An advertisement effect prediction method and system based on delivery log data fusion

Abstract

The application discloses an advertisement effect prediction method and system based on delivery log data fusion, relates to the technical field of advertisement effect prediction, and comprises the following steps: determining a plurality of delivery path information of a target advertisement, and obtaining corresponding delivery log data to form a data set; when the advertisement is loaded and displayed, recognizing an actual entering path and mode of a user, recording path entering times and user operation data; synchronously collecting advertisement display quality information, correcting the operation data, and generating corrected data; combining the corrected data with the path entering record, constructing a path entering representation sequence, establishing a historical entering times sequence based on the sequence, predicting next period entering times, generating an initial effect prediction result in combination with the corrected data; obtaining competitive advertisement exposure change data, correcting the initial result in terms of competitiveness, obtaining a final advertisement effect prediction result, realizing accurate prediction of the advertisement effect, and improving the accuracy and reliability of advertisement delivery decision.

Description

Technical Field

[0001] This invention relates to the field of advertising effectiveness prediction technology, specifically to an advertising effectiveness prediction method and system based on the fusion of campaign log data. Background Technology

[0002] With the rapid development of the internet and mobile devices, digital advertising has become a crucial marketing tool for businesses. Advertisers hope to gain more user attention and conversions through precise targeting, thereby improving advertising effectiveness. However, in actual campaigns, users may enter ads through multiple paths, such as search engines, QR code scans, and clicks on recommended ads. User behavior and ad display quality vary significantly across these paths. Traditional methods for evaluating and predicting advertising effectiveness typically rely on single metrics such as impressions, clicks, or conversion rates, lacking a comprehensive analysis of user behavior, display quality, and operational differences across multiple paths. This results in inaccurate predictions and an inability to provide reliable guidance for advertising strategies.

[0003] Furthermore, existing methods suffer from insufficient historical trend analysis and a lack of consideration for the competitive environment. Ad entry frequency and user behavior may fluctuate across different campaign periods, and changes in competitor ad exposure can also affect the target ad's performance. Existing methods fail to effectively integrate campaign logs, user behavior, and display quality information, and do not incorporate competitor ad data for correction. This leads to predictions deviating from actual campaign performance, reducing the efficiency and accuracy of ad optimization decisions. Therefore, there is an urgent need for an ad performance prediction method and system based on campaign log data fusion to achieve multi-path user behavior identification, operation correction, historical trend analysis, and competition level correction, thereby improving the accuracy and reliability of ad performance prediction. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a method and system for predicting advertising effectiveness based on the fusion of campaign log data. This technical solution resolves the shortcomings of existing methods mentioned in the background section, such as insufficient historical trend analysis and lack of consideration for the competitive environment. In different campaign periods, the number of ad entries and user behavior may fluctuate, and changes in the exposure of competing ads can also affect the effectiveness of the target ad. Existing methods fail to effectively integrate campaign logs, user behavior, and display quality information, and do not incorporate competitive ad data for correction, resulting in prediction results deviating from actual campaign performance and reducing the efficiency and accuracy of ad campaign optimization.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An advertising performance prediction method based on the fusion of delivery log data includes: Determine the multiple ad delivery paths corresponding to the target ad and obtain the delivery log data for each path to form a delivery log dataset; When an ad is loaded and displayed, the user's entry path and method into the ad are identified based on the ad delivery log dataset, and the number of times the corresponding path is entered and the user operation data generated during the ad display period are recorded to obtain the path entry record and display period operation data. During the ad display period, the ad display quality information is collected synchronously to form display quality data. Then, the operation data during the display period is corrected to obtain corrected operation data. The correction operation data is combined with the path entry record to form the path entry structure data, and the path entry representation sequence of the target advertisement is generated. Based on the path entry representation sequence, a historical entry frequency sequence of the target advertisement is established, and trend analysis is performed on the historical entry frequency sequence to obtain the entry frequency prediction result for the next period. Then, combined with the correction operation data, the initial effect prediction result of the target advertisement is generated. Based on preset monitoring channels, data on changes in the exposure of competing advertisements are obtained, and the initial effect prediction results are corrected for competitiveness to obtain the final advertising effect prediction results of the target advertisement.

[0006] In an optional embodiment, when the advertisement is loaded and displayed, the process of identifying the user's entry path and method into the advertisement based on the delivery log dataset, and recording the corresponding number of entry paths and user operation data generated during the advertisement display period, to obtain path entry records and display period operation data, specifically includes: Obtain the user's access identification information, which includes terminal identifier, network address, access timestamp, and browser type; Obtain the delivery path information corresponding to the target advertisement, and extract the path fields related to the delivery path from the delivery log dataset. The delivery path information includes the search entry, QR code entry, and recommendation click entry. Based on the ad delivery log dataset, multi-dimensional features of access behavior are constructed using access identification information. The ad delivery path field is cross-validated to determine the actual ad delivery path corresponding to the user and to record the specific path and method of entry for the user. For search entry points, identify user behavior by entering keywords or clicking on search results to access target ads; For QR code entry points, identify the behavior of users entering the target advertisement by scanning the advertisement's QR code; For recommended ad entry points, identify user behavior by clicking on a recommended ad on the page to access the target ad. Based on the deployment log dataset, access interval information, loading status information, and network status information are extracted for access validity determination, and access validity determination rules are constructed. Multiple access records under the same delivery path are accumulated, and invalid and abnormal accesses in the access logs are filtered according to the access validity judgment rules to generate the valid path entry count for that delivery path. During the ad display period, user operation data is collected in real time, including the time users stay in the ad area, the number of purchases made by users, the time users initiate customer service inquiries, and the number of times users switch between the front and back ends. The number of valid path entries is matched with the entry method corresponding to each path, and then linked and integrated with the display period operation data collected during the ad display period according to the entry event to form path entry records and display period operation data.

[0007] In an optional embodiment, the step of synchronously collecting advertisement display quality information during the advertisement display period to form display quality data, and then correcting the display period operation data to obtain corrected operation data, specifically includes: Obtain the start time and completion time of ad loading to get the ad loading time information; Obtain data on the actual display area of ​​the advertisement on the terminal screen and data on the preset display area of ​​the advertisement, and obtain the exposure area ratio information based on the ratio between the display area data and the preset area data; Obtain status data during ad playback, record the number of playback interruptions and their duration to obtain playback stuttering information; Based on loading time information, exposure area occupancy information, and playback stuttering information, display quality information is generated for each ad display. Based on the loading time information in the display quality information, the dwell time and occurrence time of various interactive behaviors in the display period operation data are corrected according to the loading delay, and the invalid segments during the loading period are deducted to obtain the preliminary correction data. Based on the exposure area ratio information in the display quality information, the dwell time, purchase frequency and consultation behavior in the preliminary correction data are corrected according to the corresponding area ratio to obtain the exposure correction data; Based on the playback stuttering information in the display quality information, the dwell time, number of purchases, consultation time, and number of front-end / back-end switching in the exposure correction data are filtered for effectiveness and corrected for time according to the stuttering segment to form the final correction data, which serves as the correction operation data.

