Advertisement putting resource scheduling management method and system based on big data analysis

Abstract

The invention discloses an advertisement putting resource scheduling management method and system based on big data analysis, and relates to the technical field of data processing, and the method comprises the steps: collecting multi-source advertisement putting data, and carrying out the batch processing and fusion calculation, and obtaining real-time user behavior characteristics; historical conversion rule data are introduced, multi-objective optimization is carried out on the real-time user behavior characteristics in combination with the historical conversion rule data, and a preliminary scheduling strategy is constructed; setting an advertisement putting constraint condition, executing the preliminary scheduling strategy to carry out resource allocation optimization, and formulating a resource scheduling scheme; and performing simulation execution on the resource scheduling scheme to perform effect tracking, obtaining simulation scheduling parameters to perform incremental updating, and constructing a closed-loop optimization report of advertisement putting resource scheduling. The technical problems that in the prior art, advertisement putting resource allocation lacks a dynamic optimization mechanism, and accurate scheduling is difficult to carry out according to real-time data are solved, and the technical effects of improving the advertisement resource allocation accuracy and the putting conversion effect are achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method and system for scheduling and managing advertising resources based on big data analysis. Background Technology

[0002] In the advertising process, resource allocation is usually based on preset rules or periodic statistical data to formulate strategies. The scheduling method is relatively fixed and it is difficult to respond in a timely manner to changes in user behavior and fluctuations in the market environment. When user interests, bidding environment, or channel performance change dynamically, the advertising strategy often cannot be adjusted synchronously, resulting in a mismatch between resource allocation and actual conversion trends, which affects the advertising effect and resource utilization efficiency. Summary of the Invention

[0003] This application provides a method and system for scheduling and managing advertising resources based on big data analysis, which is used to address the technical problems in the existing technology of lacking a dynamic optimization mechanism for the allocation of advertising resources and the difficulty in accurately scheduling them based on real-time data.

[0004] In view of the above problems, this application provides a method and system for scheduling and managing advertising resources based on big data analysis.

[0005] The first aspect of this application provides a method for scheduling and managing advertising delivery resources based on big data analysis, the method comprising: Multi-source advertising data is collected, processed in batches, and fused to obtain real-time user behavior characteristics. Historical conversion pattern data is introduced, and the real-time user behavior characteristics are combined with historical conversion pattern data for multi-objective optimization to construct a preliminary scheduling strategy. Advertising placement constraints are set, and the preliminary scheduling strategy is executed to optimize resource allocation and formulate a resource scheduling plan. The resource scheduling plan is simulated for effect tracking, and simulation scheduling parameters are obtained for incremental updates to construct a closed-loop optimization report for advertising placement resource scheduling.

[0006] A second aspect of this application provides an advertising delivery resource scheduling and management system based on big data analysis, the system comprising: The feature acquisition module is used to collect multi-source advertising data, process and merge it in batches to obtain real-time user behavior features; the multi-objective optimization module is used to introduce historical conversion pattern data, combine the real-time user behavior features with the historical conversion pattern data to perform multi-objective optimization, and construct a preliminary scheduling strategy; the resource allocation optimization module is used to set advertising constraints, execute the preliminary scheduling strategy to optimize resource allocation, and formulate a resource scheduling plan; the effect tracking module is used to simulate the execution of the resource scheduling plan to track the effect, obtain simulation scheduling parameters for incremental updates, and construct a closed-loop optimization report for advertising resource scheduling.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application collects multi-source advertising data, processes and integrates it in batches to obtain real-time user behavior characteristics; it introduces historical conversion pattern data, combines the real-time user behavior characteristics with historical conversion pattern data for multi-objective optimization, and constructs a preliminary scheduling strategy; it sets advertising placement constraints and executes the preliminary scheduling strategy to optimize resource allocation and formulate a resource scheduling plan; it simulates the execution of the resource scheduling plan to track the effect, obtains simulation scheduling parameters for incremental updates, and constructs a closed-loop optimization report for advertising resource scheduling. This invention solves the technical problems of existing technologies where advertising resource allocation lacks a dynamic optimization mechanism and is difficult to accurately schedule based on real-time data. By integrating real-time user behavior characteristics and historical conversion pattern data for multi-objective optimization and constructing a closed-loop scheduling mechanism, it achieves the technical effect of improving the accuracy of advertising resource allocation and the conversion effect of advertising placement. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic diagram of the advertising delivery resource scheduling and management method based on big data analysis provided in the embodiments of this application; Figure 2 This is a schematic diagram of the advertising resource scheduling and management system based on big data analysis provided in an embodiment of this application.

[0010] Figure labeling: Feature acquisition module 11, multi-objective optimization module 12, resource allocation optimization module 13, effect tracking module 14. Detailed Implementation

[0011] This application provides an advertising resource scheduling and management method and system based on big data analysis. It addresses the technical problems in existing technologies where advertising resource allocation lacks a dynamic optimization mechanism and is difficult to accurately schedule based on real-time data. By integrating real-time user behavior characteristics and historical conversion pattern data for multi-objective optimization and constructing a closed-loop scheduling mechanism, it achieves the technical effect of improving the accuracy of advertising resource allocation and the conversion effect of advertising.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides a method for scheduling and managing advertising resources based on big data analysis, the method comprising: Step S100: Collect multi-source advertising data, process and merge it in batches to obtain real-time user behavior characteristics.

[0015] In this embodiment, when collecting multi-source advertising data for batch processing and fusion calculation, firstly, multi-granularity parameters are set based on multiple advertising platforms to construct a multi-granularity sliding window system with fine-grained, medium-granularity, and coarse-grained windows. Click intensity data is generated by calculating clicks in the fine-grained window, browsing interest data is generated by tracking preferences in the medium-granularity window, and activity periods are identified by analyzing activity in the coarse-grained window. Subsequently, the click intensity data, browsing interest data, and activity periods are processed in batches to generate batch datasets containing micro-batch data and macro-batch data. Streaming computation is performed on the micro-batch data to generate real-time user behavior indicators, and batch computation is performed on the macro-batch data to generate historical user behavior pattern indicators.

[0016] Then, the real-time user behavior metrics and the historical user behavior pattern metrics are spliced ​​and fused to form feature fusion data, and the feature fusion data is normalized to finally obtain real-time user behavior features.

[0017] Furthermore, the method provided in the application embodiments, which involves collecting multi-source advertising data, processing and fusing it in batches to obtain real-time user behavior characteristics, also includes: Multiple granularity parameters are set based on multiple advertising platforms, and multiple granular sliding windows are set according to these parameters. These granular sliding windows include fine-grained windows, medium-grained windows, and coarse-grained windows. Click intensity data is obtained by calculating user clicks based on the fine-grained window. Browsing interest data is obtained by tracking user preferences based on the medium-grained window. Activity analysis is performed on users based on the coarse-grained window to identify multiple activity periods. The click intensity data, browsing interest data, and multiple activity periods are processed in batches to generate batch datasets, which include micro-batch data and macro-batch data. Streaming computation is performed on the micro-batch data to generate real-time user behavior indicators, and batch computation is performed on the macro-batch data to generate historical user behavior pattern indicators. The real-time user behavior indicators and historical behavior pattern indicators are then feature-concatenated and fused to obtain feature fusion data. The feature fusion data is then normalized to generate the real-time user behavior features.

[0018] In this embodiment of the application, when collecting multi-source advertising data for batch processing and fusion calculation, multiple granularity parameters are set based on multiple advertising platforms, and multiple granularity sliding windows are set according to the multiple granularity parameters. The multiple granularity sliding windows include fine-grained windows, medium-grained windows, and coarse-grained windows. Specifically, the process begins by integrating exposure, click, browsing, and login events from multiple advertising platforms. Data fields from different platforms are then standardized to ensure a consistent format for event type, user identifier, ad identifier, and timestamp fields. Next, multiple granularity parameters are defined, including window length and sliding step parameters. Event data is then segmented according to the window length parameter based on the timestamp field, and sequentially shifted according to the sliding step parameter to form continuous sliding time intervals. Based on this, fine-grained, medium-grained, and coarse-grained windows are constructed. Fine-grained windows use shorter window lengths to depict short-term behavioral fluctuations, medium-grained windows use medium window lengths to depict phased behavioral trends, and coarse-grained windows use longer window lengths to depict periodic behavioral characteristics. For example, fine-grained windows can be segmented by minute-level time intervals, medium-grained windows by hour-level time intervals, and coarse-grained windows by day-level time intervals, thus creating different time-scale behavioral statistical intervals under the same user identifier dimension.

[0019] Next, click calculations are performed based on fine-grained windows. In this process, firstly, event records with the event type "click" and timestamps within the time interval of each fine-grained window are filtered and grouped according to user identifiers. Then, click events repeatedly reported by the same user identifier under the same ad identifier within a short period are deduplicated, retaining only one valid click record. After deduplication, the valid click records within the current fine-grained window are counted to obtain the click count. Finally, the click count is converted to a unit time based on the window length parameter to obtain click intensity data. This click intensity data represents the frequency of user clicks within a unit of time. For example, if 5 valid click records are obtained within a fine-grained window with a window length parameter of 1 minute, the click intensity data is 5 clicks per minute, used to quantify the degree of real-time user interaction.

[0020] When tracking user preferences using medium-granularity windows, a method based on category frequency statistics is employed. Within each medium-granularity window, records with event types of browsing events and timestamps falling within the window's time interval are selected, and these records are categorized into their corresponding content categories based on the category field of the browsing events. After categorization, the number of views for each content category within the medium-granularity window is counted, and the percentage of views for each category is calculated by dividing the number of views for each category by the total number of browsing events within the window. This yields the interest percentage for each content category, forming browsing interest data. This data reflects the distribution of user attention to different content categories within the medium-granularity window. For example, if a certain category accounts for 70% of the total views within a one-hour medium-granularity window, then the percentage of that category in the browsing interest data is 0.7.

