification code and a unique activity identification code, the inference indicators such as real-time interaction performance, subsequent conversion tendency, churn tendency and repeated reach sensitivity are recorded, and the unique version number, channel code field, target audience label description and activity label are also recorded. The activity performance indicator set refers to the set mapping of the indicators directly related to the overall performance of the marketing activity in the above inference indicators, which is aggregated according to the unique activity identification code. This set mapping can reflect the overall reach behavior, continuous interaction behavior and transaction-related behavior of a marketing activity under a specified time window description. The user cluster score set refers to the cluster expression result of the individualized performance related to the user behavior characteristic field in the above inference indicators, which is aggregated according to the unique user identification code. The cluster expression result groups the unique user identification code into different target audience label descriptions or similar characteristic segments, which is used to describe the audience attribute, sensitivity and response form of the unique user identification code under the current marketing activity. The model version record refers to the structured information of the unique version number, channel code field, target audience label description, activity label and delivery time window description recorded for each inference output entry. The structured information is used to determine the source of the strategy scope and the training scope corresponding to the inference output entry in subsequent steps.

[0033] Specifically, the system first performs index normalization processing. The index normalization processing refers to arranging different inference indicators in the campaign performance indicator set in the same dimension description manner to form field values that can be compared horizontally, preventing different marketing campaigns from producing inconsistent field meanings due to differences in channel coding fields, differences in delivery time window descriptions, or differences in target audience label descriptions. When performing index normalization processing, the system reads the instant interaction performance, continuous interaction behavior, and transaction-related behavior recorded in the campaign performance indicator set one by one, and refers to the channel coding field and delivery time window description in the model version record. The same channel coding field and the same delivery time window description are grouped to perform the same caliber mapping for these inference indicators. For entries that cannot be directly established by the same caliber mapping, the system compares the corresponding unique version number in the entry with the historical training record. If there is a similar channel coding field and a similar delivery time window description in the historical training record, the mapping caliber in the historical training record is referenced for correction. If there is no historical training record, the entry is marked as a to-be-confirmed entry and written to the version landing registration queue, which will be landed in subsequent steps. Through index normalization processing, each inference indicator in the campaign performance indicator set is rewritten into a unified field caliber and maintains a corresponding relationship with the unique version number.

[0034] After the index normalization processing is completed, the system performs group label solidification. The group label solidification refers to generating stable group labels for the group attribution of each unique user identification code in the user group scoring set and establishing a corresponding relationship between the stable group label, the target audience label description, and the campaign label. In the group label solidification process, the system reads the group expression results related to the unique user identification code in the user group scoring set. The group expression results record the audience attributes, sensitivity, and response patterns of the unique user identification code under the current marketing campaign. Based on these group expression results, the system determines that the unique user identification code belongs to different audience categories such as the high-response group, the low-response group, or the observation group, and registers the audience category as a stable group label. The stable group label is added to the record entry of the user group scoring set to form a solidified version of the user group scoring set. The system also checks the corresponding relationship between the stable group label and the campaign label, the channel coding field, and the target audience label description. If the same stable group label is referenced by multiple campaign labels and the reach strategy descriptions of these campaign labels are different, the system registers a strategy conflict prompt information in the model version record. The strategy conflict prompt information will be used in the subsequent bid adjustment calculation, channel distribution adjustment calculation, and audience segmentation adjustment calculation processes in step S310 to avoid repeated and inconsistent strategy instructions for the same stable group label in the strategy update package.

[0035] Subsequently, the system performs version landing registration. The version landing registration refers to writing the set of activity performance indicators after index normalization and cluster tag fixation, the set of user cluster scores, and the unique version number, channel code field, target audience tag description, activity tag, and delivery time window description corresponding to the above sets into the long-term archival storage area. The long-term archival storage area is the archival unit of the evaluation model training and inference module, used to maintain version tracking capability. The version landing registration also writes the aforementioned to-be-confirmed items and strategy conflict prompt information into the long-term archival storage area, forming a complete and traceable record. After completing the above processing, the system integrates the set of activity performance indicators, the set of user cluster scores, and the model version record to generate a structured output, which is referred to as the model evaluation result structure in this implementation process. The model evaluation result structure is the output field name of step S220, and the model evaluation result structure is directly passed to step S230 for subsequent activity priority grouping, low-performing activity list extraction, and high-performing activity list extraction processing; at the same time, the model evaluation result structure is also referenced in step S410 to generate the set of activity performance indicators and execution status aggregation content in the visualization report input set.

[0036] S230, performing activity priority grouping, low-performing activity list extraction, and high-performing activity list extraction processing on the model evaluation result structure to generate a strategy update package generation input set; In step S230, the system performs active priority grouping, low-performing activity list extraction and high-performing activity list extraction processing on the model evaluation result structure output in step S220 to generate a strategy update package generation input set. The model evaluation result structure is a structured output formed via step S220, containing an activity performance indicator set, a user sub-group score set and a model version record, wherein the activity performance indicator set has undergone indicator normalization processing, the user sub-group score set has undergone sub-group label solidification, and the model version record has completed version log registration. Active priority grouping refers to the system generating a priority bit sequence representing the processing priority of each marketing activity record in the model evaluation result structure in the current period, in combination with the instant interaction performance, continuous interaction behavior and transaction-related behavior corresponding to the marketing activity in the activity performance indicator set, in combination with the stable sub-group label and response pattern corresponding to the marketing activity in the user sub-group score set, and in combination with the channel code field, the delivery time window description and the target audience label description in the model version record. The priority bit sequence is calculated by the system through a set of sorting rules, which assesses the resource sensitivity, sub-group concentration and channel resource allocation urgency of the business marketing activity within the current delivery time window description. In the active priority grouping process, the system writes the priority bit sequence into the priority field of each marketing activity record to form an activity grouping table with a priority field. The activity grouping table not only saves the priority field, but also associates the activity label, the channel code field and the stable sub-group label, and the activity grouping table will be directly used in subsequent steps.

[0037] After the activity priority grouping is completed, the system extracts a low performance activity list from the model evaluation result structure, the low performance activity list refers to a set of marketing activities that are determined to need strategic intervention after the activity priority grouping. Specifically, the system retrieves the activity grouping table, picks out the marketing activity records with the priority field in the intervention section, records the unique code of the unified activity identification, the channel code field, the target audience label description, the delivery time window description, the activity label, and the stable cluster label of the marketing activity, and reads the instant interaction performance, the continuous interaction behavior, and the transaction-related behavior in the activity performance indicator set corresponding to the marketing activity. The system combines these items into low performance activity list items. The strategy conflict prompt information is also appended to the low performance activity list items to indicate whether the marketing activity has the same stable cluster label as other marketing activities but adopts different touch strategies. At the same time, the system extracts a high performance activity list from the same model evaluation result structure, the high performance activity list refers to a set of marketing activities that are determined to be expanded or maintained in the delivery posture after the activity priority grouping. The system retrieves the activity grouping table, picks out the marketing activity records with the priority field in the maintenance section or the expansion section, and records the unique code of the unified activity identification, the channel code field, the target audience label description, the delivery time window description, the activity label, and the stable cluster label of the marketing activity, and reads the response mode in the user cluster score set corresponding to the marketing activity to describe which type of stable cluster label the marketing activity performs as a continuous interaction behavior or a transaction-related behavior more obvious.