[0008] In an optional embodiment, the step of combining the correction operation data with the path entry record to form path entry structure data and generating a path entry representation sequence for the target advertisement specifically includes: Based on the path entry information of each delivery path and the corresponding correction operation data, a path entry data set is formed. The path entry data set is divided according to different entry paths to generate first path sub-information, second path sub-information and third path sub-information. The first path sub-information, second path sub-information and third path sub-information all include the corresponding path entry times, entry methods and correction operation data during the ad display period. Based on the first path sub-information, the second path sub-information, and the third path sub-information, the number of entries and the entry method in each path sub-information are expanded into individual entry records, and then associated with the corresponding correction operation data item by item to obtain a list of entry records organized by path. Based on the list of entry records, each entry record is mapped to its corresponding correction operation data to form a behavior mapping table that can be looked up by record index; Using the behavior mapping table as input, the mapping records under the same path are arranged sequentially according to the entry time to obtain an ordered behavior flow that reflects the entry change process in the same path; Based on ordered behavior flow, multiple ordered behavior flows along the same path are merged according to the path dimension to generate a primary path entry sequence. Based on the primary path entry sequence, the number of entries, entry methods and correction operation data in the sequence are scanned item by item, and the differences between adjacent records are marked according to the preset change conditions to form sequence annotation data that can reflect the change relationship between records. Based on sequence labeling data, the primary path entry sequence is structured and organized, and a path entry representation sequence is generated according to the record order.

[0009] In an optional embodiment, the step of establishing a historical entry sequence of the target advertisement based on the path entry representation sequence, performing trend analysis on the historical entry sequence to obtain the entry prediction result for the next period, and then combining it with correction operation data to generate the initial effect prediction result of the target advertisement, specifically includes: Based on the path entry representation sequence, read the time information and entry count information corresponding to each record in the sequence; Obtain the ad display period, divide the read time information, and obtain the entry count statistics for each display period; Based on the statistics of the number of entries, they are arranged in chronological order according to the display period to form a historical entry sequence; Based on the historical entry count sequence, the entry count between each two adjacent display periods is obtained in chronological order; Based on the number of entries obtained, the entry difference between adjacent display periods is calculated to obtain the entry change in each time period; The sign of the change in the number of entries is determined to distinguish between an upward trend and a downward trend in the number of entries, thus obtaining the direction of change; After determining the trend direction, analyze the magnitude of the change in the number of entries based on the value of the change. By combining the direction and magnitude of the change, determine the trend of the number of entries in the next display cycle; Based on the changing trend, the number of entries in the next display cycle is predicted, and the entry number prediction result for the next cycle is obtained. Combined with the path information, the number of entries for each path in the next display cycle is predicted. Read the correction operation data and extract the corresponding fields of ad dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching according to each delivery path; Based on the entry count statistics for each display period and combined with path information, the historical entry count statistics for each path are obtained. Based on the historical entry count of each path and the corresponding period of the extracted fields of ad dwell time, purchase count, customer service inquiry count and front-end / back-end switching count, a linear regression model is constructed with the historical entry count of each path as the independent variable and the operation behavior field as the dependent variable. The historical entry counts for each path are input into the regression model to obtain the predicted values ​​of the operation data corresponding to each path. The predicted values ​​are then compared with the corrected operation data, and the regression model parameters are adjusted based on the difference to form a corrected regression model. The number of times each path is predicted to enter the next display cycle is input into the calibrated regression model to obtain the prediction operation data. The prediction operation data is then normalized to obtain the comprehensive effect index. The overall performance index is used as the initial performance prediction result for the target advertisement; The formula for calculating the comprehensive effect index is as follows: ; In the formula, , , and The first The values ​​for ad dwell time, purchase frequency, customer service inquiry frequency, and front-end / back-end switching frequency for each path are normalized. This represents the total number of advertising delivery paths.

[0010] In an optional embodiment, the step of acquiring exposure change data of competing advertisements based on preset monitoring channels and correcting the initial effect prediction results for competition intensity to obtain the final advertising effect prediction result of the target advertisement specifically includes: Based on preset monitoring channels, obtain data on the changes in exposure of competing ads that are in the same advertising environment as the target ad in different time periods; The data on changes in competitive ad impressions were organized in chronological order to obtain a sequence of impression changes for each time period. Based on the sequence of competitive ad exposure changes, we identify the rising, falling and stable intervals of exposure and assign corresponding competitive pressure level indicators to each interval. Based on the competitive pressure level index, the initial effect prediction results of the target advertisement are corrected for the degree of competition to obtain the final advertising effect prediction results of the target advertisement. The formula for calculating the final advertising effect prediction result is as follows: ; In the formula, The final advertising effect prediction result for the target advertisement. The initial performance prediction results for the target advertisement. For the first The level of competitive pressure over a given period of time. This represents the total number of time periods.

[0011] Furthermore, an advertising effectiveness prediction system based on the fusion of delivery log data is proposed to implement the advertising effectiveness prediction method described above, characterized by comprising: The data acquisition module is used to obtain the target advertisement's delivery path information and corresponding delivery log data, and to collect the user's display period operation data and advertisement display quality information in real time. The path identification and integration module is used to identify the actual path and method of users entering the advertisement based on the delivery log data, calculate the number of valid path entries, and integrate them with the display period operation data to form path entry records and display period operation data. An operation correction module is used to correct the operation data during the display period based on information such as ad loading time, exposure area ratio, and playback stuttering, and to generate corrected operation data. A path representation construction module is used to combine correction operation data with path entry records to construct path entry structure data and generate a path entry representation sequence for the target advertisement. The effect prediction module is used to generate a historical entry frequency sequence based on the path entry representation sequence, perform trend analysis to obtain the entry frequency prediction result for the next period, and combine the correction operation data to generate the initial effect prediction result of the target advertisement. The competition correction module is used to acquire the exposure change data of competing advertisements and perform competition correction on the initial effect prediction results to obtain the final advertising effect prediction results of the target advertisement.

[0012] In an optional embodiment, the data acquisition module includes: The delivery log collection unit is used to acquire multiple delivery path information of the target advertisement and the corresponding delivery log data. The user operation collection unit is used to collect in real time the user's dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching during the advertisement display period; The display quality acquisition unit is used to simultaneously collect information on ad loading time, exposure area ratio, and playback stuttering.

[0013] In an optional embodiment, the path representation building module includes: The path entry data combination unit is used to form a path entry data set based on the path entry information of each delivery path and the corresponding correction operation data. The path sub-information generation unit is used to divide the path entry data set into sub-information according to different entry paths, and includes the number of entry times, entry method and correction operation data during the advertising display period for each path. A behavior mapping table generation unit is used to expand the number of entries and the entry method in the path sub-information into entry records one by one, and associate them with the corresponding correction operation data to generate an indexable behavior mapping table. The path entry sequence generation unit is used to generate the final path entry representation sequence by sorting and structuring according to the path dimension based on the behavior mapping table.

[0014] In an optional embodiment, the effect prediction module includes: The historical entry count generation unit is used to read the time information and entry count information corresponding to each record based on the path entry representation sequence, and generate a historical entry count sequence according to the advertising display cycle. The trend analysis unit is used to calculate the entry difference between adjacent display periods based on the historical entry count sequence, determine the direction and magnitude of the change in the number of entries, and determine the trend of the number of entries in the next display period. The initial effect calculation unit is used to input the number of times each path is predicted for the next cycle into the calibrated linear regression model to obtain the prediction operation data corresponding to each path. The prediction operation data is then normalized, and a comprehensive effect index is calculated, including the ad dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching times, as the initial effect prediction result of the target ad.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes an advertising performance prediction method and system based on the fusion of campaign log data. By integrating log data from multiple campaign paths of the target ad, user operation behavior data, and ad display quality information, it achieves accurate identification of the user's actual entry path and method, corrects the operation data during the display period, constructs a path entry representation sequence, and predicts the number of entries in the next cycle based on trend analysis of historical entry frequency sequences. Simultaneously, it combines data on changes in the exposure of competing ads to correct for competitiveness, thereby obtaining the final performance prediction result of the target ad. This invention can improve the accuracy and reliability of advertising performance prediction, reduce prediction deviations caused by differences in user paths, uneven display quality, or changes in the competitive environment, and provide a scientific basis for advertising optimization and decision-making. Attached Figure Description