[0021] When performing user activity analysis based on coarse-grained windows, a time-based bucketing statistical method is adopted. Within each coarse-grained window, records with login events as their event type and timestamps falling within that window's time interval are selected, and a day is divided into 24-hour time intervals. Then, records are assigned to their corresponding hourly time intervals based on the login event timestamps, and the number of logins within each hourly time interval is counted. After the statistics are completed, the number of logins in each hourly time interval is compared with a preset activity threshold. The preset activity threshold is set as the average number of logins across all hourly time intervals within the current coarse-grained window plus a standard deviation. Hourly time intervals with more than this preset activity threshold are marked as active periods. For example, in a certain daily coarse-grained window, if the average number of logins across all hourly time intervals is 20 and the standard deviation is 5, then the preset activity threshold is 25. Hourly time intervals with more than 25 logins are identified as active periods.

[0022] When batch processing click intensity data, browsing interest data, and multiple activity periods, a batch processing division method based on fixed time triggering is adopted. The micro-batch triggering cycle is set to 5 minutes, and the click intensity data and browsing interest data generated within each 5-minute period are aggregated to form a micro-batch data. The macro-batch triggering cycle is set to 1 day, and the activity periods obtained from daily statistics are aggregated to form a macro-batch data. Thus, multiple micro-batch data and multiple macro-batch data are formed in the time dimension, constituting a batch dataset.

[0023] When performing streaming computation on micro-batch data, a sliding accumulation-based streaming aggregation method is adopted. The micro-batch data is read batch by batch according to the time order, and the user identifier is used as the grouping key for accumulation calculation. The click intensity data in the current micro-batch is accumulated into the current real-time statistics. At the same time, the real-time interest value of the corresponding category is updated with the browsing interest ratio in the current micro-batch. After the calculation is completed, the real-time user behavior index is output so that the real-time user behavior index reflects the behavior status in the most recent micro-batch periods.

[0024] When performing batch calculations on macro-batch data, a batch calculation method based on frequency ratio threshold determination is adopted. Within 30 consecutive macro-batch cycles, the number of times each hourly time interval is marked as an active period is counted, and the occurrence ratio is calculated with 30 as the denominator. The long-term activity determination threshold is set to 0.6. When the occurrence ratio of a certain hourly time interval is greater than or equal to 0.6, that time interval is determined as a long-term active time interval. At the same time, the arithmetic mean of the click intensity data within 30 consecutive macro-batch cycles is calculated to obtain the long-term click intensity level. Finally, the long-term active time interval and the long-term click intensity level are combined to form a user historical behavior pattern indicator, which is used to characterize the stable behavioral characteristics of users over a long period of time.

[0025] Subsequently, real-time user behavior metrics and historical behavior pattern metrics are fused together. This process begins by traversing multiple users for classification and identification, allocating storage space for each user according to their identifiers, and constructing a feature fusion mapping table to establish index relationships and field correspondences between real-time and historical features. Next, real-time user behavior metrics are written into the real-time feature regions within the multiple feature storage spaces for feature analysis, forming real-time behavior features. Simultaneously, historical feature regions within the multiple feature storage spaces are retrieved based on historical behavior pattern metrics for feature analysis, forming historical behavior features. Then, the real-time and historical behavior features are dimensionally aligned according to the feature fusion mapping table, and horizontally concatenated based on the alignment results to generate a first fused feature vector. Feature conflict identification and conflict term removal are performed on the first fused feature vector to generate a second fused feature vector. Finally, cross-validation is performed based on the second fused feature vector to output the fused feature data.

[0026] Finally, the feature fusion data is normalized. In this process, firstly, statistical analysis is performed on each dimension of the feature fusion data. Within a preset statistical period, the minimum and maximum values ​​of each dimension are calculated and recorded as normalization boundary parameters. Then, for each dimension, a linear normalization calculation formula based on the minimum and maximum values ​​is used to transform the numerical value. The original feature value is subtracted from the minimum value and then divided by the difference between the maximum and minimum values, so that the normalization result falls within the range of 0 to 1. For example, if the minimum value of a certain dimension of click intensity is 2 and the maximum value is 10 within the statistical period, when the current feature value is 6, the normalization result is calculated to be 0.5. For categorical features, the category field is first converted into a numerical code using an integer encoding method, and then the same linear normalization process is performed based on the minimum and maximum code values ​​of the corresponding code within the statistical period. When the maximum and minimum values ​​of a certain dimension are equal, the normalization result of that dimension is directly set to 0 to avoid division by zero. After all dimensions of features have been normalized, the normalization results of each dimension are arranged in the preset field order to form real-time user behavior features with a uniform scale.

[0027] Furthermore, in the method provided in the application embodiments, the method of performing feature splicing and fusion of the real-time user behavior indicators and the historical behavior pattern indicators to obtain feature fusion data further includes: Multiple users are categorized and identified, and storage is allocated according to these user identifiers to determine multiple feature storage spaces. A feature fusion mapping table is constructed based on these multiple feature storage spaces. The real-time behavior indicators of the users are written into the real-time feature regions of the multiple feature storage spaces for feature analysis to construct real-time behavior features. Based on the historical behavior pattern indicators, the historical feature regions of the multiple feature storage spaces are retrieved for feature analysis to construct historical behavior features. The real-time behavior features and the historical behavior features are dimensionally aligned according to the feature fusion mapping table, and the real-time behavior features and historical behavior features are horizontally concatenated according to the alignment results to generate a first fused feature vector. Feature conflict identification is performed on the first fused feature vector, and conflicting terms are extracted and removed to generate a second fused feature vector. Cross-validation is performed based on the second fused feature vector to generate feature fusion data.

[0028] In this embodiment, when fusing real-time user behavior metrics and historical behavior pattern metrics, multiple users are first traversed and categorized. This involves reading the user identifier field from each user data record, grouping records with the same identifier as the same user, and distinguishing records with different identifiers as different users. Then, storage is allocated to multiple users based on their identifiers. Each user identifier is assigned a unique feature storage space, ensuring all feature data for that user is written to its corresponding space, thus defining multiple feature storage spaces. Each feature storage space is internally divided into a real-time feature region and a historical feature region. The real-time feature region stores real-time user behavior metrics, while the historical feature region stores historical behavior pattern metrics. After completing the storage allocation, a feature fusion mapping table is constructed based on a field-by-field registration method. This table includes real-time field names, historical field names, field sequence numbers, and dimension index numbers. The field sequence number specifies the feature arrangement order, and the dimension index number identifies the correspondence between real-time and historical fields. For example, the click intensity field and the long-term click intensity level field are registered with the same dimension index number for subsequent dimension alignment.

[0029] Subsequently, real-time user behavior metrics are written into the real-time feature regions of multiple feature storage spaces for feature analysis, constructing real-time behavior features. That is, based on the user identifier, the corresponding feature storage space is located, and in the real-time feature region, according to the field sequence number of the real-time side field in the feature fusion mapping table, the metrics such as click intensity, browsing interest ratio, and real-time activity are sequentially filled into the corresponding field positions, and arranged according to the field sequence number to form a fixed-order field sequence. This field sequence is the real-time behavior feature. For example, if the field sequence number specifies that click intensity is the first position, browsing interest ratio is the second position, and real-time activity is the third position, then the real-time behavior features are arranged in this order.

[0030] Simultaneously, based on historical behavior pattern indicators, historical feature regions from multiple feature storage spaces are retrieved for feature analysis to construct historical behavior features. Specifically, based on the same user identifier, indicators such as long-term interest distribution, long-term active time interval, and long-term click intensity level are read from the historical feature regions of the corresponding feature storage space. These indicators are then arranged according to the field sequence number of the historical side field in the feature fusion mapping table to form a fixed-order field sequence as historical behavior features, ensuring that historical behavior features and real-time behavior features maintain consistency in field structure.

[0031] After obtaining real-time and historical behavioral features, dimensional alignment is performed according to the feature fusion mapping table. Specifically, based on the dimension index number, fields with the same dimension index number in the real-time and historical behavioral features are placed in the corresponding alignment positions to ensure that semantically consistent fields are in the same dimension. After completing the dimensional alignment, the real-time and historical behavioral features are horizontally concatenated according to the alignment result. That is, the aligned real-time field sequence and the historical field sequence are connected sequentially according to the field sequence number to generate the first fused feature vector. For example, if the real-time behavioral features include click intensity and browsing interest ratio, and the historical behavioral features include long-term click intensity level and long-term active time interval, then the first fused feature vector is arranged in the following order: click intensity, browsing interest ratio, long-term click intensity level, and long-term active time interval.

[0032] Subsequently, feature conflict identification is performed on the first fused feature vector, and conflict terms are extracted and removed to generate the second fused feature vector. Specifically, the Pearson correlation coefficient is calculated for field pairs marked with the same dimension index number in the first fused feature vector. The Pearson correlation coefficient ranges from -1 to 1. A correlation threshold of 0.95 is set. When the correlation coefficient is greater than or equal to 0.95, the field pair is determined to be highly correlated and is regarded as a conflict term. After identifying the conflict term, feature removal is performed according to a fixed retention rule, that is, real-time side fields are retained and corresponding historical side fields are deleted, thereby removing redundant information and obtaining the second fused feature vector.

[0033] Finally, cross-validation is performed based on the second fused feature vector to generate feature fusion data. Specifically, a 5-fold cross-validation method is used, dividing all samples into 5 subsets; in each round, one subset is selected as the validation subset, and the other 4 subsets are selected as the statistical analysis subset. The correlation between conversion rate and features is calculated in the statistical analysis subset, and the corresponding correlation is calculated in the validation subset. The correlation is measured using the AUC index, with an AUC value ranging from 0.5 to 1. The pass criterion is that the AUC values ​​calculated in 5 rounds are all greater than or equal to 0.70, and the difference between the maximum and minimum AUC values ​​in 5 rounds is less than or equal to 0.05. When the above criteria are met, the second fused feature vector is determined to be stable and is identified as feature fusion data.

[0034] Step S200: Introduce historical conversion pattern data, combine the real-time user behavior characteristics with the historical conversion pattern data for multi-objective optimization, and construct a preliminary scheduling strategy.

