A Real-Time Optimization Method for Ad Creatives Based on Dynamic Scene Awareness

Abstract

This invention relates to the field of internet advertising optimization technology and discloses a real-time optimization method for advertising creatives based on dynamic scene perception. This method addresses the problem in traditional methods of achieving stable, timely, and verifiable real-time optimization control of target creative units. The method first constructs an advertising optimization session in the advertising display task, continuously acquires scene signals, creative configuration status, and feedback events, performs scene acceptance determination, scene segmentation, and creative unit group division, determines whether the error is caused by scene changes or creative adjustments, then decides the timing and path for adjusting the target creative unit group, and performs verification, rollback, and cross-session acceptance to achieve accuracy, timeliness, stability, and verifiability in real-time optimization of advertising creatives.

Description

Technical Field

[0001] This invention relates to the field of internet advertising optimization technology, specifically a method for real-time optimization of advertising creatives based on dynamic scene perception. Background Technology

[0002] In the field of internet advertising technology, ad creative generation, creative analysis, and placement optimization are important technical means to improve advertising effectiveness. For example, the published invention patent application CN105956874B discloses an ad creative processing method and device, which obtains ad creative specifications and triggers multiple creative supply systems to generate ad creatives in parallel to improve creative processing efficiency and ad space utilization. However, it mainly addresses the issue of ad creative production efficiency and does not involve real-time optimization control of creatives under dynamic scene continuous changes during ad display. Another example is the published invention patent application CN112990967B, which discloses an ad creative analysis method and system. It obtains the effect data of multiple ad creatives and analyzes the contribution of ad element categories to determine the impact of different ad elements on ad effectiveness. However, this solution mainly focuses on the post-event analysis of ad creative effects and lacks online optimization control capabilities for the real-time placement process.

[0003] However, in actual advertising campaigns, the advertising display scenario is usually not static but constantly changing with media content, user environment, terminal status, time and location, and interactive behavior. Simultaneously, creative units such as ad titles, images, copy, and buttons may also undergo parallel adjustments during campaign execution, and advertising performance feedback often exhibits a certain lag. Existing technologies lack an online decoupling attribution and real-time linkage optimization mechanism for continuously changing scene signals and advertising creative units, making it difficult to accurately distinguish whether changes in advertising performance are caused by scene changes or creative adjustments. This insufficient differentiation capability can easily lead to problems such as inaccurate target creative unit selection, delayed adjustment timing, unreasonable adjustment magnitude, and difficulty in verifying optimization results, thus affecting the stability, timeliness, and controllability of real-time advertising creative optimization. Therefore, there is an urgent need to propose a real-time advertising creative optimization method based on dynamic scene perception to achieve online differentiation between the impact of scene changes and creative adjustments, and to perform stable and timely real-time optimization control of target creative units, thereby improving the accuracy, timeliness, and verifiability of advertising creative optimization results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a real-time optimization method for advertising creatives based on dynamic scene perception, which solves the problem in traditional methods of making it difficult to achieve stable, timely, and verifiable real-time optimization control of target creative units.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for real-time optimization of ad creatives based on dynamic scene awareness, comprising:

[0007] S1. Establish an ad optimization session corresponding to the target ad display task, and continuously acquire scene signals, creative configuration status and feedback events in the ad display environment during the ad optimization session;

[0008] S2. Based on the continuously acquired scene signals, determine the scene continuity and segment the current advertising display process, and divide the creative content into independently adjustable creative unit groups.

[0009] S3. Within the same scene segment, associate the change records of the creative unit group with the feedback event to determine whether the change in advertising effect is caused by scene change or creative adjustment.

[0010] S4. Based on the judgment results, combined with the stable state of the current scene segment, the adjustment state of the creative unit group, and the confirmation state of the feedback event, determine the adjustment timing and adjustment path of the target creative unit group.

[0011] S5. After completing the adjustment of the target creative unit group, perform feedback correlation, result verification and anomaly rollback for similar scenarios before and after the adjustment, and use the verified adjustment records for subsequent ad optimization sessions.

[0012] Preferably, an ad optimization session corresponding to the target ad display task is established, and during the ad optimization session, scene signals, creative configuration status, and feedback events in the ad display environment are continuously acquired, including:

[0013] Segment ad optimization sessions by target audience, ad placement, media page type, and session duration.