[0016] Figure 1 This is a flowchart of an advertising effect prediction method based on the fusion of delivery log data proposed in this invention; Figure 2 This is a flowchart of the operation correction process in this invention; Figure 3 This is a flowchart of the competition correction process in this invention; Figure 4 This is a system framework diagram of an advertising effect prediction system based on the fusion of delivery log data proposed in this invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 - Figure 4 As shown in the figure, an advertising effect prediction method based on the fusion of delivery log data in an embodiment of the present invention includes: Determine the multiple ad delivery paths corresponding to the target ad and obtain the delivery log data for each path to form a delivery log dataset; When an ad is loaded and displayed, the user's entry path and method into the ad are identified based on the ad delivery log dataset, and the number of times the corresponding path is entered and the user operation data generated during the ad display period are recorded to obtain the path entry record and display period operation data. During the ad display period, the ad display quality information is collected synchronously to form display quality data. Then, the operation data during the display period is corrected to obtain corrected operation data. The correction operation data is combined with the path entry record to form the path entry structure data, and the path entry representation sequence of the target advertisement is generated. Based on the path entry representation sequence, a historical entry frequency sequence of the target advertisement is established, and trend analysis is performed on the historical entry frequency sequence to obtain the entry frequency prediction result for the next period. Then, combined with the correction operation data, the initial effect prediction result of the target advertisement is generated. Based on preset monitoring channels, data on changes in the exposure of competing advertisements are obtained, and the initial effect prediction results are corrected for competitiveness to obtain the final advertising effect prediction results of the target advertisement.

[0019] Furthermore, when the advertisement is loaded and displayed, the system identifies the user's entry path and method into the advertisement based on the delivery log dataset, and records the corresponding number of entry paths and user operation data generated during the advertisement display period, obtaining path entry records and display period operation data, specifically including: Obtain the user's access identification information, which includes terminal identifier, network address, access timestamp, and browser type; Obtain the delivery path information corresponding to the target ad and extract path fields related to the delivery path from the delivery log dataset. The delivery path information includes the search entry, QR code entry, and recommendation click entry. Based on the ad delivery log dataset, multi-dimensional features of access behavior are constructed using access identification information. The ad delivery path field is cross-validated to determine the actual ad delivery path corresponding to the user and to record the specific path and method of entry for the user. For search entry points, identify user behavior by entering keywords or clicking on search results to access target ads; For QR code entry points, identify the behavior of users entering the target advertisement by scanning the advertisement's QR code; For recommended ad entry points, identify user behavior by clicking on a recommended ad on the page to access the target ad. Based on the deployment log dataset, access interval information, loading status information, and network status information are extracted for access validity determination, and access validity determination rules are constructed. Specifically, the first step is to obtain the user's access identification information. This information consists of a device identifier, network address, access timestamp, and browser type. The device identifier identifies the user's current device, the network address reflects the network environment from which the access originated, the access timestamp indicates the time the access occurred, and the browser type represents the user's operating environment when accessing the webpage. These access identification information collectively constitute the basic attributes of an access behavior, used for subsequent path identification and validity assessment. Next, the target advertisement's delivery path information is obtained, and path fields related to the delivery path are extracted from the delivery log dataset. The delivery path information includes three types of paths: search entry, QR code entry, and recommendation click entry. Each type of path has a corresponding path field in the delivery log, used to record the characteristic behaviors triggered by each entry point. Based on the delivery log dataset, multidimensional feature data of the access behavior is constructed using the access identification information. This multidimensional feature data includes multiple dimensions such as the user's access device attributes, network attributes, and time series attributes. Analysis of these dimensions allows for the description of the characteristic differences in user access behavior under different entry points. After constructing multidimensional features, the delivery path fields are cross-validated. Based on the correspondence of access time series, the triggering of path fields, and the consistency of access environment, the actual ad delivery path corresponding to the user is determined, and the path type and access method of the user entering the ad page are recorded.

[0020] During the path identification phase, for search entry points, the system determines whether a user triggered ad access by entering keywords or clicking search results based on the search field in the access record; for QR code entry points, the system determines whether the access behavior was triggered by QR code scanning based on the QR code scanning field in the access record; and for recommended ad click entry points, the system determines whether the access was triggered by a recommended ad click based on the recommended ad exposure and click fields in the access record. To ensure the validity of the entry path identification results, access interval information, loading status information, and network status information required for access validity determination are extracted from the campaign log dataset. Access interval information represents the time difference between adjacent access records, loading status information represents whether the page has completed loading during the access process, and network status information represents the stability and reachability of the network during the access period. Based on access interval judgment, if the time interval between two access records of the same user under the same campaign path is less than the minimum threshold, the access record is determined to be an invalid access. Based on page loading status judgment, if the access record shows that the page has not fully loaded or has failed to load, the access record is determined to be an invalid access. Based on network status judgment, if the network status corresponding to the access record is abnormal, including network interruption, severe packet loss, or latency exceeding the threshold, the access record is determined to be an invalid access. These rules enable the identification of invalid and abnormal access behaviors in access logs, providing a reliable data foundation for the accumulation of multiple accesses under the same delivery path, the generation of valid path entry counts, and the correlation of operational data during the display period.

[0021] Multiple access records under the same delivery path are accumulated, and invalid and abnormal accesses in the access logs are filtered according to the access validity judgment rules to generate the valid path entry count for that delivery path. During the ad display period, user operation data is collected in real time, including the time users stay in the ad area, the number of purchases made by users, the time users initiate customer service inquiries, and the number of times users switch between the front and back ends. The number of valid path entries is matched with the entry method corresponding to each path, and then linked and integrated with the display period operation data collected during the ad display period according to the entry event to form path entry records and display period operation data.

[0022] Specifically, when accumulating multiple access records under the same delivery path, the validity of each access record is first determined. The determination criteria include three aspects: access interval, page loading status, and network status. Access interval is used to determine if consecutive user accesses are too frequent. If the time interval between two accesses by the same user under the same delivery path is less than a preset minimum threshold, it is considered an invalid access and removed. Page loading status is used to determine if the ad page is displayed correctly. If the record shows that the page loading is incomplete or failed, the access is considered invalid. Network status is used to determine if there are any network anomalies during the access period, including network interruptions, severe packet loss, or latency exceeding a threshold; these accesses are also considered invalid. Based on these criteria, invalid and abnormal accesses are removed from the original access records, and the remaining records are the valid access records. These are then accumulated and statistically analyzed to obtain the number of valid path entries for that delivery path. During ad display, user operation data is collected in real time, including user dwell time in the ad area, number of purchases, customer service consultation time, and number of front-end / back-end switching. This data records the actual behavior of users during ad display and is used to analyze user interaction with the ad.

[0023] When matching the number of valid entry paths with the corresponding entry methods for each path, the number of entry paths is associated with the entry method based on the entry type of each path record (search entry, QR code entry, recommendation click entry). For example, if a user enters an ad three times through the search entry, these three valid entry records are marked as "search entry". During the association and integration process, each valid entry record is matched with the operation events during the ad display period, one by one. Specifically, the time interval of each entry is matched with the operation time in the display period operation data. If the operation occurs within the time interval of that entry, it is attributed to that entry record. In this way, each valid entry record corresponds to a set of corrected display period operation data, including dwell time, purchase behavior, customer service inquiries, and front-end / back-end switching, forming a complete association dataset of path entry records and display period operation data, providing an accurate data foundation for subsequent path analysis and ad performance prediction.