[0035] In this embodiment, historical conversion pattern data is first introduced, and then real-time user behavior features are combined with historical conversion pattern data for multi-objective optimization. In this process, real-time user behavior features and historical conversion pattern data are first aligned by establishing a unified field index and dimension correspondence to generate feature alignment results. Then, feature concatenation processing is performed on the real-time user behavior features and historical conversion pattern data based on the feature alignment results, forming a joint feature vector containing real-time behavior information and historical conversion path information. Next, joint optimization objectives are set, including conversion rate improvement, cost control, and revenue maximization objectives. Multi-objective optimization calculations are performed based on the joint feature vector to generate multi-objective strategy optimization results. After obtaining the multi-objective strategy optimization results, multiple rounds of deep reinforcement iterative learning are performed according to the results. Multi-objective revenue parameters are generated through iterative updates, and combined analysis is conducted based on these parameters to identify multiple strategy schemes. Finally, multiple strategy schemes are traversed, filtered, and constraint-optimized to determine the initial scheduling strategy.

[0036] Furthermore, in the method provided in the application embodiments, the process of constructing historical transformation pattern data also includes: The system retrieves the historical conversion dataset for advertising, groups it according to multiple user identifiers, and constructs complete historical behavior time series for multiple users. It then segments these complete historical behavior time series to obtain complete path sequences. Frequent sequence pattern mining and filtering are performed on these complete path sequences to identify typical conversion paths. Path feature analysis is performed on these typical conversion paths to generate path efficiency indicators. Channel conversion preference analysis is conducted on the historical behavior of multiple users to obtain user channel preference characteristics. The time-segment conversion distribution of the historical behavior of multiple users is also analyzed to obtain user time-segment response characteristics. Finally, the typical conversion paths, path efficiency indicators, user channel preference characteristics, and user time-segment response characteristics are combined and encapsulated to construct the historical conversion pattern data.

[0037] In this embodiment, the advertising historical conversion basic dataset is first retrieved, that is, the advertising historical conversion basic dataset is read from the data storage. The advertising historical conversion basic dataset includes at least an event type field, a user identifier field, a channel identifier field, and a timestamp field, and includes exposure records, click records, access records, and conversion records. Then, all records are grouped using the user identifier field as the grouping key, and records corresponding to the same user identifier are merged into the same group. Within each group, the records are sorted in ascending order according to the timestamp field, so that the exposure, click, access, and conversion events of the same user are arranged according to the occurrence time, thereby constructing a complete historical behavior time series of multiple users. The complete historical behavior time series is a list of events arranged in chronological order.

[0038] Next, when segmenting the complete historical behavior time series of multiple users, the event type field of each user's complete historical behavior time series is identified as a conversion event record, and this conversion event record is used as the termination point. Starting from the beginning event record of the user's sequence, the entire sequence is traversed backward, and all event records from the beginning event record to the termination point are extracted to form a subsequence. When the same user has multiple conversion event records, the extraction process is repeated with each conversion event record as the termination point, thereby obtaining multiple complete path sequences. The complete path sequence is a sequence of behavior path events that ends with a conversion event.

[0039] Subsequently, when performing frequent sequence pattern mining and filtering on the complete path sequences, the event type field in each complete path sequence is extracted in chronological order to form an event sequence representation. Subsequence enumeration is then performed on the event sequences, extracting consecutive or non-consecutive event combinations while maintaining the event order. Next, the occurrence frequency of each candidate subsequence in all complete path sequences is counted. The frequency is calculated by traversing all complete path sequences and checking if the event order of the candidate subsequence is included; if so, the count is incremented. The support is then calculated by dividing the occurrence frequency of the candidate subsequence by the total number of complete path sequences. A support threshold of 0.1 is set, and candidate subsequences with a support greater than or equal to 0.1 are retained. Among the retained candidate subsequences, they are sorted from highest to lowest support, and the subsequence with the highest support and the number of events reaching the preset minimum length is selected as the typical conversion path. The typical conversion path is the conversion behavior path that has a high recurrence frequency in historical data.

[0040] When performing path feature analysis on typical conversion paths, all complete path sequences are traversed, and path sequences whose event sequence contains the typical conversion path are selected as path samples. Then, the path conversion success rate is calculated by dividing the number of path samples by the total number of complete path sequences. The average path conversion time is calculated by taking the difference between the start event timestamp and the end conversion event timestamp for each path sample, and then taking the arithmetic mean of all individual path conversion times. After obtaining the path conversion success rate and the average path conversion time, a path efficiency index is calculated. The path efficiency index is calculated by dividing the path conversion success rate by the average path conversion time, so that the path efficiency index simultaneously reflects the probability level of conversion after the occurrence of the typical conversion path and the time cost required to complete the conversion.

[0041] Subsequently, channel conversion preference analysis was performed on multiple users based on their historical behavior. During this process, records with "conversion" as the event type field in the historical advertising conversion dataset were filtered, retaining only conversion records. Then, the data was double-grouped by user identifier and channel identifier fields, and the number of conversions for each user on each channel was counted. Next, the total number of conversions for the same user across all channels was summed to obtain the user's total conversion count. Then, the user's channel conversion preference value on a specific channel was calculated by dividing the user's conversion count on that channel by the user's total conversion count. Finally, the user's channel conversion preference values ​​across all channels were arranged in order of channel identifier to form the user's channel preference characteristics, which characterize the distribution of the user's conversion tendency across different channels.

[0042] When iterating through the time-period conversion distribution of multiple users' historical behaviors, records with the event type field "conversion" are filtered, retaining only conversion records. Each conversion record is mapped to an hourly time interval based on the timestamp field, by extracting the hour value corresponding to the timestamp and assigning it to an hour interval of 0 to 23. Then, users are grouped by user identifier field and hour interval, and the number of conversions for each user within each hour interval is counted. The total number of conversions for the same user across 24-hour intervals is then summed. Next, the user's time-period response value for a specific hour interval is calculated by dividing the user's conversion count in that hour interval by the user's total conversion count. Finally, the time-period response values ​​for the 24-hour intervals are arranged in hourly order to form the user's time-period response characteristics, which characterize the distribution of user conversion responses across different time periods.

[0043] Finally, when combining and encapsulating typical conversion paths, path efficiency indicators, user channel preference characteristics, and user time-period response characteristics, a historical conversion pattern data record is created for each user. The typical conversion path is written into the path field using event sequence encoding, the path efficiency indicator is written into the efficiency field, the user channel preference characteristics are written into the channel preference field in channel identifier order, and the user time-period response characteristics are written into the time-period response field in hourly order. Subsequently, the path field, efficiency field, channel preference field, and time-period response field are concatenated into a single record in a fixed field order to obtain historical conversion pattern data containing typical conversion paths, path efficiency indicators, user channel preference characteristics, and user time-period response characteristics.

[0044] Furthermore, the method provided in the application embodiment, which combines the real-time user behavior features with the historical conversion pattern data for multi-objective optimization to construct a preliminary scheduling strategy, also includes: The real-time user behavior features are aligned with the historical conversion pattern data to generate a feature alignment result. Based on the feature alignment result, the real-time user behavior features and the historical conversion pattern data are concatenated to generate a joint feature vector. A joint optimization objective is set, and multi-objective optimization is performed based on the joint feature vector to generate a multi-objective strategy optimization result. Multiple rounds of deep reinforcement iterative learning are performed according to the multi-objective strategy optimization result to generate multi-objective benefit parameters. The multi-objective benefit parameters are combined to identify multiple strategy schemes. The multiple strategy schemes are traversed for screening and optimization to determine the preliminary scheduling strategy.

[0045] In this embodiment, when aligning real-time user behavior features with historical conversion pattern data, the field structures of both are first read. Real-time user behavior features are behavioral feature vectors containing fields such as click intensity, browsing interest ratio, and real-time activity. Historical conversion pattern data are pattern feature vectors containing fields such as typical conversion paths, path efficiency indicators, user channel preference characteristics, and user time-period response characteristics. Then, using user identifiers as the matching basis, records corresponding to the real-time user behavior features are located in the historical conversion pattern data, placing the two types of data at the same user dimension. After user matching is completed, based on the predefined field correspondence, the fields in the real-time user behavior features are mapped to the fields in the historical conversion pattern data, and a unified dimension index position is assigned to each field, so that fields with the same semantics or belonging to the same category are arranged in the same dimension position. Then, the fields of the real-time user behavior features and historical conversion pattern data are rearranged according to the unified dimension index order, so that the two types of data are consistent in field order and dimension structure. After completing the field mapping and order rearrangement, an aligned data structure with one-to-one field position and unified dimension structure is obtained. This data structure is the feature alignment result.

[0046] Next, based on the feature alignment results, the real-time user behavior features are concatenated with the historical conversion pattern data. In this process, firstly, using the user identifier as the index key, a traversal search is performed on the historical conversion pattern data corresponding to the real-time user behavior features to generate search results. When a search result exists, the historical behavior features and real-time user behavior features are aligned according to the timestamp field to extract feature matching results within the corresponding time interval. When a search result does not exist, the user identifier is determined to be a new user, and global statistical analysis is performed on the historical conversion pattern data to extract global feature parameters as substitute features. Subsequently, based on the feature matching results or global feature parameters, the real-time user behavior features and historical behavior features are concatenated to generate paired features. Then, a first-category feature is extracted from the real-time user behavior features, and a second-category feature with a corresponding relationship is extracted from the historical behavior features based on the first-category feature. Cross-feature terms are generated by cross-combining the first-category feature and the second-category feature. Finally, the paired features and cross-feature terms are integrated according to a unified field order to form a joint feature vector containing basic features and cross-features.