[0014] Write scene signals, creative configuration status, and feedback events under a unified time base;

[0015] Perform initial validation and supplementary data collection on key fields;

[0016] Perform a reliability determination on the return sequence of the initial creative configuration;

[0017] Receive short-delay feedback before the session ends, and perform cache limit control and session state transition management.

[0018] Preferably, based on continuously acquired scene signals, the current advertising display process is judged for scene continuity and segmented, including:

[0019] A scene signature is generated based on continuous scene signal records in the session buffer. The continuity determination is made between the end scene of the previous session and the beginning scene of the current session, and a short-term observation is performed when the continuity is pending confirmation.

[0020] The scenarios are segmented in a scrolling manner based on page type, ad placement level, main content theme tag, terminal type, interaction status, network status, and page browsing depth, forming segment identifiers, main scenario signatures, stability levels, fluctuation markers, and comparability markers.

[0021] Preferably, the creative content is divided into independently adjustable creative unit groups, including:

[0022] Based on the creative configuration status at the start of the session, the ad creative resource table, the brand restriction table, and the ad review cache table;

[0023] Creative unit groups are divided according to the functional boundaries of information expression, visual presentation, rights and interests explanation, action guidance, and jump configuration;

[0024] Configure member version sets, adjustable tags, dependency tags, conflict tags, and priorities;

[0025] Eligibility for adjustment is determined based on candidate version availability, external business status, target resource availability, and scenario stability level.

[0026] Preferably, within the same scene segment, the change records of creative unit groups are associated with feedback events, including:

[0027] Based on scene segmentation identifiers, creative unit group status sequences, feedback event sequences, comparison statuses of similar scenes, and adjustment record tables;

[0028] Within the same scene segment, feedback events before and after the creative change are hierarchically associated according to the instant feedback window, short-delay feedback window, and delayed feedback window. The attribution of delayed feedback is completed based on the display tracking code, session identifier, scene segment identifier, and adjustment effective time window.

[0029] Preferably, determining whether changes in advertising effectiveness are caused by changes in the scenario or by adjustments to the creative content includes:

[0030] First, determine whether any changes have occurred to the creative unit groups within the current scene segment;

[0031] Key feedback is selected based on the type of advertising objective, and confidence level is determined by combining the feedback benchmark, change direction record and number of valid feedback samples from the previous segment of the same scenario.

[0032] Execute conflict observation when creative changes and scene transitions occur within the same time window;

[0033] When the interval between multiple creative changes within the same scene segment is less than a preset threshold, a mixed influence flag is executed.

[0034] When the conditions of the same type of scenario are met, generate scenario change candidates or creative adjustment candidates.

[0035] Preferably, based on the judgment result, combined with the stable state of the current scene segment, the adjustment state of the creative unit group, and the confirmation state of the feedback event, including:

[0036] Adjustment control decisions are generated based on the results of the change source determination, the stability level of the scene segment, the determination confidence level, the feedback confirmation level, the adjustable mark of the creative unit group, the conflict mark of the creative unit group, the cumulative number of adjustments in the current session, the last adjustment time, the media dynamic replacement permission mark, and the advertiser restriction rules.

[0037] Based on the scene change judgment result or the creative change judgment result, the target unit group is controlled to switch between the observation state, the preparation state, the execution state, the delayed state, the frozen state and the recovery state.

[0038] Preferably, the timing and path for adjusting the target creative unit group are determined, including:

[0039] The target creative unit groups are determined based on the principles of low coupling priority, low risk priority, rollback priority, and minimum single change priority.

[0040] The timing of adjustments is determined based on the cooling window, the time of the last adjustment, the stability level of the scene segment, and the feedback confirmation level.

[0041] Configure candidate version switching paths, switching observation durations at each level, and conditions for skipping levels according to the principle of increasing adjustment magnitude;

[0042] When the cumulative number of adjustments exceeds the limit, there are unit group conflicts, negative path accumulation, media is disabled, review fails, or feedback continues to worsen, restrictions, freezes, rollbacks, or restorations will be implemented.

[0043] Preferably, after adjusting the target creative unit group, feedback correlation, result verification, and anomaly rollback are performed on similar scene segments before and after the adjustment, including:

[0044] Based on the action adjustment identifier, creative versions before and after the adjustment, adjustment effective time, scene segment where the adjustment is located, comparison records of similar scene segments, and subsequent feedback event sequence, the attribution binding is performed according to the display tracking code, session identifier, scene segment identifier, and adjustment effective time window.