[0024] Furthermore, during the ad display period, ad display quality information is collected synchronously to form display quality data. This data is then used to correct the operational data during the display period, resulting in corrected operational data, which specifically includes: Obtain the start time and completion time of ad loading to get the ad loading time information; Obtain data on the actual display area of ​​the advertisement on the terminal screen and data on the preset display area of ​​the advertisement, and obtain the exposure area ratio information based on the ratio between the display area data and the preset area data; Obtain status data during ad playback, record the number of playback interruptions and their duration to obtain playback stuttering information; Based on loading time information, exposure area occupancy information, and playback stuttering information, display quality information is generated for each ad display. Based on the loading time information in the display quality information, the dwell time and occurrence time of various interactive behaviors in the display period operation data are corrected according to the loading delay, and the invalid segments during the loading period are deducted to obtain the preliminary correction data. Based on the exposure area ratio information in the display quality information, the dwell time, purchase frequency and consultation behavior in the preliminary correction data are corrected according to the corresponding area ratio to obtain the exposure correction data; Based on the playback stuttering information in the display quality information, the dwell time, number of purchases, consultation time, and number of front-end / back-end switching in the exposure correction data are filtered for effectiveness and corrected for time according to the stuttering segment to form the final correction data, which serves as the correction operation data.

[0025] Specifically, during ad display, the system first acquires the ad loading start time and completion time. The time difference between these two times yields the ad loading time information, used to measure the ad's loading performance on the terminal. Next, it acquires data on the actual display area of ​​the ad on the terminal screen and the preset display area. Based on the ratio of the actual display area to the preset display area, it calculates the exposure area ratio information, reflecting the ad's visibility on the screen. Finally, it acquires status data during ad playback, recording the number of playback interruptions and their duration, thus obtaining playback stuttering information to reflect the smoothness of ad playback. Based on the loading time, exposure area ratio, and playback stuttering information, it generates display quality information for each ad display.

[0026] In loading time correction, the loading start and finish times for each ad display are obtained, and the ad loading time is calculated. Dwell time and interaction events are adjusted to their starting point based on the loading delay, essentially shifting the operation time forward by the corresponding loading time. Simultaneously, periods during loading when users cannot interact with the ad are removed from the dwell time, and clicks, purchases, and other actions are marked to ensure the initial correction data only includes user behavior during the actual interactive time of the ad. In exposure ratio correction, the visible ratio is calculated based on the ratio of the actual display area of ​​the ad on the terminal screen to the preset display area. Dwell time, purchase behavior, and inquiry behavior in the initial correction data are adjusted according to this ratio. For example, if the ad only displays 70% of the preset area, the weight of related action events is reduced to 70% to reflect the actual degree of user interaction with the ad content. This operation ensures that the corrected data accurately reflects the user's actual interaction within the visible ad area. In playback stuttering correction, the number of playback interruptions and the duration of each interruption are obtained by analyzing playback status data. Exposure correction data is filtered and time-adjusted: if an operation occurs during a playback stuttering period, the validity of that operation event is set to low or removed; dwell time is corrected by dividing it into normal playback periods and stuttering periods, and only dwell behavior within the normal playback periods is counted. The number of foreground / background switches, purchases, and inquiries are also corrected according to the same rules to ensure that the final corrected data reflects real user behavior when the ad is playing smoothly.

[0027] Furthermore, the correction operation data is combined with the path entry records to form path entry structure data, and a path entry representation sequence for the target advertisement is generated, specifically including: Based on the path entry information of each delivery path and the corresponding correction operation data, a path entry data set is formed. The path entry data set is divided according to different entry paths to generate first path sub-information, second path sub-information and third path sub-information. The first path sub-information, second path sub-information and third path sub-information all include the corresponding path entry times, entry methods and correction operation data during the ad display period. Based on the first path sub-information, the second path sub-information, and the third path sub-information, the number of entries and the entry method in each path sub-information are expanded into individual entry records, and then associated with the corresponding correction operation data item by item to obtain a list of entry records organized by path. Based on the list of entry records, each entry record is mapped to its corresponding correction operation data to form a behavior mapping table that can be looked up by record index; Specifically, the entry information from each advertising path is combined with the corresponding correction operation data to form a complete entry data set. Each entry information entry includes the number of times a user entered the ad through that path and the entry method. The correction operation data includes the ad dwell time, purchase behavior, inquiry behavior, and number of front-end / back-end switches for each entry. The combination operation achieves a one-to-one correspondence by pairing the entry count under the same path with its corresponding correction data. The entry data set is divided according to different entry paths, generating first path sub-information, second path sub-information, and third path sub-information. The first path sub-information corresponds to records of ad entry through search entry, the second path sub-information corresponds to records of ad entry through QR code entry, and the third path sub-information corresponds to records of ad entry through recommendation click entry. Each path sub-information includes the number of entries, entry method, and correction operation data during ad display for that path. The segmentation operation filters by the path type field, grouping all entry records belonging to the same path into the same sub-information set, ensuring that subsequent analysis can process data independently by path dimension. In the entry record expansion stage, the entry count and entry method in each path sub-information are separated into individual entry records. For example, if a user enters an advertisement three times through a search entry, three separate records are generated, each containing the time, entry type, and entry method of each entry. Subsequently, each entry record is associated with its corresponding correction operation data item by item, achieving precise matching between entry behavior and operation data. The associated operations use time matching or sequence indexing to assign display-period operation events to specific entry records, ensuring that each entry record corresponds to a complete set of correction operation data. Based on the obtained path-organized list of entry records, a mapping relationship is established between each entry record and its corresponding correction operation data, forming a behavior mapping table. The behavior mapping table uses a unique record index to identify each entry and associates the corresponding correction operation data with this index, allowing each entry record to quickly find its corresponding operation information. Through this mapping table, user behavior across different paths and entry events can be easily retrieved, analyzed, and summarized, providing a reliable data foundation for generating ordered behavior flows and path entry representation sequences.

[0028] Using the behavior mapping table as input, the mapping records under the same path are arranged sequentially according to the entry time to obtain an ordered behavior flow that reflects the entry change process in the same path; Based on ordered behavior flow, multiple ordered behavior flows along the same path are merged according to the path dimension to generate a primary path entry sequence. Based on the primary path entry sequence, the number of entries, entry methods and correction operation data in the sequence are scanned item by item, and the differences between adjacent records are marked according to the preset change conditions to form sequence annotation data that can reflect the change relationship between records. Based on sequence labeling data, the primary path entry sequence is structured and organized, and a path entry representation sequence is generated according to the record order.

[0029] Understandably, in advertising data processing, the entry information and corresponding correction operation data for each delivery path are first combined to form a path entry data set. Each record includes the number of entries, the entry method, and corresponding correction operation data, such as dwell time, purchase behavior, customer service consultation time, and the number of front-end / back-end switches. The data set is then divided into different subsets based on the path type, such as search entry, QR code entry, and recommendation click entry. Each subset contains all entry records and correction data for that path. Next, the number of entries and entry methods in each subset are broken down into individual entry records, each corresponding to its correction operation data, forming a path-organized list of entry records. Then, a mapping relationship is established between each record and its correction data, forming a behavior mapping table. A unique index identifies each entry, allowing each entry record to quickly find its corresponding operation data. Based on the behavior mapping table, records under the same path are arranged sequentially by entry time, resulting in an ordered behavior flow. Each record retains the entry time, entry method, and correction operation data to reflect changes in user behavior within the same path. Multiple ordered behavior flows along the same path are merged in the time dimension to generate a primary path entry sequence. During the merging process, the recorded time order and correction operation data are kept unchanged to ensure that the sequence can cover the behavior of all users or multiple visits under the path.