[0047] Subsequently, a joint optimization objective is set, and multi-objective optimization is performed based on the joint feature vector. In this process, the joint optimization objectives are first determined, including conversion rate, cost of placement, and advertising revenue. Conversion rate is calculated by dividing the number of conversions by the number of impressions; cost of placement is calculated by dividing the amount spent on placement by the number of conversions; and advertising revenue is calculated by subtracting the amount spent on placement from the amount spent on placement. The number of conversions, impressions, amount spent on placement, and amount spent on placement are all calculated by replaying historical placement data, which at least includes fields for impression events, conversion events, amount spent on placement, and amount spent on placement. Then, within the sample range corresponding to the joint feature vector, adjustable placement parameters are set as bid coefficient, budget allocation ratio, and channel allocation ratio. The bid coefficient is pre-set to have a value range of 0.80 to 1.20 with a step size of 0.05, the budget allocation ratio is adjusted in 5 percentage point increments, and the channel allocation ratio is adjusted in 10 percentage point increments. All candidate placement parameter combinations are generated within the preset value ranges and step sizes. For each candidate combination of placement parameters, historical placement data playback calculations are performed. This involves filtering and measuring historical placement data based on the candidate placement parameter combination, adjusting the historical bid field proportionally according to the bid coefficient to form a simulated bid, extracting historical exposure events from different channels proportionally according to the channel allocation ratio, setting an upper limit on the cumulative placement cost according to the budget allocation ratio, and stopping the calculation when the upper limit is reached. Within the filtered historical sample, the number of exposures is counted as the number of exposure events, the number of conversions is counted as the number of conversion events, the placement cost is calculated by summing the consumption amount field within the sample, and the conversion revenue is calculated by summing the revenue amount field within the sample. This yields the conversion rate, placement cost, and advertising revenue corresponding to the candidate placement parameter combination. For all candidate placement parameter combinations, calculate the maximum and minimum values ​​of the three indicators. Normalize the conversion rate and advertising revenue by subtracting the minimum value from the current value and dividing by the maximum value minus the minimum value. Normalize the placement cost by subtracting the current value from the maximum value and dividing by the maximum value minus the minimum value, ensuring that the higher the normalized value of each indicator, the better. Then, perform a weighted summation according to pre-set and fixed weight parameters: conversion rate weight is 0.4, advertising revenue weight is 0.4, and placement cost weight is 0.2, with the sum of the three being 1, to obtain the comprehensive score for each candidate placement parameter combination. Sort the candidate placement parameter combinations from high to low according to their comprehensive scores, and select the candidate placement parameter combinations with the top 30% comprehensive scores as the multi-objective strategy optimization results.

[0048] Subsequently, during multiple rounds of deep reinforcement iterative learning based on the multi-objective strategy optimization results, each candidate placement parameter combination in the multi-objective strategy optimization results is taken as a candidate strategy, and the number of iteration rounds is pre-set to 10 rounds. In each round, historical placement data replay calculation is performed on the same candidate strategy, and the replay calculation steps are the same as described above, so that the conversion rate, placement cost and advertising revenue corresponding to the candidate strategy are obtained in each round. The maximum and minimum values ​​of the three indicators are calculated for all candidate strategies in each round and normalized. The conversion rate and advertising revenue are calculated by subtracting the minimum value from the current value and dividing by the maximum value minus the minimum value, and the placement cost is calculated by subtracting the current value from the maximum value and dividing by the maximum value minus the minimum value. The revenue value of the round is calculated and recorded with weights of 0.4, 0.4 and 0.2. The revenue values ​​of the same candidate strategy in 10 rounds are accumulated to obtain the multi-objective revenue parameter of the candidate strategy, which is used to represent the comprehensive performance value of the candidate strategy under multiple rounds of replay evaluation.

[0049] Next, when identifying multiple strategy options based on multi-objective return parameters, all candidate strategies are sorted from high to low according to the multi-objective return parameters. The largest multi-objective return parameter is taken as the benchmark value, and the screening ratio is preset to 0.90. Candidate strategies with multi-objective return parameters greater than or equal to the benchmark value multiplied by 0.90 are retained. The retained candidate strategies are grouped according to the parameter proximity rule. The parameter proximity rule is preset to be that the difference in bidding coefficients does not exceed 0.05 and the difference in the allocation ratio of the main channels does not exceed 10 percentage points. Candidate strategies that meet this rule are grouped into the same group. Within each group, the candidate strategy with the highest multi-objective return parameter is selected as the representative to obtain multiple strategy options.

[0050] Finally, when iterating through multiple strategy options for selection and optimization, historical data playback is performed again for each strategy option to calculate conversion rate, cost of placement, and advertising revenue. Within the scope of the strategy option, the maximum and minimum values ​​of the three indicators are calculated and normalized. Conversion rate and advertising revenue are calculated by subtracting the minimum value from the current value and then dividing by the maximum value minus the minimum value. Cost of placement is calculated by subtracting the current value from the maximum value and then dividing by the maximum value minus the minimum value. Then, the comprehensive score of each strategy option is calculated according to the pre-set weights of 0.4, 0.4, and 0.2. The strategy option with the highest comprehensive score is selected as the initial scheduling strategy, and the corresponding bidding coefficient, budget allocation ratio, and channel allocation ratio of the strategy option are output as the initial scheduling strategy parameters.

[0051] Furthermore, in the method provided in the application embodiments, the method of concatenating the real-time user behavior features with the historical conversion pattern data based on the feature alignment result further includes: The real-time user behavior features are used as index keys to traverse and search the historical conversion pattern data, generating search results, which are either present or absent. When a search result is present, the historical behavior features and real-time user behavior features are aligned along the time dimension, and feature matching results are extracted. When a search result is absent, a new user is generated, and the historical conversion pattern data is traversed for global analysis to extract global feature parameters. Based on the feature matching results or the global feature parameters, the real-time user behavior features and the historical behavior features are concatenated to obtain paired features. A first categorical feature is extracted based on the real-time user behavior features, and a second categorical feature is extracted from the historical behavior features according to the first categorical feature, where the first categorical feature and the second categorical feature have a corresponding relationship. The first categorical feature and the second categorical feature are cross-combined to generate cross feature terms. The paired features and cross feature terms are combined to generate the joint feature vector.

[0052] In this embodiment, when reading the user identifier field from real-time user behavior features as an index key and performing a traversal search in historical conversion pattern data to generate search results, the user identifier value corresponding to the current real-time user behavior feature is first taken. Then, the user identifier field in the historical conversion pattern data is read one by one in the record order and compared with the user identifier value. When an equality is found, the record position of the historical conversion pattern data is immediately recorded and the search result is output as an existence result. At the same time, the corresponding historical behavior feature field in the record is extracted. When no equality is found after the traversal is completed, the search result is output as a non-existent result.

[0053] When a search result is found, the historical behavior features and real-time user behavior features are aligned along the time dimension. This process involves first reading the time interval and statistical granularity fields from the real-time user behavior features, and then reading the time interval sequence corresponding to the user time-period response features from the historical behavior features. If the time interval of the real-time user behavior features is at the hour level, the response value corresponding to the same hour is directly located and retrieved from the user time-period response features. If the time interval of the real-time user behavior features is at the minute level, the corresponding hour interval is first determined using the hour value to which the minute belongs, and then the response value corresponding to that hour is located and retrieved from the user time-period response features. The response value is then placed in the same record along with fields such as click intensity, browsing interest ratio, and real-time activity from the real-time user behavior features to form the feature matching result.

[0054] When the search result is non-existent, a new user is generated, and historical conversion pattern data is traversed for global analysis. In this process, the user is first identified as a new user. Then, all historical conversion pattern data is traversed one by one to count the occurrence frequency of the typical conversion path field. The occurrence frequency of each path is divided by the total number of historical conversion pattern data records to obtain the occurrence ratio. The typical conversion path with the highest occurrence ratio is selected as the global typical conversion path. The path efficiency index field is summed one by one and divided by the total number of records to obtain the global path efficiency index. The user channel preference characteristics are summed one by one by channel dimension and divided by the total number of records to obtain the global channel preference value for each channel. The user time period response characteristics are summed one by one by hour dimension and divided by the total number of records to obtain the global time period response value for each hour. The global typical conversion path, global path efficiency index, global channel preference value for each channel, and global time period response value for each hour are combined to obtain global feature parameters, which are used as replacement data for the historical behavioral characteristics of the new user.

[0055] When concatenating real-time user behavior features with historical behavior features based on feature matching results or global feature parameters, the historical input is first selected according to the search results. If a result exists, the historical behavior feature in the feature matching result is selected; otherwise, the historical behavior feature in the global feature parameters is selected. Then, the real-time user behavior feature field is written in the preset field order, and the historical behavior feature field is written in the same order, so that the two types of fields are fixedly arranged in the same record, thus obtaining paired features. The paired features include real-time behavior fields and historical pattern fields, and the field order is kept constant.

[0056] When extracting the first categorical feature based on real-time user behavior features and extracting the second categorical feature from historical behavior features according to the first categorical feature, the channel identifier field and time interval field are first read from the real-time user behavior features as the first categorical feature. Then, based on the channel identifier field, the preference value corresponding to the same channel in the user channel preference feature is located in the historical behavior features, and this preference value is used as the second categorical feature. Next, based on the time interval field, the response value corresponding to the same time interval in the user time period response feature is located, and this response value is used as another second categorical feature, so that the channel identifier corresponds to the channel preference value, and the time interval corresponds to the time period response value, thereby completing the corresponding extraction of the first categorical feature and the second categorical feature.

[0057] When combining the first and second categorical features, a joint coding term is generated for each pair of correspondences. The joint coding term is obtained by concatenating the values ​​of the first and second categorical features. Then, the occurrence and conversion frequency of the joint coding term are counted within the sample range. The occurrence frequency is calculated based on the number of records in the sample where the joint coding term appears, and the conversion frequency is calculated based on the number of conversion events in the record corresponding to the joint coding term. The conversion rate of the joint coding term is obtained by dividing the conversion frequency by the occurrence frequency. This conversion rate is used as the value of the cross-feature term to obtain the channel cross-feature term and the time period cross-feature term, respectively.

[0058] Finally, when combining paired features and cross features, the paired features are first used as the first half of the joint feature vector. Then, the channel cross features and time period cross features are added after the paired features in the preset field order as the second half. The position of each field in the vector is kept fixed, and finally a joint feature vector containing real-time user behavior feature fields, historical behavior feature fields and cross feature field is formed.

[0059] Step S300: Set advertising placement constraints, execute the preliminary scheduling strategy to optimize resource allocation, and formulate a resource scheduling plan.