[0045] When the conditions of similar scenarios are met, the key feedback improvement threshold and side effect judgment rules are used to verify and classify the data, and the verification conclusions are recorded in the stable acceptance record, the observation acceptance record, the frozen record or the abnormal record.

[0046] When the verification result is negative, the intervention level of the adjusted path will be adjusted to implement immediate rollback, short observation recovery, or freeze.

[0047] Preferably, the verified adjustment records will be used for subsequent ad optimization sessions, including:

[0048] Read the official stable records, observe the acceptance records and freeze records according to the scenario signature matching relationship, and determine the acceptance path of the subsequent advertising optimization session based on the number of confirmations, the most recent confirmation time and the risk of freezing.

[0049] For delayed feedback within the supplementary window that can be clearly attributed, the verification conclusion will be adjusted according to the adjacent conclusion levels.

[0050] Feedback exceeding the recovery window is written into the long-term performance archive;

[0051] Perform exclusion processing or empty path processing for abnormal and uncleared paths, frozen paths, and paths without successors.

[0052] Compared with existing technologies, this invention provides a real-time optimization method for advertising creatives based on dynamic scene perception, which has the following beneficial effects:

[0053] 1. This invention establishes an advertising optimization session to achieve unified time alignment, continuous reception, and structured management of scene signals, creative configuration status, and feedback events. It also distinguishes the correspondence between advertising effect changes and scene changes and creative adjustments by combining scene reception judgment, scene segmentation, creative unit grouping, and comparison with similar scenes. Furthermore, it hierarchically associates creative change records and feedback events within the same scene segment, and combines judgment confidence grading, conflict observation, mixed influence marking, adjustment timing control, path progressive switching, result verification, anomaly rollback, and cross-session reception record updates. This solves the problems in existing technologies where it is difficult to accurately distinguish the source of effect changes under conditions of continuous scene changes, parallel creative adjustments, and asynchronous feedback arrival. This leads to inaccurate target creative unit selection, delayed adjustment timing, unreasonable adjustment magnitude, and difficulty in verifying optimization results. The invention improves the accuracy, timeliness, stability, and verifiability of the real-time advertising creative optimization process.

[0054] 2. This invention, by hierarchically verifying the adjusted feedback results and forming stable acceptance records, observation acceptance records, and frozen records according to different verification conclusions, and combining delayed feedback replenishment, abnormal record updates, and acceptance path reading in subsequent advertising optimization sessions, continuously accumulates and orderly inherits historical adjustment information. This solves the problems in the prior art of difficulty in accumulating single optimization results, easy interruption of cross-session adjustment basis, repeated use of abnormal paths, and lack of stable reference for subsequent optimization, and achieves improved continuity, historical reusability, and abnormal avoidance capabilities in the advertising creative optimization process. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the process of a real-time optimization method for advertising creatives based on dynamic scene perception according to the present invention.

[0056] Figure 2 This is a schematic diagram of the scene acceptance determination and rolling scene segmentation of the present invention;

[0057] Figure 3 This is a diagram showing the division of creative unit groups and the attribution of feedback events in this invention.

[0058] Figure 4 The target creative unit group of this invention is adjusted to control state flow diagram;

[0059] Figure 5 This is the adjusted verification, abnormal rollback, and cross-session acceptance closed-loop diagram for this invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1: Figures 1-5 A method for real-time optimization of ad creatives based on dynamic scene awareness is presented, including:

[0062] S1. Establish an ad optimization session corresponding to the target ad display task, and continuously acquire scene signals, creative configuration status and feedback events in the ad display environment during the ad optimization session;

[0063] S2. Based on the continuously acquired scene signals, determine the scene continuity and segment the current advertising display process, and divide the creative content into independently adjustable creative unit groups.

[0064] S3. Within the same scene segment, associate the change records of the creative unit group with the feedback event to determine whether the change in advertising effect is caused by scene change or creative adjustment.

[0065] S4. Based on the judgment results, combined with the stable state of the current scene segment, the adjustment state of the creative unit group, and the confirmation state of the feedback event, determine the adjustment timing and adjustment path of the target creative unit group.