[0030] In the initial path entry sequence, the number of entries, entry methods, and correction operation data are scanned one by one, and adjacent records are labeled according to preset change conditions. These preset change conditions include: changes in entry methods (e.g., switching from a search entry to a QR code entry), changes in the number of entries (e.g., the difference in entries between adjacent records exceeding a threshold), changes in dwell time (e.g., the dwell time of adjacent records exceeding or falling below the average by a certain percentage), changes in the number of purchases, changes in the number of inquiries, and abnormal increases or decreases in the number of front-end / back-end switching. Adjacent records that meet these conditions are labeled, and the type and magnitude of the change are recorded, forming sequence labeling data that reflects the differences between records.

[0031] After tagging, the sequence is organized, merging consecutive, unchanged records into a single time period. The start and end times are recorded, and the cumulative number of entries and corrective actions for that period are calculated, such as total dwell time, total purchases, and total inquiries. Change nodes are retained separately to show points in time when user behavior changes significantly. After organization, the records along the entire path are rearranged chronologically to form the final path entry representation sequence. Each record or time period clearly displays the entry time, path type, entry method, cumulative number of entries, and corresponding corrective actions, fully presenting the evolution of user behavior along that path. This provides an accurate data foundation for subsequent historical entry sequence analysis and advertising effectiveness prediction.

[0032] Furthermore, a historical entry sequence of the target ad is established based on the path entry representation sequence, and trend analysis is performed on the historical entry sequence to obtain the entry prediction result for the next period. Then, combined with the correction operation data, the initial effect prediction result of the target ad is generated, specifically including: Based on the path entry representation sequence, read the time information and entry count information corresponding to each record in the sequence; Obtain the ad display period, divide the read time information, and obtain the entry count statistics for each display period; Based on the statistics of the number of entries, they are arranged in chronological order according to the display period to form a historical entry sequence; Based on the historical entry sequence, the entry count is obtained in chronological order between every two adjacent display periods; Based on the number of entries obtained, the entry difference between adjacent display periods is calculated to obtain the entry change in each time period; The sign of the change in the number of entries is determined to distinguish between an upward trend and a downward trend in the number of entries, thus obtaining the direction of change; After determining the trend direction, analyze the magnitude of the change in the number of entries based on the value of the change. By combining the direction and magnitude of the change, determine the trend of the number of entries in the next display cycle; Based on the changing trend, the number of entries in the next display cycle is predicted, and the entry number prediction result for the next cycle is obtained. Combined with the path information, the number of entries for each path in the next display cycle is predicted. Specifically, the entry time and corresponding number of entries for each record are extracted from the path entry representation sequence to describe the distribution of user behavior during ad display. Records are divided into various periods based on the ad display cycle, such as by hour, day, or preset ad phase. Each entry record is assigned to its corresponding display cycle, and the total number of entries for each cycle is calculated to form the entry count statistics for each cycle. The statistics are arranged chronologically to obtain the historical entry count sequence, with each record reflecting the cumulative number of entries within a cycle. By comparing the entry counts of two adjacent display cycles, the difference between them is calculated to obtain the change in entry count for each time period. The sign of the change is determined: a positive difference indicates an increase in entry count, and a negative difference indicates a decrease in entry count, thus determining the direction of change.

[0033] After determining the direction of change, the magnitude of the change is analyzed to measure the amplitude of the change in the number of entries. A larger amplitude indicates significant fluctuations in the number of entries, while a smaller amplitude indicates stable entry numbers. After obtaining the direction and amplitude of the change in the number of entries, the number of entries for the next display period can be extrapolated and predicted based on the identified trend. In operation, the number of entries for the current display period is used as the base value, and adjustments are generated for extrapolation based on the direction and amplitude obtained from the previous trend analysis: when the trend is continuously rising, the current number of entries is adjusted positively according to the confirmed upward amplitude; when the trend is declining, it is corrected negatively according to the downward amplitude; when the trend fluctuates or the amplitude is small, a stabilization approach is adopted, performing small-amplitude smoothing adjustments to avoid excessive deviation in the predicted value. To ensure the reliability and engineering feasibility of the adjustment, before generating the adjustment, outlier detection is performed on the change in the number of entries from the previous period. Sudden peaks, jump records, or data that do not conform to the access validity rules are identified and processed, and short-term anomalies are eliminated through methods such as removal and compression to eliminate interference from trend judgment. After obtaining the processed effective change, the adjustment amount used for extrapolation is determined by the average level, typical amplitude, or median change of the change over multiple periods, making the prediction basis more stable and controllable. Subsequently, boundary constraints are introduced before forming the final prediction value. By setting upper and lower limits for change, the prediction value is kept within a reasonable range consistent with historical fluctuation characteristics, avoiding unreasonable sharp increases or decreases. This ensures that the extrapolation prediction process has both trend sensitivity and result stability. Through the above process, the trend direction, change amplitude, and effective change amount can be transformed into a specific prediction value for the next display period. Based on the current number of entries, the determined adjustment amount is superimposed according to the trend direction, choosing to increase, decrease, or smooth the adjustment. This forms a prediction mechanism that is directly deployable, interpretable, and controllable in engineering.

[0034] For example, if the number of entries in the last three display periods increases by 20, 15, and 18 times respectively, without triggering outlier detection, the typical level of these changes can be determined as 17. With the current number of entries at 150, 17 can be added to the current value as an extrapolation increment, resulting in a prediction of approximately 167 entries for the next display period. If a continuous downward trend in the number of entries is identified, such as a decrease of 30 and 25 times in the previous two periods respectively, and these decreases are within the allowable range of boundary constraints, the typical decrease can be determined as 28, and this decrease can be subtracted from the current value to generate a prediction reflecting the downward trend. If the changes in the number of entries in the previous few periods are small, such as an overall fluctuation of less than five times, or if some changes were not adopted due to triggering anomaly detection rules, then three to five small smoothing adjustments can be performed. This ensures that the prediction remains responsive to trends while being unaffected by short-term minor fluctuations or outliers, resulting in more robust overall performance and more interpretable and engineering-applicable predictions.

[0035] Read the correction operation data and extract the corresponding fields of ad dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching according to each delivery path; Based on the entry count statistics for each display period and combined with path information, the historical entry count statistics for each path are obtained. Based on the historical entry count of each path and the corresponding period of the extracted fields of ad dwell time, purchase count, customer service inquiry count and front-end / back-end switching count, a linear regression model is constructed with the historical entry count of each path as the independent variable and the operation behavior field as the dependent variable. The historical entry counts for each path are input into the regression model to obtain the predicted values ​​of the operation data corresponding to each path. The predicted values ​​are then compared with the corrected operation data, and the regression model parameters are adjusted based on the difference to form a corrected regression model. The number of times each path is predicted to enter the next display cycle is input into the calibrated regression model to obtain the prediction operation data. The prediction operation data is then normalized to obtain the comprehensive effect index. The overall performance index is used as the initial performance prediction result for the target advertisement; Understandably, the process begins by acquiring historical display data for the target ad, including the historical entry count for each delivery path, and the ad dwell time, purchase count, customer service inquiry count, and front-end / back-end switching count extracted for each path within the corresponding period. For each path, the historical entry count is used as the input variable, and the various operational behavior data for the corresponding period are used as the output variable. The input and output data for each display period are compiled into a training sample set, and then the relationship between the input and output is fitted. During the fitting process, for the training samples of each path, the difference between the operational behavior values ​​predicted by the regression model and the actual observed corrected operational data is compared to calculate the sum of squared errors for each sample. A set of model parameters is then sought that minimizes the sum of squared errors for all training samples. Through continuous iteration and optimization, a regression model that can accurately reflect the relationship between the input entry count and the output operational behavior is obtained. The output of the regression model includes ad dwell time, purchase count, customer service inquiry count, and front-end / back-end switching count. Data for each path is trained separately to form an initial mapping model for each path, realizing the function of mapping historical entry counts to operational behaviors.