[0060] In this embodiment, when optimizing resource allocation by executing a preliminary scheduling strategy after setting advertising placement constraints, real-time bidding data is first introduced. Based on the real-time bidding data, environmental analysis is performed according to advertising channels to extract indicators such as average bidding level, competition intensity, and exposure capacity of each advertising channel, generating real-time market environment parameters. Subsequently, the preliminary scheduling strategy is encoded and converted according to the advertising placement constraints, transforming parameters such as budget ratio, bidding coefficient, and channel allocation ratio into encoded forms, forming multiple encoded populations, with each encoded individual corresponding to a candidate resource allocation scheme. Then, fitness calculation is performed on multiple candidate resource allocation schemes in conjunction with the real-time market environment parameters, considering factors such as revenue indicators, A multidimensional fitness assessment is constructed using dimensions such as cost control indicators and budget satisfaction indicators. Based on this fitness assessment, multiple encoded populations are traversed and compared, and then sorted in descending order according to the benefit dimension to form a benefit ranking sequence. Random sampling is performed on these populations to construct multiple competitive groups, and cyclic cross-matching is performed within each group to generate multiple offspring parameters. Subsequently, mutation and iterative evaluation are performed on these offspring parameters, and a target population is constructed through multiple rounds of screening. Individuals within the target population are then decoded and restored to obtain multiple analytical parameters. Finally, these analytical parameters are integrated to form a resource scheduling scheme that satisfies advertising constraints and adapts to the real-time market environment.

[0061] Furthermore, in the method provided in the application embodiments, setting advertising placement constraints, executing the preliminary scheduling strategy to optimize resource allocation, and formulating a resource scheduling scheme further includes: Real-time bidding data is introduced, and environmental analysis is performed based on the real-time bidding data according to advertising channels to generate real-time market environment parameters. The initial scheduling strategy is executed according to the advertising placement constraints to perform encoding conversion, obtaining multiple encoding populations, each containing multiple candidate resource allocation schemes. Fitness is calculated on the multiple candidate resource allocation schemes based on the real-time market environment parameters to obtain multi-dimensional evaluation fitness. The multiple encoding populations are traversed and sorted in descending order according to the multi-dimensional evaluation fitness to obtain a profit ranking sequence. Based on the profit ranking sequence, multiple encoding populations are randomly selected to determine multiple competitive groups, and cyclic cross-matching is performed to generate multiple offspring parameters. Mutation iteration evaluation is performed based on the multiple offspring parameters to construct a target population for individual decoding and reconstruction, obtaining multiple analytical parameters. The multiple analytical parameters are integrated to construct the resource scheduling scheme.

[0062] In this embodiment, real-time bidding data is first introduced, and environmental analysis is performed based on the real-time bidding data according to advertising channels. In this process, real-time bidding data is first read from the bidding system interface. The real-time bidding data includes at least the following fields: channel identifier, bid amount, transaction amount, number of bidding participations, number of transactions, and number of impressions. Then, the real-time bidding data is grouped and statistically analyzed using the channel identifier field as the grouping key. For each advertising channel, the average transaction price is calculated, which is equal to the sum of the transaction amounts divided by the sum of the number of transactions. The bidding success rate is calculated, which is equal to the sum of the number of transactions divided by the sum of the number of bidding participations. The exposure capacity per unit time is calculated, which is equal to the sum of the number of impressions within a fixed statistical time window. The average transaction price, bidding success rate, and exposure capacity are combined to form real-time market environment parameters.

[0063] Next, the initial scheduling strategy is executed and encoded according to the advertising placement constraints. In this process, the bidding coefficient, budget allocation ratio, and channel allocation ratio in the initial scheduling strategy are read first, along with the budget cap, the lower and upper limits of the single bid range, and the upper and lower limits of the channel ratios in the advertising placement constraints. Under the premise of meeting the advertising placement constraints, the bidding coefficient is perturbed at a fixed step size to generate multiple candidate values; the budget allocation ratio is divided into multiple ratio combinations within the budget cap range at a fixed ratio; and the channel allocation ratio is generated into multiple allocation combinations within the upper and lower limits of each channel ratio at a fixed step size. Each set of bidding coefficient, budget allocation ratio, and channel allocation ratio is encoded into a numerical sequence according to a fixed field order; this numerical sequence is a coded individual. All coded individuals are aggregated to form multiple coded populations, with each coded individual corresponding to a candidate resource allocation scheme.

[0064] Subsequently, when calculating the fitness of multiple candidate resource allocation schemes based on real-time market environment parameters, revenue and cost prediction calculations are performed for each coded individual. First, the bidding coefficient of the coded individual is compared with the average transaction price of the corresponding advertising channel. When the bidding coefficient multiplied by the historical benchmark bid is greater than or equal to the average transaction price in the real-time market environment parameters, the bidding success rate of that channel is multiplied by the exposure capacity to obtain the expected number of transactions. Then, the expected number of transactions is multiplied by the historical average conversion revenue to obtain the expected revenue, and the expected number of transactions is multiplied by the average transaction price to obtain the expected expenditure amount. The revenue dimension score is calculated by dividing the expected revenue by the maximum expected revenue among all coded individuals. The cost control dimension score is calculated by dividing the minimum expected expenditure amount by the current expected expenditure amount. The constraint satisfaction score is calculated by taking a value of 1 when the expected expenditure amount is less than or equal to the budget limit, otherwise, the budget limit is divided by the expected expenditure amount to calculate the proportion value. The revenue dimension score, cost control dimension score, and constraint satisfaction score are weighted and summed according to preset weights of 0.5, 0.3, and 0.2 to obtain the multi-dimensional evaluation fitness.

[0065] Next, the multiple coding populations are traversed and sorted in descending order according to the multidimensional evaluation fitness. All coding individuals are sorted from high to low according to the multidimensional evaluation fitness value. The sorting algorithm adopts the numerical comparison sorting method, and the coding individuals with larger fitness values ​​are placed first, resulting in a fitness-sorted sequence.

[0066] Subsequently, multiple coded populations are randomly selected based on the profit ranking sequence to determine multiple competitive groups for cyclic cross-matching. In this process, two coded individuals are randomly selected with equal probability from the top 50% of the profit ranking sequence to form a competitive group. Within the competitive group, certain parameter segments in the coded sequence are exchanged at fixed positions. The exchanged parameter segments include the bid coefficient position, budget allocation ratio position, or channel allocation ratio position. After the exchange, two new coded sequences are generated, each of which is a child parameter. This random selection and parameter exchange process is repeated multiple times until multiple child parameters are generated.

[0067] Next, based on multiple offspring parameters, iterative mutation evaluation is performed to construct the target population for individual decoding and restoration. For each offspring parameter, a single-point mutation is performed with a preset mutation probability of 0.1. Single-point mutation involves randomly selecting a parameter position in the coding sequence and increasing or decreasing it by a fixed step size within the allowed value range. The fitness of the mutated offspring parameters is recalculated and compared with the fitness of the corresponding parent parameters. The coding individuals with higher fitness are retained to enter the next generation population. After repeating the crossover and mutation process for several rounds, the target population is obtained. Each coding individual in the target population is decoded in field order to restore the numerical sequence to the bidding coefficient, budget allocation ratio, and channel allocation ratio. The restored parameters are the analytical parameters.

[0068] Finally, multiple analytical parameters are integrated, and the analytical parameters with the highest fitness in multidimensional evaluation are selected from the target population as the optimal parameter combination. The bidding coefficient, budget allocation ratio and channel allocation ratio in the analytical parameters are then allocated and mapped according to the advertising channels to form the final resource scheduling scheme.

[0069] Step S400: Simulate the execution of the resource scheduling scheme to track the effect, obtain the simulation scheduling parameters for incremental updates, and construct a closed-loop optimization report for advertising resource scheduling.

[0070] In this embodiment, when simulating the execution of a resource scheduling scheme for effect tracking, an advertising simulation environment is first constructed, and the bidding configuration, budget allocation ratio, and channel allocation structure in the resource scheduling scheme are mapped to the advertising simulation environment for simulated placement, generating a simulated placement effect dataset containing indicators such as impressions, clicks, conversions, cost, and revenue. Then, historical resource scheduling effect data is introduced, and the conversion rate, cost control level, and revenue level within the historical period are statistically analyzed, with target effect data set as a benchmark. Next, the simulated placement effect dataset and the target effect data are compared item by item, calculating multiple effect deviation indicators such as conversion rate deviation, cost deviation, and revenue deviation. Subsequently, based on these multiple effect deviation indicators, a deviation attribution analysis is performed, identifying the scheduling deviation parameters that cause the deviation by associating with resource allocation ratios, bidding parameters, and channel configuration structures. Finally, the scheduling deviation parameters are associated and stored with the corresponding multiple effect deviation indicators to form simulation scheduling parameters.

[0071] After generating simulation scheduling parameters, the resource scheduling scheme is incrementally updated based on these parameters. Specifically, the scheduling deviation parameters recorded in the simulation scheduling parameters, along with their corresponding conversion rate deviation, cost deviation, and revenue deviation values, are read. The bidding configuration, budget allocation ratio, and channel allocation structure in the resource scheduling scheme are then quantitatively corrected according to the deviation direction. When the conversion rate deviation is less than zero, the budget allocation ratio of the corresponding channel is lowered according to a preset adjustment ratio, while the budget allocation ratio of channels with positive conversion rate deviations is simultaneously increased to keep the total budget constant. When the cost deviation is greater than zero, the bidding coefficient of the corresponding channel is lowered by a fixed step size, which is a pre-set value. When the revenue deviation is less than zero, a portion of the budget is transferred proportionally from channels with negative revenue deviations to channels with positive revenue deviations. After completing the parameter correction, an updated resource scheduling scheme is generated.