[0038] After the low performance activity list extraction and the high performance activity list extraction are completed, the system merges the activity grouping table, the low performance activity list entries and the high performance activity list entries to form a strategy update package generation input set. The strategy update package generation input set is registered as an output field name of step S230 in the present implementation process, and the strategy update package generation input set is directly read by step S310. After obtaining the strategy update package generation input set, step S310 performs bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation processing for each low performance activity list entry and high performance activity list entry respectively to generate a strategy update package. The strategy update package generation input set will also be indirectly referenced by the execution back record link returned by subsequent step S320, step S330, because the strategy update package needs to be issued to the delivery execution end, and the execution state registration and return channel binding processing of the delivery execution end form an execution back record, and the execution back record is finally used for the collection constraint configuration update of step S330. The strategy update package generation input set will also be used as the upstream data source of the visual report input set in the multi-dimensional analysis index aggregation, strategy change record aggregation and execution state aggregation processing of step S410. Understandably, the technical effect of the present step is summarized as follows: step S230 integrates the activity performance index set, the user group score set and the model version record in the model evaluation result structure into a data entry that can directly drive strategy updating, and outputs the strategy update package generation input set that can be directly used by the bid adjustment calculation, the channel distribution adjustment calculation and the audience segmentation adjustment calculation processing of step S310 through the activity priority grouping, the low performance activity list extraction and the high performance activity list extraction processing, so that the generation of the subsequent strategy update package, the generation of the execution back record and the update of the collection constraint configuration have continuous field sources and clear version tracking relationship.

[0039] Step S300 at least includes steps S310-S330: S310, obtain the strategy update package generation input set, and perform bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation processing to obtain a strategy update package; In step S310, the system acquires the strategy update package generation input set, sequentially performs bid adjustment calculation, channel allocation adjustment calculation and audience segmentation adjustment calculation processing on the strategy update package generation input set, and outputs the strategy update package. The strategy update package generation input set is output by the previous step S230, and the strategy update package generation input set includes the activity grouping table, the low-performing activity list item and the high-performing activity list item. The activity grouping table records the priority field of each marketing activity in the current delivery time window description, and associates the priority field with the activity label, the channel code field, the target audience label description, the stable group label, the unique code of the unified activity identification and the delivery time window description; the low-performing activity list item refers to a set of marketing activities that are determined by the activity priority grouping to need strategy intervention, and carries the unique code of the unified activity identification, the channel code field, the target audience label description, the delivery time window description, the activity label, the stable group label and the set of activity performance indicators related to the marketing activity, including the instant interaction performance, the continuous interaction behavior and the transaction-related behavior; the high-performing activity list item refers to a set of marketing activities that are determined by the activity priority grouping to maintain or increase the delivery intensity, and is also bound with the unique code of the unified activity identification, the channel code field, the target audience label description, the delivery time window description, the activity label and the stable group label. The strategy update package generation input set also includes the strategy conflict prompt information generated in the version landing registration process by step S220, which records the case that multiple marketing activities simultaneously act on the same stable group label but have different reach strategies, and is used to handle resource conflicts in the calculation stage.

[0040] Specifically, the system first performs bid adjustment calculation on the low-performing campaign list entries in step S310. The bid adjustment calculation refers to generating a bid instruction matched with the marketing campaign based on the unique code of the uniform activity identifier recorded in each low-performing campaign list entry, in combination with the instant interaction performance, continuous interaction behavior and transaction-related behavior in the set of activity performance indicators corresponding to the entry, without breaking the budget allocation interval and the delivery time window description of the entry. The budget allocation interval refers to the budget block description registered in the marketing campaign data, which can be understood as the fee allocation range allowed for the marketing campaign within the current delivery time window description; the bid instruction refers to the adjustment mode of the single touch bid behavior of the marketing campaign under the corresponding channel code field. The system reads the strategy conflict prompt information in the bid adjustment calculation process, and if the same stable sub-group label has been reached by other marketing campaigns with a higher intensity and the reach strategy description is different, the system sets a reduction mark for the bid instruction of the marketing campaign. The reduction mark is a resource dilution mark used to prompt the subsequent steps, and the subsequent step S320 will write the reduction mark into the budget allocation instruction set and deliver it to the delivery execution end. Subsequently, the system performs channel allocation adjustment calculation on the high-performing campaign list entries in step S310. The channel allocation adjustment calculation refers to determining whether the marketing campaign is suitable for being expanded to other channel code fields or similar channel code fields under the same target audience label description according to the channel code field, target audience label description and stable sub-group label in the high-performing campaign list entry. For example, when the same stable sub-group label shows continuous interaction behavior and stable transaction-related behavior under multiple channel code fields, the system can generate a channel expansion instruction in the channel allocation adjustment calculation, which is used to indicate that the marketing campaign can be allocated to more channel code fields or increase the exposure frequency in the existing channel code fields. For stable sub-group labels with strategy conflict prompt information, the system marks the channel expansion instructions of these marketing campaigns with a restriction mark in the channel allocation adjustment calculation, which is used to record the restriction reason in the execution state registration of the subsequent step S320, and generate a visual display entry in the strategy change record aggregation of the subsequent step S410. Then, the system performs audience segmentation adjustment calculation on the campaign grouping table in step S310. The audience segmentation adjustment calculation refers to re-splitting the stable sub-group label to combine the uniform user identifier unique codes with similar response patterns under one stable sub-group label into a sub-group description. The sub-group description is a subset of the delivery audience description, which is used to guide the subsequent delivery execution end to only reach the set of uniform user identifier unique codes corresponding to the sub-group description in the delivery process. The system binds the sub-group description with the specific time period, channel code field and activity label in the audience segmentation adjustment calculation, forming an audience instruction.After the bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation are completed, the system collates the bid instruction, channel expansion instruction, limit identifier, reduction mark, sub-group description and audience instruction to generate a systematic delivery instruction set for the delivery execution end. The systematic delivery instruction set is referred to as a strategy update package in this embodiment. The strategy update package is the output field name of step S310, and the strategy update package is directly used as an input in subsequent step S320 to extract the delivery instruction set, budget allocation instruction set and audience instruction set and carry out execution end delivery. At the same time, the strategy update package will be used as source information of the strategy change record in the subsequent step S410 multi-dimensional analysis index aggregation, strategy change record aggregation and execution state aggregation processing, and will be incorporated into the visualization report input set and finally displayed as a strategy change record entry in the multi-dimensional analysis report structure.

[0041] S320, extracting the delivery instruction set, budget allocation instruction set and audience instruction set from the strategy update package, performing execution end delivery, execution state registration and return channel binding processing, and generating an execution return record; In step S320, the system extracts the delivery instruction set, budget allocation instruction set and audience instruction set from the strategy update package output by step S310, performs execution end delivery, execution state registration and return channel binding processing on the above instruction sets, and outputs an execution return record. The strategy update package is the output field name of step S310, and the strategy update package contains instruction elements such as bid instruction, channel expansion instruction, limit identifier, reduction mark, sub-group description and audience instruction. The delivery instruction set is a delivery action description uniquely coded for each unified activity identifier. The delivery action description describes whether the marketing activity needs to maintain the current reach, improve the reach or shrink the reach under a certain channel code field. The budget allocation instruction set is a budget reallocation action description uniquely coded for each unified activity identifier. The budget reallocation action description describes whether the budget block description of the marketing activity within a specific delivery time window description needs to be reallocated between different channel code fields or needs to be staged within the same channel code field. The audience instruction set is an audience screening action description given for each stable group label and its sub-group description. The audience screening action description describes which set of unified user identifiers should be reached in the marketing activity to be delivered, and whether these sets are limited to a certain time period, a certain channel code field or a certain target population label description.