[0066] S5. After completing the adjustment of the target creative unit group, perform feedback correlation, result verification and anomaly rollback for the same scene segments before and after the adjustment, and use the verified adjustment records for the subsequent ad optimization session handling.

[0067] Ad optimization sessions are established separately according to ad target, ad placement, media page type, and session time range. During the ad optimization session, scene signals, creative configuration status, and feedback events in the ad display environment are continuously acquired as a unified input basis for subsequent processing. The reason for adopting the above establishment method is that there are differences in the display rhythm, feedback latency, and attribution criteria corresponding to different traffic sources, different ad placement levels, and different display formats. If the relevant data is directly mixed, it is easy to cause inconsistencies in the data sources within the same ad optimization session, thereby affecting subsequent time alignment, feedback attribution, and result comparison.

[0068] The input structure of an ad optimization session includes at least the following fields: ad task identifier, ad object identifier, ad plan identifier, ad placement identifier, media page type, terminal type, display start time, current creative version, strategy version, parameter version, and session validity period. The ad task identifier identifies the current execution process, the ad object identifier identifies the displayed ad content, the ad plan identifier identifies the ad object's execution plan, the ad placement identifier limits the feedback scope, the current creative version identifies the creative configuration that actually took effect at the start of the session, and the strategy and parameter versions ensure traceability of subsequent processing. The session identifier for the ad optimization session can be a unique code ranging from 24 to 64 bits, preferably 32 bits. This setting satisfies the unique identification requirements across ad objects, ad placements, page types, and time windows, while also controlling the index length for easier caching and log retrieval. A length below 24 bits increases the risk of duplication in high-concurrency environments, while a length above 64 bits increases transmission burden and indexing overhead, hindering real-time processing.

[0069] The session validity period is used to limit the continuous observation range of the current ad optimization session, and can be set from 10 seconds to 300 seconds. This parameter is determined in conjunction with the ad refresh rate, page dwell characteristics, and the frequency of scene changes. When the ad refresh cycle is short, the content switching frequency is high, or the user dwell time is short, the session validity period is set from 10 seconds to 60 seconds. When the ad dwell time is at a medium level and the page browsing behavior is relatively continuous, the session validity period is set from 30 seconds to 120 seconds. When the ad display is accompanied by a long interaction process, the page dwell time is long, or there are multiple consecutive operations, the session validity period is set from 60 seconds to 300 seconds. For in-feed ads, the session validity period is, for example, 30 seconds to 120 seconds. For short videos... For feed ads, the session validity period is, for example, 10 to 60 seconds; for interactive landing page ads, the session validity period is, for example, 60 to 300 seconds. This is because continuous viewing of feed ads typically lasts for tens of seconds, a single effective view of short video ads typically lasts for several seconds to tens of seconds, and effective dwell time on interactive landing pages typically lasts for tens of seconds to several minutes. Therefore, the session validity period needs to cover a continuous display observation interval while avoiding crossing too many scene change nodes. When the session validity period is less than 10 seconds, it is difficult to completely inherit the continuous state in the same display process; when the session validity period is more than 300 seconds, too many scene drifts are easily superimposed within the same session, resulting in unstable subsequent comparison criteria.

[0070] After the session is established, scene signal data streams, creative configuration status data streams, and feedback event data streams are continuously received and written to the session buffer according to a unified time base. The unified time base is aligned using standard timestamps within the same session. Scene signals, creative configuration status, and feedback events are all converted to a unified time scale based on the collection time or reporting time, so that subsequent scene segmentation, creative status comparison, and feedback attribution binding can be performed on the same time axis. The scene signal data stream includes at least page type, content theme tag, content theme main tag, ad placement level, ad placement preceding content tag, terminal type, screen orientation, network level, and time period. The data stream includes fields such as tag, geographic region label, user real-time interaction status, and current page browsing depth; the creative configuration status data stream includes at least fields such as creative version identifier, title version, main image version, cover version, rights text version, button copy version, landing page anchor point version, layout configuration version, and effective time of each version; the feedback event data stream includes at least fields such as event identifier, event type, event occurrence time, event reporting time, associated ad object identifier, associated display identifier, and feedback delay level; all the above data are stored in a structured record format to facilitate subsequent completion of scene transition, creative status comparison, and feedback attribution binding in chronological order;