[0072] The updated resource scheduling scheme is then mapped back to the advertising simulation environment for simulation playback. A new dataset of simulated campaign performance is generated, and the new conversion rate deviation, cost deviation, and revenue deviation are calculated by comparing each item with the target performance data. When the absolute value of the new conversion rate deviation is less than the preset conversion rate threshold, the absolute value of the cost deviation is less than the preset cost threshold, and the absolute value of the revenue deviation is less than the preset revenue threshold, the resource scheduling scheme is considered to have reached a stable state. If any deviation exceeds the corresponding threshold, incremental updates are performed based on the new simulation scheduling parameters, and the simulation verification process is repeated until all deviations meet the threshold conditions.

[0073] After the resource scheduling scheme reaches a stable state, the initial resource scheduling scheme, the incremental update records of each round, the change data of simulation scheduling parameters in each round, the change trajectory of multiple performance deviation indicators, and the final stable conversion rate, placement cost and advertising revenue indicators are structured, organized and archived to form an advertising placement resource scheduling closed-loop optimization report containing data of the entire process of strategy generation, simulation verification, deviation correction and parameter convergence.

[0074] Furthermore, in the method provided in the application embodiments, simulating the execution of the resource scheduling scheme to track its effects and obtain simulation scheduling parameters further includes: An advertising simulation environment is constructed, and the resource scheduling scheme is mapped to the advertising simulation environment for simulated advertising, generating a simulated advertising effect dataset. Historical resource scheduling effects are introduced for expected analysis, and target effect data is set. The simulated advertising effect dataset and the target effect data are compared and analyzed item by item to calculate multiple effect deviation indicators. Based on the multiple effect deviation indicators, deviation attribution is performed to determine scheduling deviation parameters. The scheduling deviation parameters are combined with the multiple effect deviation indicators and stored in association to generate the simulation scheduling parameters.

[0075] In this embodiment, when constructing the advertising simulation environment, a historical log replay method is used to reproduce the bidding and behavior chain. Specifically, historical exposure logs, click logs, and conversion logs are retrieved. The historical exposure logs include fields for request time, request identifier, user identifier, channel identifier, advertising identifier, historical transaction price, and cost. The click logs include fields for click time, request identifier, and click identifier. The conversion logs include fields for conversion time, request identifier, conversion identifier, and conversion cost. Subsequently, the historical exposure logs are sorted in ascending order by the request time field to form a replay sequence. The click logs and conversion logs are matched and associated with the historical exposure logs using the request identifier field as the association key. This allows each exposure request record to be associated with the corresponding click record and conversion record, thereby forming a replayable chain containing exposure, click, and conversion under the same request identifier dimension. After the association is completed, an index is created for the replay sequence by the channel identifier field, enabling the simulation environment to perform replay scheduling and statistics by advertising channel. The index is used to quickly locate the exposure request record of the corresponding channel during replay, thus obtaining the advertising simulation environment.

[0076] Next, the resource scheduling plan is mapped to the advertising simulation environment for simulated campaign execution. Specifically, the bidding configuration, budget allocation ratio, and channel allocation structure from the resource scheduling plan are read, and the bidding configuration is written to the channel parameter table of the simulation environment, the budget allocation ratio is written to the channel budget table of the simulation environment, and the channel allocation structure is written to the channel delivery table of the simulation environment. Then, historical exposure requests are played back one by one according to the playback sequence. For each exposure request, the bidding configuration of the corresponding channel is read according to the channel identifier field, and the simulated bid is calculated. The simulated bid is equal to the historical benchmark bid multiplied by the bidding coefficient. The simulated bid is compared with the historical transaction price field corresponding to the exposure request. When the simulated bid is greater than or equal to the historical transaction price field, it is determined as a simulated transaction and is counted in the exposure count. The cost of the exposure request is included in the cumulative cost. After determining it to be a simulated transaction, the related click and conversion logs are retrieved using the request identifier field. If a click record exists, it is counted as a click; if a conversion record exists, it is counted as a conversion, and the conversion amount is included in the cumulative revenue. During the replay, budget constraints are applied to the cumulative cost of each channel. When the cumulative cost of a channel reaches the channel budget limit corresponding to the budget allocation ratio, the counting of simulated transactions for subsequent exposure requests on that channel is stopped. After the replay ends, statistical results such as the number of exposures, clicks, conversions, cost, and revenue are output to form a simulated campaign performance dataset.

[0077] Next, when introducing historical resource scheduling effects for expected analysis, historical resource scheduling effect records are retrieved. These records include at least the following fields: historical period, actual exposure count, actual click count, actual conversion count, actual cost, and actual revenue. For each historical period, the historical conversion rate is calculated (actual conversion count divided by actual exposure count). The historical cost of goods sold is calculated (actual cost divided by actual conversion count). The historical revenue is calculated (actual revenue minus actual cost). Subsequently, the arithmetic mean of the historical conversion rates over N consecutive historical periods is used to obtain the target conversion rate. The arithmetic mean of the historical cost of goods sold over N consecutive historical periods is used to obtain the target cost of goods sold. The arithmetic mean of the historical revenue over N consecutive historical periods is used to obtain the target revenue. Here, N is a pre-defined number of historical periods. The target conversion rate, target cost of goods sold, and target revenue are combined to form the target performance data.

[0078] When comparing and analyzing the simulated campaign performance dataset with the target performance data item by item, the process first involves performing user behavior simulation analysis based on both types of data. This involves calculating the probability of clicks and conversions to generate simulated response probability parameters, and then calculating the ratio of these simulated response probability parameters to preset response probability parameters to obtain a response deviation index. Next, a market bidding simulation analysis is performed based on the simulated campaign performance dataset and the target performance data. This involves statistically analyzing the number of transactions and bidding attempts to generate simulated bidding conversion rate parameters, and then calculating the ratio of these simulated bidding conversion rate parameters to preset simulated bidding conversion rate parameters to obtain a bidding deviation index. Then, a channel response simulation analysis is performed based on both types of data. This involves statistically analyzing the number of impressions on each channel to generate simulated exposure parameters, and then calculating the ratio of these simulated exposure parameters to preset exposure parameters to obtain an exposure deviation index. Finally, the response deviation index, bidding deviation index, and exposure deviation index are integrated to form multiple performance deviation indices.

[0079] Subsequently, when attributing deviations based on multiple performance deviation metrics, the following steps were taken: First, the channel identifier, time interval, impression count, click count, conversion count, bidding count, and transaction count fields were extracted from the simulated campaign performance dataset. These fields were then grouped and statistically analyzed according to the channel identifier and time interval fields. For each group, a simulated response probability parameter was calculated (equal to the conversion count field divided by the click count field). A simulated bidding conversion rate parameter was also calculated (equal to the transaction count field divided by the bidding count field). A simulated impression count parameter was also calculated (equal to the impression count field). Next, channel-level response deviation metrics were calculated (equal to the simulated response probability parameter divided by the preset response probability parameter). Similarly, channel-level bidding deviation metrics were calculated (equal to the simulated bidding conversion rate parameter divided by the preset simulated bidding conversion rate parameter). Finally, channel-level impression deviation metrics were calculated. The deviation index is calculated as follows: the channel-level exposure deviation index equals the simulated exposure parameter divided by the preset exposure parameter; subsequently, based on the channel identifier field, the bidding configuration, budget allocation ratio, and channel allocation structure corresponding to the same channel are located in the resource scheduling scheme, and the bidding coefficient in the bidding configuration, the channel budget ratio in the budget allocation ratio, and the channel allocation ratio in the channel allocation structure are extracted as candidate scheduling parameters; when the channel-level bidding deviation index is less than 1, the bidding coefficient is determined as the scheduling deviation parameter, and the bidding deviation difference is recorded as 1 minus the channel-level bidding deviation index; when the channel-level exposure deviation index is less than 1, the channel budget ratio is determined as the scheduling deviation parameter, and the exposure deviation difference is recorded as 1 minus the channel-level exposure deviation index; when the channel-level response deviation index is less than 1, the channel allocation ratio is determined as the scheduling deviation parameter, and the response deviation difference is recorded as 1 minus the channel-level response deviation index, thus obtaining the scheduling deviation parameters.

[0080] When storing scheduling deviation parameters in conjunction with the aforementioned multiple performance deviation indicators, a simulation scheduling parameter record is generated for each scheduling deviation parameter. This simulation scheduling parameter record includes a channel identifier field, a time interval field, a parameter type field, a parameter value field, a response deviation indicator field, a bidding deviation indicator field, an exposure deviation indicator field, a response deviation difference field, a bidding deviation difference field, and an exposure deviation difference field. The parameter type field identifies the bidding coefficient, channel budget ratio, or channel allocation ratio, and the parameter value field contains the corresponding numerical value. The channel-level response deviation indicator, channel-level bidding deviation indicator, and channel-level exposure deviation indicator are respectively written into the response deviation indicator field, bidding deviation indicator field, and exposure deviation indicator field. The response deviation difference, bidding deviation difference, and exposure deviation difference are respectively written into the corresponding difference field. Finally, the simulation scheduling parameter record is written into the simulation scheduling parameter storage table, completing the generation of the simulation scheduling parameters.

[0081] Furthermore, in the method provided in the application embodiments, the method further includes comparing and analyzing the simulated delivery effect dataset with the target effect data item by item, and calculating multiple effect deviation indicators, and also includes: User behavior simulation analysis is performed based on the simulated campaign performance dataset and the target performance data to generate simulated response probability parameters; the ratio of the simulated response probability parameters to the preset response probability parameters is calculated to generate a response deviation index; market bidding simulation analysis is performed based on the simulated campaign performance dataset and the target performance data to generate simulated bidding conversion rate parameters; the ratio of the simulated bidding conversion rate parameters to the preset simulated bidding conversion rate parameters is calculated to generate a bidding deviation index; channel response simulation analysis is performed based on the simulated campaign performance dataset and the target performance data to generate simulated exposure parameters; the ratio of the simulated exposure parameters to the preset exposure parameters is calculated to generate an exposure deviation index; the response deviation index, the bidding deviation index, and the exposure deviation index are integrated to construct multiple performance deviation indices.