[0042] Specifically, the system groups the delivery instruction set, the budget allocation instruction set and the audience instruction set according to the unique coding of the unified activity identifier in the execution end delivery process, and formats each group of instructions into a delivery execution request recognizable by the delivery execution end. The delivery execution end is an execution system facing real marketing delivery channels, and the delivery execution end is responsible for actually applying the bidding instruction, channel expansion instruction and audience instruction on various delivery channels, ad positions or distribution networks. When the system generates the delivery execution request, it will also carry the limit identifier and the reduction mark from step S310 to indicate that the delivery execution end should not overreach the stable group label involved in the strategy conflict prompt information when processing the stable group label. After the delivery execution request is issued, the system starts the execution state registration. The execution state registration refers to the system reading the delivery execution feedback from the delivery execution end. The delivery execution feedback is the processing state description returned by the delivery execution end after accepting the delivery execution request. The processing state description includes the request receiving state, the applied state, the partially applied state, the rejection reason state and the scheduling state. The system binds the processing state description with the corresponding unique coding of the unified activity identifier, the channel code field, the target audience label description, the stable group label, the sub-group description and the delivery time window description to form an execution state entry. If the delivery execution end feeds back the partially applied state or the rejection reason state for a certain budget allocation instruction set, the system will record the corresponding budget block description in the state exception field. The state exception field will be referenced in the sampling frequency update calculation, the filtering threshold update calculation and the acquisition end pulling rule update processing in the subsequent step S330 to determine whether the sampling frequency of the channel code field needs to be increased or the delivery behavior associated with the budget block description needs to be set with a stricter filtering threshold in the next round of sampling.

[0043] After completing the execution status registration, step S320 performs return channel binding processing. Return channel binding processing means that the system assigns a return path to each execution status entry. The return path specifies which acquisition channel should be used to write back subsequent real-time feedback. The acquisition channel can be any one or more of the following: the browser-side embedded data acquisition program, transaction service logs, or customer service data entry channels. The specific selection is related to the unified activity identifier unique code, channel code field, and delivery time window description. Return channel binding processing writes this return path into the execution status entry, forming a data return pointer for subsequent monitoring cycles. At the end of step S320, the system aggregates all execution status entries to generate an execution return record. The execution return record is the output field name of step S320. The execution return record will be directly called by step S330 for sampling frequency update calculation, filter threshold update calculation, and acquisition end pull rule update processing. Meanwhile, the execution feedback records will also be referenced in the execution status aggregation step of the subsequent step S410 as the execution status source field of the visualization report input set, so as to present the execution status, budget application status and audience reach status of each unified activity identifier unique code in the multi-dimensional analysis report structure.

[0044] S330. Perform sampling frequency update calculation, filtering threshold update calculation, and acquisition end pull rule update processing on the execution feedback record to generate acquisition constraint configuration; In step S330, the system performs sampling frequency update calculation, filtering threshold update calculation, and collection end pulling rule update processing on the execution back record output in step S320, and outputs the collection constraint configuration. The execution back record is the output field name of step S320, and the execution back record is composed of multiple execution state entries, each execution state entry containing a binding uniform activity identification unique code, a channel code field, a target crowd label description, a stable group label, a sub-group description, a delivery time window description, a request receiving state, an applied state, a partial application state, a rejection reason state, a scheduling state, a state exception field, and a return channel binding processing generated backhaul path. The sampling frequency update calculation refers to the system automatically adjusting the collection frequency field in the execution state entry according to the request receiving state, the applied state, and the partial application state. The collection frequency field is part of the collection constraint configuration and is used to limit the minimum sampling interval of subsequent behavior collection under the same uniform user identification unique code and the same uniform activity identification unique code combination. Specifically, when the execution state entry shows that the delivery instruction set of a certain marketing activity under a certain channel code field is completely executed and feedbacks the applied state, the system determines that the marketing activity is in a high attention state within the next delivery time window description, which indicates that the marketing activity needs more intensive behavior feedback collection. Accordingly, the system shortens the collection frequency field interval corresponding to the uniform activity identification unique code and the channel code field in the sampling frequency update calculation. Conversely, when the execution state entry shows that a certain budget allocation instruction set is marked as a scheduling state or a partial application state, and the state exception field indicates that the budget block description has a resource limitation, the system extends the collection frequency field interval of the uniform activity identification unique code under the channel code field in the sampling frequency update calculation, avoiding a large number of redundant repeated behavior reports in the subsequent collection phase. For the execution state entry marked by the rejection reason state, the system establishes a binding relationship between the rejection reason state and the target crowd label description and the stable group label, and records it as a sampling frequency exception rule, which will be directly written into the collection constraint configuration in the next round of collection, preventing invalid high-frequency collection for the target crowd label description.

[0045] After the sampling frequency update calculation is completed, the system performs a filtering threshold update calculation on the same copy of the execution back record. The filtering threshold update calculation refers to the system reading the state exception field, the drop flag, the restriction identifier, and the backhaul path in the execution status entry, jointly analyzing these entries, and updating the filtering threshold field. The filtering threshold field belongs to part of the collection constraint configuration, and describes which records in the subsequent collection behavior segment should be regarded as noise segments and directly removed in the noise segment removal process in step S120. Specifically, when a certain execution status entry carries a drop flag, it means that the marketing activity has performed a resource contraction action on a certain stable segment label. In the filtering threshold update calculation, the system increases the filtering threshold field of the stable segment label under the same channel code field in the subsequent collection, so that the extremely short dwell time record, the repeated click record, or the transaction-related behavior inconsistent with the target audience label description collected in the subsequent collection is more likely to be determined as a noise segment and removed. On the other hand, when a certain execution status entry carries a restriction identifier and indicates that the restriction identifier comes from the policy conflict prompt information, the system identifies the conflict scenario that the restriction identifier points to multiple marketing activities acting on the same stable segment label at the same time. In the filtering threshold update calculation, the system reduces the filtering threshold field of these stable segment labels, so as to more completely retain the cross-response behavior of these stable segment labels between different marketing activities in the subsequent collection, and prevent important interaction behaviors from being filtered out and unable to be identified in the next round of policy adjustment. The result of the filtering threshold update calculation is immediately written into a set of filtering threshold fields to be effective, which will be merged into the collection constraint configuration at the end of the current step.