[0071] The scene signal data stream is continuously updated with a sampling step size of 1 to 5 seconds. The sampling step size is determined based on the scene change speed and session cache pressure corresponding to the ad type. When the content switching frequency is high and the viewing pace is fast, the sampling step size is 1 to 2 seconds; when the browsing pace is medium and the page changes are relatively stable, the sampling step size is 2 to 3 seconds; when the page dwell time is long and the interface changes are relatively gentle, the sampling step size is 3 to 5 seconds. For short video ads, the sampling step size can be 1 second; for feed ads, the sampling step size can be 2 seconds; for interactive page ads, the sampling step size can be 3 seconds. The reason for this setting is that a sampling step size of less than 1 second will generate a large number of duplicate records, significantly increasing the risk of data loss. The caching pressure has limited impact on scene recognition accuracy; sampling steps exceeding 5 seconds may cross scene change nodes, leading to insufficient scene drift recording; to facilitate subsequent scene segmentation, the scene signal sampling step is smaller than the minimum judgment window for scene segmentation, avoiding single sampling intervals exceeding the adjacent scene change recognition interval; the creative configuration status data stream writes the initial state when the session is established, and immediately writes the changed state when creative adjustments occur; the feedback event data stream is written using an event-triggered method, recording immediately upon receiving feedback such as exposure, visibility, dwell time, click, bounce, conversion, favorite, inquiry, and close, to ensure that the order of event arrival and the order of creative status changes are aligned within the same session;

[0072] To ensure session initialization quality, an initialization verification mechanism is implemented for optimized ad sessions. After session establishment, the session first enters the initialization state. When the basic ad data, initial creative version, and ad placement context are complete, the session transitions to the data collection state. If any key fields are missing, the session enters the verification pending state. Key fields include at least the ad object identifier, ad placement identifier, media page type, current creative version, and display start time. If any of these key fields are missing, the current session enters the verification pending state. After supplementary data collection, when all key fields are complete, the current session transitions to the data collection state. If one or more key fields are still missing after supplementary data collection, the current session is marked as a low-completeness session. A low-completeness session indicates incomplete input fields; only session records and observation permissions are retained, and the session does not enter the real-time adjustment process.

[0073] In the pending verification state, one to three supplementary data collections are allowed, with a supplementary data collection interval ranging from 100 milliseconds to 1000 milliseconds, preferably 300 milliseconds. The consideration for this setting is that one supplementary data collection may not be sufficient to cover short-term interface jitter, while more than three supplementary data collections will significantly prolong the session initialization time and affect the start-up of the real-time link. When the supplementary data collection interval is less than 100 milliseconds, the interface state is usually not yet recovered, while when it is more than 1000 milliseconds, it is easy to cause excessive lag in the session start-up judgment. The preferred interval of 300 milliseconds is because this time can reserve the necessary buffer for interface recovery without significantly prolonging the initialization processing time. After the supplementary data collection is completed, if the key fields are still incomplete, the existence and missing status records of the session are retained and will not be entered into subsequent real-time optimization processing.

[0074] The ad optimization session also sets an initial creative configuration return timeout to determine whether the current session meets the trusted conditions to enter the real-time optimization link. If the media interface does not return the initial creative configuration status within 0.5 to 2 seconds after the session is established, the current session is marked as a low-trust session, which is only allowed to enter the observation link and not the real-time adjustment link. A low-trust session means that although the input fields are complete, the return time of the key creative status exceeds the timeout. The reason for this setting is that the initial creative configuration is the basic data in the session startup phase. If it does not return within 2 seconds, it is difficult to bind subsequent feedback events with the same standard as the initial creative version, which can easily affect the judgment of the source of subsequent changes. If the time is less than 0.5 seconds, normal network fluctuations and interface queuing may also be misjudged as abnormal, which is not conducive to practical application. Low-completeness sessions and low-trust sessions do not enter the real-time adjustment link, but only retain observation processing permissions. The former corresponds to incomplete input fields, and the latter corresponds to the timeout of the return of the key creative status.