[0082] In this embodiment, when performing user behavior simulation analysis based on the simulated campaign performance dataset and the target performance data, the click count field and conversion count field are extracted from the simulated campaign performance dataset and summarized according to the statistical period field to obtain the total number of simulated clicks and the total number of simulated conversions. The simulated response probability parameter is calculated using a probability calculation formula, which is equal to the total number of simulated conversions divided by the total number of simulated clicks. Simultaneously, the target click count field and target conversion count field are extracted from the target performance data and summarized according to the same statistical period field to obtain the total number of target clicks and the total number of target conversions. The preset response probability parameter is calculated using the same probability calculation formula, which is equal to the total number of target conversions divided by the total number of target clicks. Subsequently, the response deviation index is calculated, which is equal to the simulated response probability parameter divided by the preset response probability parameter.

[0083] When conducting market bidding simulation analysis based on the simulated campaign performance dataset and target performance data, the number of bids and the number of transactions are extracted from the simulated campaign performance dataset and summarized according to the statistical period field to obtain the total number of simulated bids and the total number of simulated transactions. The simulated bidding conversion rate parameter is calculated using a proportional calculation formula, which equals the total number of simulated transactions divided by the total number of simulated bids. Simultaneously, the target bid count and target transaction count fields are extracted from the target performance data and summarized according to the same statistical period field to obtain the total number of target bids and the total number of target transactions. The preset simulated bidding conversion rate parameter is calculated using the same proportional calculation formula, which equals the total number of target transactions divided by the total number of target bids. Finally, the bidding deviation index is calculated, which equals the simulated bidding conversion rate parameter divided by the preset simulated bidding conversion rate parameter.

[0084] When performing channel response simulation analysis based on the simulated campaign performance dataset and the target performance data, the channel identifier field and the number of impressions field are extracted from the simulated campaign performance dataset, and the results are grouped and summed according to the channel identifier field to obtain the simulated exposure parameters for each channel. At the same time, the target channel identifier field and the target number of impressions field are extracted from the target performance data, and the results are grouped and summed according to the channel identifier field to obtain the preset exposure parameters for each channel. Then, the data is matched one by one according to the same channel identifier field, and the exposure deviation index is calculated. The exposure deviation index is equal to the simulated exposure parameters divided by the preset exposure parameters.

[0085] After generating response deviation, bidding deviation, and exposure deviation metrics separately, these metrics are integrated according to the statistical period field and channel identifier field, and written into a unified data record in a fixed field order to form multiple performance deviation metrics that include response deviation, bidding deviation, and exposure deviation metrics fields.

[0086] In summary, the embodiments of this application have at least the following technical effects: This application collects multi-source advertising data, processes and integrates it in batches to obtain real-time user behavior characteristics; it introduces historical conversion pattern data, combines the real-time user behavior characteristics with historical conversion pattern data for multi-objective optimization, and constructs a preliminary scheduling strategy; it sets advertising placement constraints and executes the preliminary scheduling strategy to optimize resource allocation and formulate a resource scheduling plan; it simulates the execution of the resource scheduling plan to track the effect, obtains simulation scheduling parameters for incremental updates, and constructs a closed-loop optimization report for advertising resource scheduling. This invention solves the technical problems of existing technologies where advertising resource allocation lacks a dynamic optimization mechanism and is difficult to accurately schedule based on real-time data. By integrating real-time user behavior characteristics and historical conversion pattern data for multi-objective optimization and constructing a closed-loop scheduling mechanism, it achieves the technical effect of improving the accuracy of advertising resource allocation and the conversion effect of advertising placement.

[0087] Example 2 is based on the same inventive concept as the advertising resource scheduling and management method based on big data analysis in the previous examples, such as... Figure 2 As shown, this application provides an advertising resource scheduling and management system based on big data analysis. The system and method embodiments in this application are based on the same inventive concept. The system includes: The feature acquisition module 11 is used to collect multi-source advertising data, process and merge it in batches to obtain real-time user behavior features; the multi-objective optimization module 12 is used to introduce historical conversion pattern data, combine the real-time user behavior features with the historical conversion pattern data to perform multi-objective optimization, and construct a preliminary scheduling strategy; the resource allocation optimization module 13 is used to set advertising placement constraints, execute the preliminary scheduling strategy to optimize resource allocation, and formulate a resource scheduling scheme; the effect tracking module 14 is used to simulate the execution of the resource scheduling scheme to track the effect, obtain simulation scheduling parameters for incremental updates, and construct a closed-loop optimization report for advertising placement resource scheduling.

[0088] Furthermore, the system is also used to implement the following functions: Multiple granularity parameters are set based on multiple advertising platforms, and multiple granular sliding windows are set according to these parameters. These granular sliding windows include fine-grained windows, medium-grained windows, and coarse-grained windows. Click intensity data is obtained by calculating user clicks based on the fine-grained window. Browsing interest data is obtained by tracking user preferences based on the medium-grained window. Activity analysis is performed on users based on the coarse-grained window to identify multiple activity periods. The click intensity data, browsing interest data, and multiple activity periods are processed in batches to generate batch datasets, which include micro-batch data and macro-batch data. Streaming computation is performed on the micro-batch data to generate real-time user behavior indicators, and batch computation is performed on the macro-batch data to generate historical user behavior pattern indicators. The real-time user behavior indicators and historical behavior pattern indicators are then feature-concatenated and fused to obtain feature fusion data. The feature fusion data is then normalized to generate the real-time user behavior features.

[0089] Furthermore, the system is also used to implement the following functions: Multiple users are categorized and identified, and storage is allocated according to these user identifiers to determine multiple feature storage spaces. A feature fusion mapping table is constructed based on these multiple feature storage spaces. The real-time behavior indicators of the users are written into the real-time feature regions of the multiple feature storage spaces for feature analysis to construct real-time behavior features. Based on the historical behavior pattern indicators, the historical feature regions of the multiple feature storage spaces are retrieved for feature analysis to construct historical behavior features. The real-time behavior features and the historical behavior features are dimensionally aligned according to the feature fusion mapping table, and the real-time behavior features and historical behavior features are horizontally concatenated according to the alignment results to generate a first fused feature vector. Feature conflict identification is performed on the first fused feature vector, and conflicting terms are extracted and removed to generate a second fused feature vector. Cross-validation is performed based on the second fused feature vector to generate feature fusion data.

[0090] Furthermore, the system is also used to implement the following functions: The system retrieves the historical conversion dataset for advertising, groups it according to multiple user identifiers, and constructs complete historical behavior time series for multiple users. It then segments these complete historical behavior time series to obtain complete path sequences. Frequent sequence pattern mining and filtering are performed on these complete path sequences to identify typical conversion paths. Path feature analysis is performed on these typical conversion paths to generate path efficiency indicators. Channel conversion preference analysis is conducted on the historical behavior of multiple users to obtain user channel preference characteristics. The time-segment conversion distribution of the historical behavior of multiple users is also analyzed to obtain user time-segment response characteristics. Finally, the typical conversion paths, path efficiency indicators, user channel preference characteristics, and user time-segment response characteristics are combined and encapsulated to construct the historical conversion pattern data.

[0091] Furthermore, the system is also used to implement the following functions: The real-time user behavior features are aligned with the historical conversion pattern data to generate a feature alignment result. Based on the feature alignment result, the real-time user behavior features and the historical conversion pattern data are concatenated to generate a joint feature vector. A joint optimization objective is set, and multi-objective optimization is performed based on the joint feature vector to generate a multi-objective strategy optimization result. Multiple rounds of deep reinforcement iterative learning are performed according to the multi-objective strategy optimization result to generate multi-objective benefit parameters. The multi-objective benefit parameters are combined to identify multiple strategy schemes. The multiple strategy schemes are traversed for screening and optimization to determine the preliminary scheduling strategy.

[0092] Furthermore, the system is also used to implement the following functions: The real-time user behavior features are used as index keys to traverse and search the historical conversion pattern data, generating search results, which are either present or absent. When a search result is present, the historical behavior features and real-time user behavior features are aligned along the time dimension, and feature matching results are extracted. When a search result is absent, a new user is generated, and the historical conversion pattern data is traversed for global analysis to extract global feature parameters. Based on the feature matching results or the global feature parameters, the real-time user behavior features and the historical behavior features are concatenated to obtain paired features. A first categorical feature is extracted based on the real-time user behavior features, and a second categorical feature is extracted from the historical behavior features according to the first categorical feature, where the first categorical feature and the second categorical feature have a corresponding relationship. The first categorical feature and the second categorical feature are cross-combined to generate cross feature terms. The paired features and cross feature terms are combined to generate the joint feature vector.

[0093] Furthermore, the system is also used to implement the following functions: Real-time bidding data is introduced, and environmental analysis is performed based on the real-time bidding data according to advertising channels to generate real-time market environment parameters. The initial scheduling strategy is executed according to the advertising placement constraints to perform encoding conversion, obtaining multiple encoding populations, each containing multiple candidate resource allocation schemes. Fitness is calculated on the multiple candidate resource allocation schemes based on the real-time market environment parameters to obtain multi-dimensional evaluation fitness. The multiple encoding populations are traversed and sorted in descending order according to the multi-dimensional evaluation fitness to obtain a profit ranking sequence. Based on the profit ranking sequence, multiple encoding populations are randomly selected to determine multiple competitive groups, and cyclic cross-matching is performed to generate multiple offspring parameters. Mutation iteration evaluation is performed based on the multiple offspring parameters to construct a target population for individual decoding and reconstruction, obtaining multiple analytical parameters. The multiple analytical parameters are integrated to construct the resource scheduling scheme.

[0094] Furthermore, the system is also used to implement the following functions: An advertising simulation environment is constructed, and the resource scheduling scheme is mapped to the advertising simulation environment for simulated advertising, generating a simulated advertising effect dataset. Historical resource scheduling effects are introduced for expected analysis, and target effect data is set. The simulated advertising effect dataset and the target effect data are compared and analyzed item by item to calculate multiple effect deviation indicators. Based on the multiple effect deviation indicators, deviation attribution is performed to determine scheduling deviation parameters. The scheduling deviation parameters are combined with the multiple effect deviation indicators and stored in association to generate the simulation scheduling parameters.