[0046] Subsequently, the system performs a collection end pulling rule updating process based on the execution feedback record output in step S320. The collection end pulling rule updating process refers to dynamically adjusting the pulling priority, feedback path, and access frequency band of three types of collection channels, namely, a browser end embedded point collection program, a transaction service log, and a customer service supplement recording channel, without modifying the behavior collection channel hardware capability. The browser end embedded point collection program refers to a collection channel that reports interaction actions in a user browsing interface according to a unique user identifier code. The transaction service log refers to a collection channel that records completed transaction related behaviors, cancellation actions, and after-sales actions in a transaction processing service according to a unique activity identifier code. The customer service supplement recording channel refers to a collection channel that supplements abnormal interactions or supplements missing completed transaction records in a human intervention scenario. The system analyzes the feedback path field in the execution feedback record to determine which type of collection channel the subsequent behavior feedback of each execution status entry should flow into, and rearranges the pulling channel priority field of the collection channel in the collection end pulling rule updating process. If an execution status entry shows an applied state and the corresponding feedback path points to the browser end embedded point collection program, the system improves the priority of the browser end embedded point collection program in the pulling channel priority field, so that the browser end embedded point collection program is preferentially scheduled for access in the next collection period. If an execution status entry shows a rejection reason state and the feedback path points to the transaction service log, the system will lower the priority of the transaction service log in the pulling channel priority field, and record the rejection reason state to the pulling channel shielding table. The pulling channel shielding table will indicate that the collection range of the transaction service log needs to be reduced in the next collection, in order to avoid collecting a large number of records that cannot be accepted by the execution end. If an execution status entry shows a partial application state and the feedback path points to the customer service supplement recording channel, the system will keep the priority of the customer service supplement recording channel unchanged in the collection end pulling rule updating process, but will record the supplement time period interval of the customer service supplement recording channel in the collection constraint configuration. The supplement time period interval refers to the upper and lower limits of the time period during which the customer service supplement recording channel can supplement, and is used to limit the backtracking length of the customer service supplement recording channel.

[0047] After the sampling frequency update calculation, filter threshold update calculation and collection end pull rule update processing are completed, the system will update the sampling frequency field, filter threshold field and pull channel priority field, together with the backfill time interval, sampling frequency exception rule and pull channel shielding table, to form a new version of the collection constraint configuration. The collection constraint configuration is the output field name of step S330, and the collection constraint configuration is directly passed to step S410 of the next main step. After obtaining the collection constraint configuration, step S410 carries out multi-dimensional analysis index aggregation, policy change record aggregation and execution state aggregation processing to obtain a visual report input set, and finally displays the execution state and policy change record when generating a multi-dimensional analysis report structure. At the same time, the collection constraint configuration will be written back to step S110. When obtaining user behavior data, marketing activity data and collection constraint configuration, step S110 will use the collection constraint configuration as the current effective version of the sampling frequency field, filter threshold field and pull channel priority field to re-drive the unified time reference binding, user identifier binding and activity identifier binding processing, so that the subsequent generation of the aligned sampling data set and the training input packaging structure is consistent with the latest delivery execution state and pull strategy. Understandably, the technical effect of this step is summarized as follows: step S330 dynamically generates the collection constraint configuration based on the execution of the back transmission record, and through the sampling frequency update calculation, filter threshold update calculation and collection end pull rule update processing, the execution end issued result and the delivery execution feedback are solidified into the pre-constraint of the next round of collection and training, so that the strategy update package in step S310, the strategy update package in step S320, and the collection constraint configuration in step S330 form a continuous closed loop.

[0048] Step S400 at least contains steps S410-S430: S410, obtain the collection constraint configuration, and perform multi-dimensional analysis index aggregation, policy change record aggregation and execution state aggregation processing to obtain a visual report input set; In step S410, the system acquires the collection constraint configuration and performs multi-dimensional analysis index aggregation, policy change record aggregation and execution state aggregation processing on the collection constraint configuration in the business marketing activity analysis scene, and outputs a visual report input set. The collection constraint configuration is output and written back by step S330, which includes the collection frequency field, the filtering threshold field, the pull channel priority field, the backfill time interval, the sampling frequency exception rule and the pull channel shielding table. The collection frequency field records the time interval of the behavior collection under the specific unified activity identification unique code and the specific channel code field; the filtering threshold field records the exclusion criterion of noise segments such as extremely short stay time records and repeated click records in the subsequent collection stage; the pull channel priority field records the scheduling priority order of the three types of collection channels of the browser embedded point collection program, the transaction service log and the customer service supplement recording channel; the backfill time interval records the upper and lower limits of the time period that the customer service supplement recording channel can backfill; the sampling frequency exception rule records the special collection frequency strategy for specific target person group tag description due to rejection reason state or resource limitation; the pull channel shielding table records the channel segment that is limited to collect in a certain delivery time window description. Step S410 first matches the collection constraint configuration with the execution feedback record output by step S320, and the execution feedback record contains request receiving state, applied state, partial application state, rejection reason state and scheduling state, and also contains unified activity identification unique code, channel code field, target person group tag description, stable group tag, sub-group description and delivery time window description bound with these states. Through the matching, the system can read the content of the collection constraint configuration and the actual issued execution situation under the same time reference, forming an aggregable basic unit.

[0049] Specifically, the multi-dimension analysis index aggregation processing of step S410 refers to the system around the effectiveness information of the business marketing activity, and aggregates the activity performance index set, the user cluster score set and the execution state entry related to the marketing activity for each unified activity identification unique code. The activity performance index set comes from the model evaluation result structure of step S220, describes the instant interaction performance, continuous interaction behavior and transaction related behavior of the marketing activity under the current delivery time window description; the user cluster score set comes from the stable cluster label and response form generated for the unified user identification unique code in the model evaluation result structure of step S220, records the touch characteristics of the marketing activity on different stable cluster labels; the execution state entry comes from the execution state registration of step S320, indicates the processing state of the delivery instruction set, the budget allocation instruction set and the audience instruction set at the delivery execution end. In the multi-dimension analysis index aggregation processing, the system aggregates the above three types of information according to the unified activity identification unique code, the channel code field and the delivery time window description, and records the aggregation result as an activity index aggregation block. The unique version number and the activity label are written in the activity index aggregation block at the same time, the unique version number comes from the version number registration of step S130, and the activity label comes from the activity label binding of step S130, which is used to indicate which version of the marketing activity definition is referred to in the subsequent display of the activity index aggregation block. Further, step S410 performs strategy change record aggregation processing. The strategy change record aggregation processing refers to the system extracting the bid instruction, the channel expansion instruction, the restriction identifier, the discount mark, the sub-group description and the audience instruction from the strategy update package generated in step S310, and generating a strategy change link entry by comparing these contents with the execution state entry. The strategy change link entry records the marketing activity state before the strategy is issued, the change content when the strategy update package is issued, the processing state fed back in the execution state entry and the possible restriction identifier or discount mark, and embodies the strategy switching process experienced by the marketing activity within the delivery time window description. The strategy change link entry will be included in the visualization aggregation process together with the activity index aggregation block. Subsequently, step S410 performs execution state aggregation processing. The execution state aggregation processing refers to sorting the request receiving state, the applied state, the partially applied state, the rejection reason state and the scheduling state in the time dimension for each execution state entry, and combining the backfill time interval to mark which state needs to be supplemented by the customer service supplement recording channel, which state is blocked by the pulling channel shielding table, and which state is relaxed in the sampling frequency field due to the sampling frequency exception rule. Through the above sorting and marking, the system generates an execution state sequence description for each unified activity identification unique code, which will be used by the downstream display module to present the actual landing situation of the marketing activity at the delivery execution end.After the multi-dimensional analysis index aggregation, the policy change record aggregation and the execution status aggregation are completed, the system will regularize the activity index aggregation block, the policy change link entry and the execution status sequence description, and generate a data set for display. In this embodiment, the data set is referred to as a visualization report input set. The visualization report input set is the output field name of step S410, and the visualization report input set is directly called by step S420 for subsequent chart configuration generation, text summary paragraph generation and time series comparison entry generation processing.