[0036] Subsequently, the historical entry counts for each path are used as input and sequentially fed into the trained regression model to obtain the predicted operational data values ​​for the corresponding path, including ad dwell time, purchase counts, customer service inquiry counts, and front-end / back-end switching counts. The predicted values ​​are compared with the observed corrected operational data to analyze the differences between the predicted and actual values. Based on these differences, the parameters of the regression model are adjusted so that the model can more accurately reflect the actual situation when predicting operational behavior for each path. Through continuous iterative parameter updates, a corrected regression model is finally obtained, which can better map the entry counts for each path to the corresponding operational behavior. Then, the entry counts for each path are predicted based on historical trends and path distribution, and these predicted entry counts are sequentially fed into the corrected regression model to obtain the predicted operational data values ​​for each path. The predicted operational data for each path are normalized to unify the dimensions and numerical ranges of different indicators, allowing dwell time, purchase counts, inquiry counts, and back-end switching counts to be compared and synthesized under the same evaluation standard. After normalization, the operational data for each path are accumulated according to the path, and the back-end switching count is weighted as a negative influence to obtain a comprehensive performance index. This comprehensive performance indicator takes into account multiple factors such as ad dwell time, user purchase behavior, inquiry behavior, and number of times the ad is switched to the background. It can intuitively reflect the overall performance of the target ad in the next display cycle and serve as the initial performance prediction result for the target ad, providing a reference for subsequent ad placement optimization and strategy adjustment.

[0037] The formula for calculating the comprehensive performance index is as follows: ; In the formula, , , and The first The values ​​for ad dwell time, purchase frequency, customer service inquiry frequency, and front-end / back-end switching frequency for each path are normalized. This represents the total number of advertising delivery paths.

[0038] Understandably, the comprehensive performance index calculated using the formula can be used to measure the overall interaction of user groups across different ad delivery paths. A higher index indicates longer user dwell time, more purchases, and more inquiries during ad display, coupled with fewer front-end / back-end switching, suggesting strong ad appeal and high conversion potential. Conversely, a lower index indicates insufficient ad appeal or low user engagement. Specifically, dwell time reflects user attention to the ad, purchase frequency directly reflects conversion ability, customer service inquiries reflect user interest or questions, and front-end / back-end switching is a negative factor indicating distracted user attention. This index is normalized, allowing for comparison and synthesis of different types of behavior on the same scale, resulting in a clear overall performance evaluation. In practical applications, the index for the current period can be compared with historical data or similar ad metrics to determine whether ad performance is high, medium, or low. Trends can also be observed: an upward trend indicates improved ad performance, a downward trend suggests potential impact and the need for strategy adjustments, while stability indicates stable ad performance. In this way, the comprehensive performance index not only quantifies ad engagement and conversion potential but also provides a clear reference for ad optimization.

[0039] Furthermore, based on pre-defined monitoring channels, data on changes in the exposure of competing advertisements are obtained, and the initial performance prediction results are corrected for competitiveness to obtain the final performance prediction results for the target advertisement, specifically including: Based on preset monitoring channels, obtain data on the changes in exposure of competing ads that are in the same advertising environment as the target ad in different time periods; The data on changes in competitive ad impressions were organized in chronological order to obtain a sequence of impression changes for each time period. Based on the sequence of competitive ad exposure changes, we identify the rising, falling and stable intervals of exposure and assign corresponding competitive pressure level indicators to each interval. Based on the competitive pressure level index, the initial effect prediction results of the target advertisement are corrected for the degree of competition to obtain the final advertising effect prediction results of the target advertisement. Specifically, through pre-defined monitoring channels, data on the exposure changes of competing ads placed in the same environment as the target ad across different time periods are acquired. This data reflects the display status and user contact frequency of competing ads at different times. These exposure data are organized chronologically to form a continuous exposure change sequence, with each time period corresponding to a specific exposure value, facilitating the analysis of dynamic trends in competing ads. When analyzing the exposure change sequence of competing ads, the exposure of each time period is first normalized, mapping the exposure of each time period to a fixed range (e.g., 0 to 100) to compare the relative intensity of different time periods. Based on the normalized values, the deviation of the exposure of each time period from the sequence mean is first calculated. The deviation is obtained by calculating the absolute value of the difference between the exposure of that time period and the exposure mean of the sequence. The deviation of all time periods is then normalized according to the largest deviation, resulting in a base pressure value ranging from 0 to 1, which reflects the competitive strength of the exposure of that time period relative to the overall level. Subsequently, a trend factor is determined by combining the change in exposure of that time period with the previous time period. When the exposure volume in a given time period is greater than that in the previous time period, it indicates an upward trend in exposure, and the trend factor is set to a preset positive value. When the exposure volume in a given time period is less than that in the previous time period, it indicates a downward trend in exposure, and the trend factor is set to a preset negative value. When the difference between the exposure volume in a given time period and that in the previous time period is within a preset threshold range, it indicates that the exposure volume remains stable, and the trend factor is set to zero. By adding the trend factor to the baseline pressure value, the competitive pressure level index for that time period is obtained, allowing the index to simultaneously reflect both the absolute level of exposure and its changing trend.

[0040] For example, periods with exposure significantly higher than the sequence mean are assigned higher competitive pressure values ​​(e.g., 80-100), periods with exposure close to the mean are assigned medium competitive pressure values ​​(e.g., 40-60), and periods with exposure significantly lower than the mean are assigned low competitive pressure values ​​(e.g., 0-20). To more precisely reflect changing trends, the index is adjusted based on the rising or falling ranges of the sequence. When exposure shows an upward trend over a continuous period, the pressure value is allocated according to the magnitude of the increase; when it shows a continuous decline, the pressure value is allocated according to the magnitude of the increase; and a constant pressure value is maintained in the stable range. For ease of engineering implementation, the competitive pressure value can be mapped to a numerical range, for example, high pressure is 80-100, medium pressure is 40-60, and low pressure is 0-20. The specific values ​​obtained for each time period can be directly used for subsequent correction operations. In this way, a specific competitive pressure index is obtained for each time period, which can quantify the impact of competing advertisements on the potential diversion of traffic to the target advertisement during that time period, providing a basis for adjusting the initial effect prediction of the target advertisement.

[0041] After obtaining the competitive pressure indicators for each time period, these are combined with the initial performance predictions of the target advertisement to execute specific correction operations. For example, for time periods with high competitive pressure, the initial performance prediction value can be reduced proportionally to mitigate the potential impact of traffic diversion; for time periods with medium pressure, the prediction value can be slightly adjusted to reflect some traffic diversion; for time periods with low pressure or stability, the prediction value is maintained or fine-tuned to ensure normal advertising performance. In practice, the competitive pressure indicators can be directly mapped to adjustment ratios or adjustment magnitudes. For example, when the pressure value is 80-100, the initial performance is reduced by 10%-15%; when the pressure value is 40-60, it is reduced by 3%-7%; and when the pressure value is 0-20, the original value is maintained or fine-tuned within 1%. Through this mechanism, the final advertising performance prediction result not only integrates the target advertisement's own historical entry and correction operation data but also fully considers the dynamic changes in the external competitive environment, making the prediction results more accurate, actionable, and engineerable.

[0042] The formula for calculating the final advertising effectiveness prediction result is as follows: ; In the formula, The final advertising effect prediction result for the target advertisement. The initial performance prediction results for the target advertisement. For the first The level of competitive pressure over a given period of time. This represents the total number of time periods.