[0095] Furthermore, the system is also used to implement the following functions: User behavior simulation analysis is performed based on the simulated campaign performance dataset and the target performance data to generate simulated response probability parameters; the ratio of the simulated response probability parameters to the preset response probability parameters is calculated to generate a response deviation index; market bidding simulation analysis is performed based on the simulated campaign performance dataset and the target performance data to generate simulated bidding conversion rate parameters; the ratio of the simulated bidding conversion rate parameters to the preset simulated bidding conversion rate parameters is calculated to generate a bidding deviation index; channel response simulation analysis is performed based on the simulated campaign performance dataset and the target performance data to generate simulated exposure parameters; the ratio of the simulated exposure parameters to the preset exposure parameters is calculated to generate an exposure deviation index; the response deviation index, the bidding deviation index, and the exposure deviation index are integrated to construct multiple performance deviation indices.

[0096] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for scheduling and managing advertising resources based on big data analysis, characterized in that, The method includes: Collect multi-source advertising data, process and integrate it in batches to obtain real-time user behavior characteristics; By introducing historical conversion pattern data, the real-time user behavior characteristics are combined with historical conversion pattern data for multi-objective optimization, and a preliminary scheduling strategy is constructed. Set advertising placement constraints, execute the preliminary scheduling strategy to optimize resource allocation, and formulate a resource scheduling plan. The resource scheduling scheme is simulated and its effects are tracked. The simulation scheduling parameters are obtained and incrementally updated to construct a closed-loop optimization report for advertising resource scheduling.

2. The advertising resource scheduling and management method based on big data analysis as described in claim 1, characterized in that, The method involves collecting multi-source advertising data, processing and integrating it in batches to obtain real-time user behavior characteristics. Multiple granularity parameters are set based on multiple advertising platforms, and multiple granularity sliding windows are set according to the multiple granularity parameters. The multiple granularity sliding windows include fine-grained windows, medium-grained windows, and coarse-grained windows. Click intensity data is obtained by calculating the user's clicks based on the fine-grained window. Based on the medium-granularity window, user preferences are tracked to obtain browsing interest data; Based on the coarse-grained window, user activity is analyzed to identify multiple activity periods; The click intensity data, the browsing interest data, and the multiple activity time periods are processed in batches to generate batch datasets, which include micro-batch data and macro-batch data. The micro-batch data is processed by streaming computation to generate real-time user behavior metrics, and the macro-batch data is processed by batch computation to generate historical user behavior pattern metrics. The real-time user behavior metrics and the historical behavior pattern metrics are spliced ​​and fused to obtain feature fusion data; The fused feature data is normalized to generate the real-time user behavior features.

3. The advertising resource scheduling and management method based on big data analysis as described in claim 2, characterized in that, The method involves fusing the real-time user behavior metrics with the historical behavior pattern metrics to obtain feature fusion data, including: Multiple users are traversed and classified, and storage is allocated to multiple users according to multiple user identifiers to determine multiple feature storage spaces. A feature fusion mapping table is constructed based on the multiple feature storage spaces. The user's real-time behavior metrics are written into the real-time feature region of the multiple feature storage spaces for feature analysis to construct real-time behavior features; Based on the historical behavior pattern indicators, historical feature regions of the multiple feature storage spaces are retrieved for feature analysis to construct historical behavior features; The real-time behavior features and the historical behavior features are aligned dimensionally according to the feature fusion mapping table. Based on the alignment result, the real-time behavior features and the historical behavior features are horizontally concatenated to generate the first fused feature vector. The first fused feature vector is subjected to feature conflict identification, conflict terms are extracted and feature removal is performed, and a second fused feature vector is generated. Cross-validation is performed based on the second fused feature vector to generate feature fusion data.

4. The advertising resource scheduling and management method based on big data analysis as described in claim 3, characterized in that, The process and methods for constructing historical transformation pattern data include: Retrieve the historical conversion dataset of the advertisement, group the historical conversion dataset of the advertisement according to the multiple user identifiers, and construct a complete historical behavior time series of multiple users; The complete historical behavior time series of multiple users are segmented to obtain the complete path sequence; Frequent sequence pattern mining and filtering are performed on the complete path sequence to determine typical transformation paths; Path feature analysis is performed on the typical conversion path to generate path efficiency indicators; By iterating through multiple users and analyzing their historical behavior to determine their channel conversion preferences, we can obtain user channel preference characteristics. By iterating through the time-period conversion distribution of multiple users' historical behaviors, the user's time-period response characteristics can be obtained. The typical conversion path, the path efficiency index, the user channel preference characteristics, and the user time period response characteristics are combined and encapsulated to construct the historical conversion pattern data.

5. The advertising resource scheduling and management method based on big data analysis as described in claim 1, characterized in that, The method involves combining the real-time user behavior characteristics with the historical conversion pattern data for multi-objective optimization to construct a preliminary scheduling strategy, including: The real-time user behavior features are aligned with the historical conversion pattern data to generate feature alignment results. Based on the feature alignment result, the real-time user behavior features and the historical conversion pattern data are concatenated to generate a joint feature vector; A joint optimization objective is set, and multi-objective optimization is performed based on the joint feature vector to generate multi-objective strategy optimization results. Based on the optimization results of the multi-objective strategy, multiple rounds of deep reinforcement iterative learning are performed to generate multi-objective benefit parameters. Based on the combination of the multi-objective benefit parameters, multiple strategy schemes are identified. The multiple strategy options are traversed and optimized to determine the initial scheduling strategy.

6. The advertising resource scheduling and management method based on big data analysis as described in claim 5, characterized in that, Based on the feature alignment result, the real-time user behavior features are concatenated with the historical conversion pattern data to generate a joint feature vector. The method includes: The real-time user behavior features are used as index keys to traverse and search the historical conversion pattern data to generate search results, which are either present or absent. When the search result is a result, the historical behavior features and real-time user behavior features are aligned in the time dimension, and the feature matching result is extracted. When the search result is no result, new user historical conversion pattern data is generated for global analysis to extract global feature parameters. Based on the feature matching result or the global feature parameter, the real-time user behavior feature and the historical behavior feature are concatenated to obtain paired features; Based on the real-time user behavior features, a first categorical feature is extracted, and a second categorical feature is extracted from the historical behavior features according to the first categorical feature. The first categorical feature and the second categorical feature have a corresponding relationship. The first categorical feature and the second categorical feature are cross-combined to generate cross feature terms; The paired features and cross features are combined to generate the joint feature vector.

7. The advertising resource scheduling and management method based on big data analysis as described in claim 1, characterized in that, The method includes setting advertising placement constraints, executing the preliminary scheduling strategy to optimize resource allocation, and formulating a resource scheduling plan, including: Real-time bidding data is introduced, and environmental analysis is performed based on the real-time bidding data according to advertising channels to generate real-time market environment parameters. The initial scheduling strategy is executed according to the advertising placement constraints to perform encoding conversion, thereby obtaining multiple encoding populations, which contain multiple candidate resource allocation schemes. Based on the real-time market environment parameters, the fitness of the multiple candidate resource allocation schemes is calculated to obtain a multi-dimensional evaluation fitness. The multidimensional evaluation fitness is used to traverse multiple encoded races and sort them in descending order to obtain the profit ranking sequence; Based on the profit ranking sequence, multiple coded populations are randomly selected to determine multiple competitive groups, and cyclic cross-matching is performed to generate multiple offspring parameters; Based on the multiple offspring parameters, mutation iteration evaluation is performed, a target population is constructed, and individual decoding and reconstruction are performed to obtain multiple analytical parameters; The resource scheduling scheme is constructed by integrating the multiple parsing parameters.

8. The advertising resource scheduling and management method based on big data analysis as described in claim 1, characterized in that, The method for simulating the execution of the resource scheduling scheme and tracking its effects to obtain simulation scheduling parameters includes: Construct an advertising delivery simulation environment, map the resource scheduling scheme to the advertising delivery simulation environment for simulated delivery, and generate a simulated delivery effect dataset; Introduce historical resource scheduling effects for expected analysis and set target effect data; The simulated delivery effect dataset is compared and analyzed item by item with the target effect data to calculate multiple effect deviation indicators; Based on the aforementioned multiple performance deviation indicators, deviation sources are traced and attributed to determine scheduling deviation parameters. The scheduling deviation parameters are associated and stored with the multiple effect deviation indicators to generate the simulation scheduling parameters.

9. The advertising resource scheduling and management method based on big data analysis as described in claim 8, characterized in that, The simulated delivery effect dataset is compared and analyzed item by item with the target effect data to calculate multiple effect deviation indicators. The method includes: Based on the simulated delivery effect dataset and the target effect data, user behavior simulation analysis is performed to generate simulated response probability parameters; Calculate the ratio of the simulated response probability parameter to the preset response probability parameter to generate a response deviation index; Based on the simulated campaign performance dataset and the target performance data, a market bidding simulation analysis is performed to generate simulated bidding conversion rate parameters. Calculate the ratio of the simulated bidding conversion rate parameter to the preset simulated bidding conversion rate parameter to generate a bidding deviation index; Based on the simulated delivery effect dataset and the target effect data, channel response simulation analysis is performed to generate simulated exposure parameters. Calculate the ratio of the simulated exposure parameter to the preset exposure parameter to generate an exposure deviation index; The response deviation index, the bidding deviation index, and the exposure deviation index are integrated to construct multiple performance deviation indexes.

10. An advertising resource scheduling and management system based on big data analysis, characterized in that, The system is used to execute the advertising delivery resource scheduling and management method based on big data analysis as described in any one of claims 1-9, and the system includes: The feature acquisition module is used to collect multi-source advertising data, process and merge it in batches to obtain real-time user behavior features; The multi-objective optimization module is used to introduce historical conversion pattern data, combine the real-time user behavior characteristics with the historical conversion pattern data to perform multi-objective optimization, and construct a preliminary scheduling strategy. The resource allocation optimization module is used to set advertising placement constraints, execute the preliminary scheduling strategy to optimize resource allocation, and formulate a resource scheduling plan. The effect tracking module is used to simulate the execution of the resource scheduling plan to track the effect, obtain simulation scheduling parameters for incremental updates, and construct a closed-loop optimization report for advertising placement resource scheduling.

Images