[0050] S420, extract the activity performance index set, the policy change record and the execution status from the visualization report input set, and generate a multi-dimensional analysis report structure through chart configuration generation, text summary paragraph generation and time series comparison entry generation; In step S420, the system extracts the activity performance indicator set, the policy change record and the execution status from the visualization report input set output in step S410, and performs chart configuration generation, text summary paragraph generation and time sequence comparison item generation around the above extraction results, and outputs a multi-dimensional analysis report structure. The visualization report input set contains an activity indicator aggregation block, a policy change link item and an execution status sequence description. The activity performance indicator set carried in the activity indicator aggregation block corresponds to the behavior performance under the specific unified activity identifier, channel code field and delivery time window description. The policy change link item records the process of converting the policy update package generation input set into a policy update package and landing at the delivery execution end. The execution status sequence description depicts the execution continuity in the order of request receiving state, applied state, partial application state, rejection reason state and scheduling state. Step S420 first performs chart configuration generation. Chart configuration generation refers to the system assembling configuration items for visualization display components according to the activity performance indicator set in the activity indicator aggregation block, the user group score set and the unique version number. Visualization display components include trend curve chart, group distribution chart, delivery channel allocation comparison chart and other visual presentation objects. The trend curve chart is used to show the change trajectory of the activity performance indicator set within the delivery time window description. The group distribution chart is used to show the distribution difference of stable group labels and sub-group descriptions under different target population label descriptions. The delivery channel allocation comparison chart is used to show the correspondence between the channel code field and the budget allocation instruction set. When generating the above chart configuration, the system writes the unique version number into the metadata area of each configuration item, which is used to trace back the training and delivery caliber corresponding to the configuration item when viewed later. Since the term User Interface Configuration (UIC) first appears, the system internally records all the parameters of this chart configuration generation as User Interface Configuration (UIC). User Interface Configuration is an interface description set for presenting business marketing activity evaluation results, including chart type, data source field name, display time range, group display strategy and label style.

[0051] Subsequently, step S420 performs text summary paragraph generation. The text summary paragraph generation refers to that the system generates a structured description paragraph of the policy switching behavior uniquely coded in the current delivery time window description based on the policy change link entry of the same unified activity identification. Specifically, the system reads the bidding instruction, channel expansion instruction, limit identification, drop mark, sub-population description and audience instruction in the policy change link entry, sorts these records to present the sequence of policy actions in time order along the delivery time window description; the system simultaneously references the execution state sequence description to mark which of the above policy actions are registered as applied state, which are registered as partially applied state, which are registered as rejection reason state or in-scheduling state by the execution state entry. The system writes the above sorting results into a reading-oriented description paragraph, and associates the description paragraph with the activity label, channel coding field and target population label description to form a policy change record that can be manually interpreted. Since the term Natural Language Summary (NLS) first appears, the system registers this description paragraph as a Natural Language Summary (NLS), which is used to express the policy change link entry and the execution state sequence description in words, reflecting the adjustment process of the business marketing activity on different stable cluster labels and sub-population descriptions. Further, step S420 performs timing comparison entry generation. The timing comparison entry generation refers to that the system divides the delivery time window description into multiple adjacent sections based on the execution state sequence description, and records the application state of the budget allocation instruction set, the delivery state of the audience instruction set and the effective state of the sampling frequency field in each section to obtain a set of timing comparison entries. The timing comparison entry will be used as the time axis type presentation content in the subsequent display structure, for displaying the corresponding relationship between the policy action and the execution state side by side. The system associates the user interface configuration generated by the chart configuration generation, the natural language summary generated by the text summary paragraph generation, and the timing comparison entries generated by the timing comparison entry generation to form a set of data entities that can be displayed externally. This data entity is called a multi-dimensional analysis report structure in this embodiment. The multi-dimensional analysis report structure is the output field name of step S420, which will be read by step S430 for performing policy update tracking record extraction, collecting policy rollback record extraction and activity configuration version record extraction processing, and at the same time, the multi-dimensional analysis report structure will also be used as the final visualization result on the display end of the business marketing activity management system, which can be read by the marketing party or the operator in a human-computer interface manner.

[0052] S430, performing policy update tracking record extraction, collecting policy rollback record extraction and activity configuration version record extraction processing on the multi-dimensional analysis report structure to generate a collection constraint configuration write-back structure; In step S430, the system performs policy update tracking record extraction, collection policy backtracking record extraction and active configuration version record extraction processing on the multi-dimensional analysis report structure output in step S420, and outputs a collection constraint configuration write-back structure. The multi-dimensional analysis report structure is generated in step S420 and contains user interface configuration, natural language summary and time series comparison item. The user interface configuration records the configuration items in the chart configuration generation stage, the natural language summary records the narrative results of the policy change link item and the execution state sequence description, and the time series comparison item records the segment-level evolution of the budget allocation instruction set, the audience instruction set, the effective sampling frequency field and other elements within the delivery time window description. The policy update tracking record extraction refers to extracting the core action of each policy update package from the natural language summary and the time series comparison item, and associating the core action with the corresponding unified activity identifier, channel coding field, stable group label and sub-group description to form a policy update tracking record. The policy update tracking record is a traceable item set in this embodiment, which records which bidding instructions, channel expansion instructions and audience instructions each marketing activity has experienced under a specific delivery time window description, and maps these instructions with the applied state, partial application state, rejection reason state and scheduling state in the execution state item. The mapping result can be used to look back why a stable group label is assigned a specific collection frequency field or filter threshold field in the subsequent collection stage. Subsequently, the system performs collection policy backtracking record extraction. Collection policy backtracking record extraction refers to analyzing the actual execution of the effective sampling frequency field, the filter threshold field and the pull channel priority field in the time series comparison item by segment, corresponding these fields with the corresponding execution state sequence description, and generating collection policy backtracking records. The collection policy backtracking record is used to describe how the collection constraint configuration is applied to specific stable group labels and sub-group descriptions in the real delivery period, which scenarios trigger the sampling frequency exception rule, which scenarios trigger the pull channel shielding table, and which scenarios require the customer service to record behavior fragments in the backfill period interval. Through this processing, the system can restore the effective segment, restricted segment and backfill segment of the current collection policy without re-scanning all original behavior data, and record it as a backtracking item.

[0053] Further, step S430 performs an action as an activity configuration version record extraction. The activity configuration version record extraction refers to that the system reads the version related information such as the unique version number, the activity label, the channel coding field, the target audience label description and the delivery time window description from the user interface configuration, splices the information with the policy update tracking record and the collection strategy backtracking record, and generates the activity configuration version record. The activity configuration version record is used to describe that in a certain delivery time window description, a certain marketing activity is delivered with which unique version number, targeted with which target audience label description, and resource allocation is performed with which channel coding field, and meanwhile, it is marked whether the unique version number has experienced the restriction identifier or the reduction mark adjustment. After the system completes the policy update tracking record extraction, the collection strategy backtracking record extraction and the activity configuration version record extraction, the policy update tracking record, the collection strategy backtracking record and the activity configuration version record are merged to generate the collection constraint configuration backwrite structure. The collection constraint configuration backwrite structure is the output field name of step S430, and the collection constraint configuration backwrite structure is directly transmitted to step S110 and used as a new collection constraint configuration source in the stage of obtaining user behavior data, marketing activity data and collection constraint configuration by step S110. That is to say, step S110 will directly read the collection frequency field, the filtering threshold field, the pull channel priority field, the backfill time interval, the sampling frequency exception rule and the pull channel shielding table registered in the collection constraint configuration backwrite structure in subsequent running, and write these fields as the current constraint into the aligned sampling data set in the unified time reference binding, user identification binding and activity identification binding processing, so as to form a stable loop among the subsequent training input packaging structure, the model training input set structure, the preliminary reasoning result set, the model evaluation result structure, the policy update package generation input set, the policy update package and the visualization report input set. Understandably, the technical effect of this step is summarized as follows: step S430 deposits the multi-dimensional analysis report structure into the collection constraint configuration backwrite structure, and directly binds the policy update tracking record, the collection strategy backtracking record and the activity configuration version record with the next round of collection constraint configuration, so that the aligned sampling data set generation action of step S110 and the policy update package generation action of step S310 maintain a backtraceable and closed-loop parameter transmission link.