[0043] Furthermore, an advertising effectiveness prediction system based on the fusion of delivery log data is proposed to implement the advertising effectiveness prediction method described above, characterized by comprising: The data acquisition module is used to obtain the target advertisement's delivery path information and corresponding delivery log data, and to collect user's display period operation data and advertisement display quality information in real time. The path identification and integration module is used to identify the actual path and method by which users enter the advertisement based on the delivery log data, calculate the number of valid path entries, and integrate them with the display period operation data to form path entry records and display period operation data. The operation correction module is used to correct the operation data during the display period based on the ad loading time information, exposure area ratio information and playback stuttering information, and generate corrected operation data. The path representation building module combines the correction operation data with the path entry record to construct the path entry structure data and generate the path entry representation sequence of the target advertisement. The effect prediction module is used to generate a historical entry frequency sequence based on the path entry representation sequence, perform trend analysis to obtain the entry frequency prediction result for the next period, and combine the correction operation data to generate the initial effect prediction result for the target advertisement. The competition correction module is used to obtain the exposure change data of competing advertisements and perform competition correction on the initial effect prediction results to obtain the final advertising effect prediction results of the target advertisement.

[0044] Furthermore, the data acquisition module includes: The delivery log collection unit is used to obtain information on multiple delivery paths of the target advertisement and the corresponding delivery log data. The user operation data collection unit is used to collect data in real time on the user's dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching during the ad display period. The display quality acquisition unit is used to simultaneously collect information such as ad loading time, exposure area ratio, and playback stuttering.

[0045] Furthermore, the path representation building modules include: The path entry data combination unit is used to form a path entry data set based on the path entry information of each delivery path and the corresponding correction operation data. The path sub-information generation unit is used to divide the path entry data set into sub-information according to different entry paths, and includes the number of entry times, entry method and correction operation data during the advertising display period for each path. The behavior mapping table generation unit is used to expand the number of entries and the entry method in the path sub-information into individual entry records, and associate them with the corresponding correction operation data to generate an indexable behavior mapping table. The path entry sequence generation unit is used to generate the final path entry representation sequence by sorting and structuring according to the path dimension based on the behavior mapping table.

[0046] Furthermore, the effect prediction module includes: The historical entry count generation unit is used to read the time information and entry count information corresponding to each record based on the path entry representation sequence, and generate a historical entry count sequence according to the advertising display cycle. The trend analysis unit is used to calculate the entry difference between adjacent display periods based on the historical entry frequency sequence, determine the direction and magnitude of the change in entry frequency, and determine the trend of entry frequency change in the next display period. The initial effect calculation unit is used to input the number of times each path is predicted for the next cycle into the calibrated linear regression model to obtain the prediction operation data corresponding to each path. The prediction operation data is then normalized, and a comprehensive effect index is calculated, including ad dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching times, which serves as the initial effect prediction result for the target ad.

[0047] In summary, the advantages of this invention are as follows: by integrating log data from multiple target ad delivery paths, user operation behavior data, and ad display quality information, it achieves accurate identification of the user's actual entry path and method, corrects the operation data during the display period, constructs a path entry representation sequence, performs trend analysis based on the historical entry frequency sequence to predict the entry frequency in the next cycle, and combines competitive ad exposure change data for competition degree correction, thereby obtaining the final effect prediction result of the target ad. This invention can improve the accuracy and reliability of ad effect prediction, reduce prediction deviations caused by differences in user paths, uneven display quality, or changes in the competitive environment, and provide a scientific basis for ad placement optimization and decision-making.

[0048] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for predicting advertising effectiveness based on the fusion of delivery log data, characterized in that, include: Determine the multiple ad delivery paths corresponding to the target ad and obtain the delivery log data for each path to form a delivery log dataset; When an ad is loaded and displayed, the user's entry path and method into the ad are identified based on the ad delivery log dataset, and the number of times the corresponding path is entered and the user operation data generated during the ad display period are recorded to obtain the path entry record and display period operation data. During the ad display period, the ad display quality information is collected synchronously to form display quality data. Then, the operation data during the display period is corrected to obtain corrected operation data. The correction operation data is combined with the path entry record to form the path entry structure data, and the path entry representation sequence of the target advertisement is generated. Based on the path entry representation sequence, a historical entry frequency sequence of the target advertisement is established, and trend analysis is performed on the historical entry frequency sequence to obtain the entry frequency prediction result for the next period. Then, combined with the correction operation data, the initial effect prediction result of the target advertisement is generated. Based on preset monitoring channels, data on changes in the exposure of competing advertisements are obtained, and the initial effect prediction results are corrected for competitiveness to obtain the final advertising effect prediction results of the target advertisement.

2. The advertising effect prediction method based on the fusion of delivery log data according to claim 1, characterized in that, When an ad is loaded and displayed, the system identifies the user's entry path and method into the ad based on the ad delivery log dataset, and records the corresponding number of entry attempts and user operation data generated during ad display, obtaining path entry records and display period operation data, specifically including: Obtain the user's access identification information, which includes terminal identifier, network address, access timestamp, and browser type; Obtain the delivery path information corresponding to the target advertisement, and extract the path fields related to the delivery path from the delivery log dataset. The delivery path information includes the search entry, QR code entry, and recommendation click entry. Based on the ad delivery log dataset, multi-dimensional features of access behavior are constructed using access identification information. The ad delivery path field is cross-validated to determine the actual ad delivery path corresponding to the user and to record the specific path and method of entry for the user. For search entry points, identify user behavior by entering keywords or clicking on search results to access target ads; For QR code entry points, identify the behavior of users entering the target advertisement by scanning the advertisement's QR code; For recommended ad entry points, identify user behavior by clicking on a recommended ad on the page to access the target ad. Based on the deployment log dataset, access interval information, loading status information, and network status information are extracted for access validity determination, and access validity determination rules are constructed. Multiple access records under the same delivery path are accumulated, and invalid and abnormal accesses in the access logs are filtered according to the access validity judgment rules to generate the valid path entry count for that delivery path. During the ad display period, user operation data is collected in real time, including the time users stay in the ad area, the number of purchases made by users, the time users initiate customer service inquiries, and the number of times users switch between the front and back ends. The number of valid path entries is matched with the entry method corresponding to each path, and then linked and integrated with the display period operation data collected during the ad display period according to the entry event to form path entry records and display period operation data.

3. The advertising effect prediction method based on the fusion of delivery log data according to claim 1, characterized in that, The process of synchronously collecting ad display quality information during ad display to form display quality data, followed by correcting the display period operation data to obtain corrected operation data, specifically includes: Obtain the start time and completion time of ad loading to get the ad loading time information; Obtain data on the actual display area of ​​the advertisement on the terminal screen and data on the preset display area of ​​the advertisement, and obtain the exposure area ratio information based on the ratio between the display area data and the preset area data; Obtain status data during ad playback, record the number of playback interruptions and their duration to obtain playback stuttering information; Based on loading time information, exposure area occupancy information, and playback stuttering information, display quality information is generated for each ad display. Based on the loading time information in the display quality information, the dwell time and occurrence time of various interactive behaviors in the display period operation data are corrected according to the loading delay, and the invalid segments during the loading period are deducted to obtain the preliminary correction data. Based on the exposure area ratio information in the display quality information, the dwell time, purchase frequency and consultation behavior in the preliminary correction data are corrected according to the corresponding area ratio to obtain the exposure correction data; Based on the playback stuttering information in the display quality information, the dwell time, number of purchases, consultation time, and number of front-end / back-end switching in the exposure correction data are filtered for effectiveness and corrected for time according to the stuttering segment to form the final correction data, which serves as the correction operation data.