[0054] Embodiment Two Figure 2 A structural block diagram of an AI-based marketing effect analysis system according to an embodiment of the present application is shown. As shown in the structural block diagram, the structure can include: Figure 2 ​The data access and alignment module 01 is configured to obtain user behavior data, marketing activity data and collection constraint configuration, perform unified time reference binding, user identification binding, activity identification binding, missing segment only marking without value estimation processing, noise segment elimination processing, field dictionary unification processing, activity label binding, version number registration, delivery channel mapping processing, generate model training input set structure, and provide the model training input set structure to the evaluation model training and inference module. Specifically, the data access and alignment module receives user behavior data, marketing activity data and collection constraint configuration as input. The user behavior data records interactive behavior, transaction behavior and after-sales behavior from browsing end embedded point collection programs, transaction service logs and customer service supplementary recording channels, and includes timestamp, source channel, trigger entry, stay duration, order action, cancellation action, return action, forwarding action and other information in the entries. The marketing activity data records delivery channel description, reach strategy description, target audience label description, creative version description, budget block description and delivery time window description, and forms an activity main file record for each marketing activity. The collection constraint configuration records collection frequency field, filter threshold field, pull channel priority field, backfill time interval, sampling frequency exception rule and pull channel shielding table. The data access and alignment module first performs unified time reference binding action to unify the timestamps recorded by different source channels. The original timestamp is corrected with the master clock reference to obtain the corrected timestamp, which is registered as a synchronized timestamp field. Subsequently, the data access and alignment module performs user identification binding action to merge and register account identification, device identification and number identification related to the same natural person or the same terminal device, and generates a unified user identification unique code. At the same time, the activity identification binding action is performed to merge and register different presentation forms of the same marketing activity in different delivery channels, different creative versions and different budget blocks, and generate a unified activity identification unique code. After the above binding is completed, the data access and alignment module performs deduplication and splicing on the original records according to the synchronized timestamp field, the unified user identification unique code and the unified activity identification unique code, forms an aligned sampling segment, and eliminates duplicate entries in the same time slice. After obtaining the aligned sampling segment, the data access and alignment module performs missing segment only marking without value estimation processing on the aligned sampling segment. For the segment with a timestamp gap, no reference value is pushed, but a missing marker and source channel trace information are registered. At the same time, the data access and alignment module performs noise segment elimination processing on the aligned sampling segment. According to the filter threshold field, extremely short stays, repeated clicks, abnormal transactions and behaviors outside the delivery time window description are discarded, and the elimination reason is registered. The data access and alignment module then performs field dictionary unification processing on the retained segment, maps the field names used by the source channel to the unified field dictionary, so that the user behavior data field, the marketing activity data field and the synchronized timestamp field adopt a unified naming rule, and the above fields are arranged into a training input packaging structure.After generating the training input packaging structure, the data access and alignment module performs an activity label binding action to write the activity label in the activity master record into the training input packaging structure; performs a version number registration action to register the collection frequency field, the filtering threshold field, the pull channel priority field, the target audience label description, the activity label and other information corresponding to the training input packaging structure as the version number; and performs a delivery channel mapping processing action to map the delivery channel description corresponding to each record into the channel code field, and to establish a corresponding relationship between the channel code field, the target audience label description and the delivery time window description. After the above processing, the data access and alignment module combines the user behavior data field, the marketing activity data field, the synchronization timestamp field, the activity label, the version number, the channel code field, the missing marker and the rejection reason to form the model training input set structure. The model training input set structure is provided by the data access and alignment module to the evaluation model training and inference module as an output product, and is used as an input for subsequent model loading, historical version reference registration and inference session initialization processing. At the same time, the data access and alignment module retains the corresponding relationship between the version number and the channel code field at this stage, and after the sampling frequency update calculation, the filtering threshold update calculation and the collection end pull rule update processing link of the strategy generation and execution feedback module form the collection constraint configuration, receives the collection constraint configuration update content and refreshes the subsequent collection constraint configuration according to the same, which is used in the next round of input stage.

[0055] The evaluation model training and inference module 02 is configured to perform model loading, historical version reference registration, inference session initialization processing on the model training input set structure, generate a preliminary inference result set, extract an activity performance indicator set, a user cluster score set and a model version record from the preliminary inference result set, perform indicator normalization processing, cluster label solidification, version registration, generate a model evaluation result structure, and provide the model evaluation result structure to the strategy generation and execution feedback module. Specifically, the evaluation model training and inference module takes the model training input set structure as input, where each record contains a unique user identification code, a unique activity identification code, a synchronization timestamp field, an activity label, a channel code field, a target audience label description, a version number, a missing flag and a rejection reason. The evaluation model training and inference module first performs model loading, loads the training parameter record, indicator range record and constraint record corresponding to the version number according to the version number, and registers the version number as the current inference context version. The evaluation model training and inference module then performs historical version reference registration, loads the historical training record and historical evaluation record associated with the version number into the inference context for subsequent explanation of the source range and range difference. Next, the evaluation model training and inference module performs inference session initialization processing on the model training input set structure, groups the inference input items according to the combination order of the unique user identification code and the unique activity identification code, and registers only the missing state when encountering the missing flag, and rejects the introduction of the noise segment item when encountering the rejection reason, until a runnable inference input item sequence is formed. The evaluation model training and inference module performs inference after obtaining the inference input item sequence, obtains inference indicators such as real-time interaction performance, continuous interaction behavior, transaction-related behavior, response tendency, loss tendency and repeated reach sensitivity, and registers them in association with the activity label, channel code field, target audience label description and version number, thereby forming a preliminary inference result set. Subsequently, the evaluation model training and inference module extracts an activity performance indicator set aggregated for the unique activity identification code from the preliminary inference result set, extracts a user cluster score set aggregated for the unique user identification code, and extracts a model version record corresponding to the above sets. For the activity performance indicator set, the evaluation model training and inference module performs indicator normalization processing, refers to the channel code field, target audience label description and delivery time window description, performs uniform range conversion on the indicator values of different behavior sources, and registers the binding relationship between the conversion range and the version number. For the user cluster score set, the evaluation model training and inference module performs cluster label solidification, applies stable cluster labels to the unique user identification code according to the response tendency, loss tendency and repeated reach sensitivity in the user cluster score set, and forms a one-to-many mapping between the stable cluster label and the target audience label description, activity label and channel code field to describe the user identification code's population attribution status under the current marketing activity.For the model version record, the model training and inference module performs a version registration action, writes the active performance indicator set, the user cluster score set, the stable cluster label, the channel code field, the target audience label description, the activity label, the delivery time window description, and the version number into the long-term archiving area, and registers the time sequence marker. After the above processing, the model training and inference module integrates the active performance indicator set, the user cluster score set, and the model version record to output a model evaluation result structure. The model evaluation result structure is provided as an output product to the strategy generation and execution feedback module, serving as an input for activity priority grouping, low-performance activity list extraction, and high-performance activity list extraction; at the same time, the model training and inference module completes the binding registration of the version number and the behavior scope in the version registration action, providing a traceable active performance indicator set source for the visualization report generation module in the multi-dimensional analysis indicator aggregation process.

[0056] The strategy generation and execution feedback module 03 is configured to perform activity priority grouping, low performance activity list extraction and high performance activity list extraction on the model evaluation result structure, generate a strategy update package generation input set, perform bid adjustment calculation, channel distribution adjustment calculation and audience segmentation adjustment calculation after obtaining the strategy update package generation input set, generate a strategy update package, extract a delivery instruction set, a budget allocation instruction set and an audience instruction set from the strategy update package, perform execution end delivery, execution state registration and return channel binding processing, generate an execution feedback record, perform sampling frequency update calculation, filtering threshold update calculation and collection end pull rule update processing according to the execution feedback record, generate collection constraint configuration, and provide the collection constraint configuration to the visualization report generation module and the data access and alignment module, respectively. Specifically, the strategy generation and execution feedback module receives the model evaluation result structure, which includes an activity performance indicator set, a user cluster score set and a model version record. The strategy generation and execution feedback module parses the model evaluation result structure piece by piece, records a priority field for each marketing activity according to the records of the activity performance indicator set and the user cluster score set in the current delivery time window description, and forms an activity priority grouping. The strategy generation and execution feedback module extracts marketing activities in the intervention section from the activity priority grouping, records the unique code of the uniform activity identifier, the channel code field, the target audience label description, the stable cluster label, the delivery time window description and the corresponding activity performance indicator set entry, forms a low performance activity list, and extracts marketing activities in the maintenance section or the release section at the same time, records the corresponding fields, and forms a high performance activity list. The strategy generation and execution feedback module combines the activity priority grouping, the low performance activity list and the high performance activity list into a strategy update package generation input set. After obtaining the strategy update package generation input set, the strategy generation and execution feedback module performs bid adjustment calculation on the low performance activity list, records the bid instruction for each unique code of the uniform activity identifier, performs channel distribution adjustment calculation on the high performance activity list, records the channel expansion instruction and resource allocation direction, and performs audience segmentation adjustment calculation on the stable cluster label involved in the above two lists, records the audience instruction and generates a sub-group description. After the above processing, the strategy generation and execution feedback module packs the bid instruction, the channel expansion instruction and the audience instruction set to form a strategy update package.After receiving the strategy update package, the strategy generation and execution feedback module splits the delivery instruction set, budget allocation instruction set, and audience instruction set according to the unique unified activity identifier code, forming a delivery request, which is then sent to the delivery execution end. The delivery execution end provides feedback on the request reception status, application status, partial application status, rejection reason status, and scheduling status. Based on this feedback, the strategy generation and execution feedback module performs an execution status registration action, recording the unique unified activity identifier code, channel code field, target audience tag description, stable segment tag, audience instruction set, budget allocation instruction set, and corresponding status result in the execution status entry. In the return channel binding processing action, it registers the feedback path for each execution status entry, indicating that subsequent behavioral feedback flows to the browser-side event tracking program, transaction service logs, or customer service supplementary entry channel. The strategy generation and execution feedback module aggregates the above execution status entries to generate execution feedback records. The strategy generation and execution feedback module performs sampling frequency update calculations, filtering threshold update calculations, and data acquisition rule update processing on the execution feedback record to form a data acquisition constraint configuration. This configuration records the sampling frequency field, filtering threshold field, data acquisition channel priority field, backfill time interval, sampling frequency exception rules, and data acquisition channel masking table. The strategy generation and execution feedback module then delivers the data acquisition constraint configuration to the visualization report generation module, which uses it as input for multi-dimensional analysis indicator aggregation, strategy change record aggregation, and execution status aggregation. Simultaneously, it delivers the data acquisition constraint configuration to the data integration and alignment module, serving as the source of data acquisition parameters for the next round of user behavior data, marketing campaign data, and data acquisition constraint configuration. This drives unified time base binding, user identifier binding, and campaign identifier binding actions in subsequent input stages.

[0057] The visualization report generation module 04 is configured to perform multi-dimensional analysis index aggregation, strategy change record aggregation and execution state aggregation processing after obtaining the collection constraint configuration, generate a visualization report input set, extract the activity performance index set, the strategy change record and the execution state from the visualization report input set, perform chart configuration generation, text summary paragraph generation and timing comparison item generation, generate a multi-dimensional analysis report structure, generate a collection constraint configuration write-back structure after performing strategy update tracking record extraction, collection strategy backtracking record extraction and activity configuration version record extraction on the multi-dimensional analysis report structure, and feed back the collection constraint configuration write-back structure to the data access and alignment module. Specifically, the visualization report generation module receives the collection constraint configuration, the execution back record, the activity performance index set, the user group score set, the model version record, the activity priority grouping, the low performance activity list and the high performance activity list. The visualization report generation module first performs multi-dimensional analysis index aggregation processing, arranges the activity performance index set and the user group score set according to the unified activity identifier unique code, the channel code field and the delivery time window description, and attaches the version number in the model version record to form an activity index aggregation block. The visualization report generation module then performs strategy change record aggregation processing, extracts the bid instruction, the channel expansion instruction, the audience instruction, the budget allocation instruction set, the delivery instruction set, the restriction identifier and the reduction mark from the strategy update package generation input set and the strategy update package comparison execution state item, records the corresponding order of the strategy action and the execution state item, and forms the strategy change record. The visualization report generation module further performs execution state aggregation processing, registers the request receiving state, the applied state, the partially applied state, the rejection reason state and the scheduling state in the execution state item in chronological order, synchronizes the order with the backhaul path, the pull channel priority field and the backfill time interval to obtain an execution state sequence description. The visualization report generation module combines the activity index aggregation block, the strategy change record and the execution state sequence description to form the visualization report input set. After obtaining the visualization report input set, the visualization report generation module extracts the activity performance index set, the strategy change record and the execution state therefrom, performs chart configuration generation, and gives a chart configuration for the display end; performs text summary paragraph generation, and writes corresponding text paragraph descriptions of the strategy action and the execution state according to the strategy change record and the execution state sequence description; and performs timing comparison item generation, divides adjacent sections according to the delivery time window description, gives the budget allocation instruction set, the audience instruction set and the collection frequency field state in each section to form multi-section timing comparison items. The visualization report generation module combines the chart configuration, the text paragraph and the timing comparison item to generate a multi-dimensional analysis report structure.The visual report generation module, after obtaining the multi-dimensional analysis report structure, performs a policy update tracking record extraction action on the multi-dimensional analysis report structure, extracts each policy action and its corresponding execution state, and records it as a policy update tracking record; performs a collection policy rollback record extraction action, forms a collection policy rollback record from the actual effective state of the collection frequency field, the filtering threshold field, the pull channel priority field, the backfill time interval, the sampling frequency exception rule, and the pull channel shielding table in each section; performs an activity configuration version record extraction action, and arranges the version number, the activity label, the channel code field, the target audience label description, and the delivery time window description into an activity configuration version record. The visual report generation module combines the policy update tracking record, the collection policy rollback record, and the activity configuration version record to generate a collection constraint configuration write-back structure. The visual report generation module returns the collection constraint configuration write-back structure to the data access and alignment module, which refreshes the collection constraint configuration accordingly. The refresh result takes effect in the next round of user behavior data, marketing activity data, and collection constraint configuration input stages, thereby closing the transmission relationship between the aforementioned data access and alignment module, the evaluation model training and reasoning module, the policy generation and execution feedback module, and the visual report generation module.