4. The advertising effect prediction method based on the fusion of delivery log data according to claim 1, characterized in that, The step of combining the correction operation data with the path entry record to form path entry structure data and generating the path entry representation sequence of the target advertisement specifically includes: Based on the path entry information of each delivery path and the corresponding correction operation data, a path entry data set is formed. The path entry data set is divided according to different entry paths to generate first path sub-information, second path sub-information and third path sub-information. The first path sub-information, second path sub-information and third path sub-information all include the corresponding path entry times, entry methods and correction operation data during the ad display period. Based on the first path sub-information, the second path sub-information, and the third path sub-information, the number of entries and the entry method in each path sub-information are expanded into individual entry records, and then associated with the corresponding correction operation data item by item to obtain a list of entry records organized by path. Based on the list of entry records, each entry record is mapped to its corresponding correction operation data to form a behavior mapping table that can be looked up by record index; Using the behavior mapping table as input, the mapping records under the same path are arranged sequentially according to the entry time to obtain an ordered behavior flow that reflects the entry change process in the same path; Based on ordered behavior flow, multiple ordered behavior flows along the same path are merged according to the path dimension to generate a primary path entry sequence. Based on the primary path entry sequence, the number of entries, entry methods and correction operation data in the sequence are scanned item by item, and the differences between adjacent records are marked according to the preset change conditions to form sequence annotation data that can reflect the change relationship between records. Based on sequence labeling data, the primary path entry sequence is structured and organized, and a path entry representation sequence is generated according to the record order.

5. The advertising effect prediction method based on the fusion of delivery log data according to claim 1, characterized in that, The process involves establishing a historical entry sequence for the target ad based on the path entry representation sequence, performing trend analysis on the historical entry sequence to obtain the entry prediction result for the next period, and then combining it with correction operation data to generate the initial effect prediction result for the target ad. Specifically, this includes: Based on the path entry representation sequence, read the time information and entry count information corresponding to each record in the sequence; Obtain the ad display period, divide the read time information, and obtain the entry count statistics for each display period; Based on the statistics of the number of entries, they are arranged in chronological order according to the display period to form a historical entry sequence; Based on the historical entry sequence, the entry count is obtained in chronological order between every two adjacent display periods; Based on the number of entries obtained, the entry difference between adjacent display periods is calculated to obtain the entry change in each time period; The sign of the change in the number of entries is determined to distinguish between an upward trend and a downward trend in the number of entries, thus obtaining the direction of change; After determining the trend direction, analyze the magnitude of the change in the number of entries based on the value of the change. By combining the direction and magnitude of the change, determine the trend of the number of entries in the next display cycle; Based on the changing trend, the number of entries in the next display cycle is predicted, and the entry number prediction result for the next cycle is obtained. Combined with the path information, the number of entries for each path in the next display cycle is predicted. Read the correction operation data and extract the corresponding fields of ad dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching according to each delivery path; Based on the entry count statistics for each display period and combined with path information, the historical entry count statistics for each path are obtained. Based on the historical entry count of each path and the corresponding period of the extracted fields of ad dwell time, purchase count, customer service inquiry count and front-end / back-end switching count, a linear regression model is constructed with the historical entry count of each path as the independent variable and the operation behavior field as the dependent variable. The historical entry counts for each path are input into the regression model to obtain the predicted values ​​of the operation data corresponding to each path. The predicted values ​​are then compared with the corrected operation data, and the regression model parameters are adjusted based on the difference to form a corrected regression model. The number of times each path is predicted to enter the next display cycle is input into the calibrated regression model to obtain the prediction operation data. The prediction operation data is then normalized to obtain the comprehensive effect index. The overall performance index is used as the initial performance prediction result for the target advertisement; The formula for calculating the comprehensive effect index is as follows: ； In the formula, , , and The first The values ​​for ad dwell time, purchase frequency, customer service inquiry frequency, and front-end / back-end switching frequency for each path are normalized. This represents the total number of advertising delivery paths.

6. The advertising effect prediction method based on the fusion of delivery log data according to claim 1, characterized in that, The process involves acquiring data on changes in the exposure of competing advertisements based on preset monitoring channels, and then correcting the initial performance prediction results for competition to obtain the final performance prediction result for the target advertisement. Specifically, this includes: Based on preset monitoring channels, obtain data on the changes in exposure of competing ads that are in the same advertising environment as the target ad in different time periods; The data on changes in competitive ad impressions were organized in chronological order to obtain a sequence of impression changes for each time period. Based on the sequence of competitive ad exposure changes, we identify the rising, falling and stable intervals of exposure and assign corresponding competitive pressure level indicators to each interval. Based on the competitive pressure level index, the initial effect prediction results of the target advertisement are corrected for the degree of competition to obtain the final advertising effect prediction results of the target advertisement. The formula for calculating the final advertising effect prediction result is as follows: ； In the formula, The final advertising effect prediction result for the target advertisement. The initial performance prediction results for the target advertisement. For the first The level of competitive pressure over a given period of time. This represents the total number of time periods.

7. An advertising effectiveness prediction system based on the fusion of delivery log data, used to implement the advertising effectiveness prediction method as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to obtain the target advertisement's delivery path information and corresponding delivery log data, and to collect the user's display period operation data and advertisement display quality information in real time. The path identification and integration module is used to identify the actual path and method of users entering the advertisement based on the delivery log data, calculate the number of valid path entries, and integrate them with the display period operation data to form path entry records and display period operation data. An operation correction module is used to correct the operation data during the display period based on information such as ad loading time, exposure area ratio, and playback stuttering, and to generate corrected operation data. A path representation construction module is used to combine correction operation data with path entry records to construct path entry structure data and generate a path entry representation sequence for the target advertisement. The effect prediction module is used to generate a historical entry frequency sequence based on the path entry representation sequence, perform trend analysis to obtain the entry frequency prediction result for the next period, and combine the correction operation data to generate the initial effect prediction result of the target advertisement. The competition correction module is used to acquire the exposure change data of competing advertisements and perform competition correction on the initial effect prediction results to obtain the final advertising effect prediction results of the target advertisement.

8. The advertising effect prediction system based on delivery log data fusion according to claim 7, characterized in that, The data acquisition module includes: The delivery log collection unit is used to acquire multiple delivery path information of the target advertisement and the corresponding delivery log data. The user operation collection unit is used to collect in real time the user's dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching during the advertisement display period; The display quality acquisition unit is used to simultaneously collect information on ad loading time, exposure area ratio, and playback stuttering.

9. The advertising effect prediction system based on the fusion of delivery log data according to claim 7, characterized in that, The path representation construction module includes: The path entry data combination unit is used to form a path entry data set based on the path entry information of each delivery path and the corresponding correction operation data. The path sub-information generation unit is used to divide the path entry data set into sub-information according to different entry paths, and includes the number of entry times, entry method and correction operation data during the advertising display period for each path. A behavior mapping table generation unit is used to expand the number of entries and the entry method in the path sub-information into entry records one by one, and associate them with the corresponding correction operation data to generate an indexable behavior mapping table. The path entry sequence generation unit is used to generate the final path entry representation sequence by sorting and structuring according to the path dimension based on the behavior mapping table.

10. The advertising effect prediction system based on delivery log data fusion according to claim 7, characterized in that, The effect prediction module includes: The historical entry count generation unit is used to read the time information and entry count information corresponding to each record based on the path entry representation sequence, and generate a historical entry count sequence according to the advertising display cycle. The trend analysis unit is used to calculate the entry difference between adjacent display periods based on the historical entry count sequence, determine the direction and magnitude of the change in the number of entries, and determine the trend of the number of entries in the next display period. The initial effect calculation unit is used to input the number of times each path is predicted for the next cycle into the calibrated linear regression model to obtain the prediction operation data corresponding to each path. The prediction operation data is then normalized, and a comprehensive effect index is calculated, including the ad dwell time, number of purchases, number of customer service inquiries, and number of front-end and back-end switching times, as the initial effect prediction result of the target ad.

Images