Claims

1. A marketing effectiveness analysis method based on AI, characterized in that, include: Acquire user behavior data, marketing campaign data, and collection constraint configurations; bind them to a unified time base and user identifiers and campaign identifiers; perform missing segment labeling without valuation and noise segment removal processing, including the sampling frequency field and filtering threshold field in the collection constraint configuration; and perform unified field dictionary mapping processing to generate the model training input set structure. The evaluation model training, inference, and result normalization processes are performed. Combined with the unique version number, historical version reference registration and group label solidification are performed to generate a model evaluation result structure containing a set of activity performance indicators, a set of user group scores, and model version records. A strategy update package is also constructed to generate the input set. Perform bid adjustment calculations, channel allocation adjustment calculations, and audience segmentation adjustment calculations. Combine strategy conflict prompts to generate a strategy update package with a reduction mark and a restriction mark. Execute the delivery instruction, perform status registration and return channel binding processing, and generate a collection constraint configuration containing status exception fields. The system performs multi-dimensional analysis indicator aggregation, strategy change record aggregation, and execution status aggregation to generate a set of visual report inputs. It also constructs a data collection constraint configuration write-back structure through chart configuration generation, natural language summary generation, and time series comparison item generation. This structure is used to write back the sampling frequency field, filter threshold field, and pull channel priority field.

2. The method according to claim 1, characterized in that, The process of generating the data collection constraint configuration that includes the status exception field also includes: The strategy update package generates an input set, performs bid adjustment calculations for low-performing campaigns, performs channel allocation adjustment calculations for high-performing campaigns, and performs audience segmentation adjustment calculations for stable segment tags in conjunction with strategy conflict prompts, generating a strategy update package containing reduction flags and restriction flags; Based on the strategy update package, the execution terminal is deployed and the execution status registration and return channel binding process containing the unique code of the unified activity identifier is performed, generating an execution return record with an exception status field; Based on the execution feedback records, the sampling frequency update calculation, filtering threshold update calculation, and data acquisition rule update processing are performed. The sampling frequency field, filtering threshold field, and data acquisition channel priority field are dynamically adjusted in combination with the status anomaly field to generate a new version of the data acquisition constraint configuration. Among them, the execution feedback record consists of multiple execution status entries. Each execution status entry includes a unique unified activity identifier, channel code field, target audience tag description, stable subgroup tag, subgroup description, delivery time window description, request reception status, applied status, partially applied status, rejection reason status, scheduled status, status anomaly field, and feedback path generated by the return channel binding processing.

3. The method according to claim 2, characterized in that, The collection frequency field is automatically adjusted based on the request receiving status, applied status, and partial application status in the execution status entry. When the execution status entry shows the applied status, the collection frequency field interval of the corresponding unified activity identifier unique code and channel code field is shortened.

4. The method according to claim 2, characterized in that, When the execution status entry shows that the budget allocation instruction set is marked as scheduled or partially applied and the status exception field indicates that resources are limited, the interval of the collection frequency field under the corresponding unified activity identifier unique code in the channel code field is extended.

5. The method according to claim 2, characterized in that, The rejection reason status specifically includes: For execution status entries marked with rejection reason status, the rejection reason status is bound to the target population label description and stable subgroup label, recorded as a sampling frequency exception rule, and directly written into the collection constraint configuration in the next round of collection.

6. The method according to claim 2, characterized in that, Read the status exception field, reduction flag, limit identifier and return path, jointly analyze the execution status entries, update the filter threshold field, and write the results into the filter threshold field set to be effective.

7. The method according to claim 2, characterized in that, When an execution status entry carries a decrease flag, the system increases the filtering threshold field of the corresponding stable group label in order to remove records with extremely short dwell time, repeated click records, or transaction behaviors that are inconsistent with the target group label description.

8. The method according to claim 2, characterized in that, Based on the analysis of the return path field of the execution return record, it is determined whether the behavior feedback flows into the browser terminal's data collection program, transaction service logs, or customer service supplementary entry channel, and the priority field of the pull channel is rearranged.

9. The method according to claim 2, characterized in that, After completing the calculation of sampling frequency update, filtering threshold update, and acquisition end pull rule update, the updated sampling frequency field, filtering threshold field, pull channel priority field, sampling frequency exception rules, and pull channel masking table are merged together to form a new version of the acquisition constraint configuration.

10. An AI-based marketing effectiveness analysis system, applied to the method described in any one of claims 1-9, characterized in that, include: The data access and alignment module is used to acquire user behavior data, marketing activity data and collection constraint configuration, perform unified time base binding, user identifier binding, activity identifier binding, missing segment marking without evaluation processing, noise segment removal processing, unified field dictionary processing, activity tag binding, version number registration, and distribution channel mapping processing, generate model training input set structure, and provide the model training input set structure to the evaluation model training and inference module; The evaluation model training and inference module is used to perform model loading, historical version reference registration, and inference session initialization processing on the model training input set structure to generate a preliminary inference result set. It is used to extract the activity performance indicator set, user group score set, and model version record from the preliminary inference result set, perform indicator normalization processing, group label solidification, and version disk registration to generate the model evaluation result structure and provide the model evaluation result structure to the strategy generation and execution feedback module. The strategy generation and execution feedback module is used to perform activity priority grouping, low-performance activity list extraction, and high-performance activity list extraction on the model evaluation results structure, generate a strategy update package input set, and perform bid adjustment calculation, channel allocation adjustment calculation, and audience segmentation adjustment calculation after obtaining the strategy update package input set, generate a strategy update package, extract the delivery instruction set, budget allocation instruction set, and audience instruction set from the strategy update package, perform execution terminal distribution, execution status registration, and return channel binding processing, generate execution feedback records, and perform sampling frequency update calculation, filter threshold update calculation, and data acquisition terminal pull rule update processing based on the execution feedback records, generate data acquisition constraint configuration, and provide data acquisition constraint configuration to the visualization report generation module and the data access and alignment module respectively; The visualization report generation module is used to perform multi-dimensional analysis indicator aggregation, strategy change record aggregation, and execution status aggregation after obtaining the data collection constraint configuration, and generate a visualization report input set. It is used to extract the activity performance indicator set, strategy change record, and execution status from the visualization report input set, and to perform chart configuration generation, text summary paragraph generation, and time series comparison item generation to generate a multi-dimensional analysis report structure. After extracting strategy update tracking records, data collection strategy backtracking records, and activity configuration version records from the multi-dimensional analysis report structure, it generates a data collection constraint configuration write-back structure and feeds back the data collection constraint configuration write-back structure to the data access and alignment module.

Images