[0075] To ensure that short-delayed feedback before the session closes is correctly received, the ad optimizes the session settings with a pending closure mechanism and a delayed feedback buffer mechanism. When the session expires, the ad leaves the screen, or the media returns to the display termination mark, the session transitions from the data collection state to the pending closure state. Short-delayed feedback events continue to be received in the pending closure state. A first delay threshold is used to determine the upper bound of short-delayed feedback to be received and included in the current session before it closes; its value ranges from 2 to 30 seconds, preferably 10 seconds. This threshold is set based on common feedback reporting delays. Feedback such as exposure, visibility, short dwell time, and instant clicks are usually reported within seconds after the display ends, thus allowing for continued... Feedback that is significantly longer delayed is more likely to be a subsequent action or delayed conversion and should not be directly included in the current session. If the first delay threshold is less than 2 seconds, normal short-delayed feedback may be misjudged as a mismatch event. If it is more than 30 seconds, too many subsequent actions will be included in the current session, weakening the clarity of the session boundary. 10 seconds is preferred because this duration can cover the normal reporting delay of most instant feedback, while avoiding premature inclusion of long-delayed actions in the current session. Feedback events exceeding the first delay threshold are transferred to the delayed feedback buffer for use in delayed feedback binding processing after the session is closed. After the buffer is empty, the session enters the closed state.

[0076] To control resource consumption per session, the session buffer sets an upper limit on the number of scene signal records. This upper limit ranges from 50 to 1000 records, with 200 records for regular feed ads, 100 for short video ads, and 300 for interactive page ads. This upper limit is determined by the session validity period and sampling step size, and is typically no less than the number of records obtained by dividing the session validity period by the sampling step size, with 10% to 50% reserved for supplementary recording and abnormal records. For example, in feed ads, with a session validity period of 60 seconds and a sampling step size of 2 seconds, the basic number of records is approximately 30. Combined with continuous observation... For supplementary data collection and exception caching, the upper limit can be set to 200 records. In short video ads, with a session validity period of 30 seconds and a sampling step size of 1 second, the basic number of records is approximately 30. Considering the short viewing time and concentrated state changes, the upper limit can be set to 100 records. In interactive page ads, with a session validity period of 180 seconds and a sampling step size of 3 seconds, the basic number of records is approximately 60. Considering the long page dwell time and numerous interactive states, the upper limit can be set to 300 records. If the upper limit of the number of records is too low, it may not be able to fully retain the main scene changes within the session period. If the upper limit is too high, it will increase cache usage and is not conducive to real-time processing.

[0077] The session cache sets a single-session upper limit for the number of feedback events, ranging from 100 to 2000, with 300 to 1000 being preferred. The upper limit is determined based on the feedback density corresponding to the ad type. Click-driven ads have relatively low feedback density and can use a lower upper limit; interactive ads and high-frequency exposure ads have relatively high feedback density and can use a higher upper limit. This setting is because a single session needs to retain key feedback such as impressions, clicks, dwell times, conversions, closes, and inquiries, while controlling the number of duplicate sampled feedback events. If the upper limit is too low, key events may be lost, affecting subsequent attribution; if the upper limit is too high, it will significantly increase the single-session cache pressure and reduce real-time processing efficiency. After exceeding the feedback event upper limit, key feedback such as impressions, clicks, conversions, closes, and explicit inquiries are prioritized for retention, while duplicate dwell sampling data and data with consecutive identical states are compressed to ensure that key behaviors are not lost.

[0078] The ad optimization session also sets state transition rules; the session state includes at least the initialization state, the collection state, the pending verification state, the pending closure state, and the closed state; the initialization state is used to complete the basic field verification of the session and obtain the initial creative configuration; the collection state is used to continuously receive scene signals, creative configuration status, and feedback events; the pending verification state is used to handle the supplementary collection process after the key fields are missing; the pending closure state is used to receive short-delay feedback and perform a buffer before closing; the closed state is used to end the real-time reception of the current session and transfer the delayed feedback to the subsequent binding process; the above states flow in sequence according to the aforementioned conditions, and it is not allowed to skip the pending verification state and directly enter the closed state, nor is it allowed to re-receive real-time scene signals in the closed state, so as to ensure clear session boundaries and clear state changes.

[0079] Specifically, such as Figure 2 As shown: After the ad optimization session is established, continuous scene signal records are read from the session cache of the current ad optimization session, and the scene continuity judgment and scene segmentation are performed on the current ad display process. Based on this, the creative content is divided into creative unit groups that can be adjusted independently.

[0080] Scene continuity determination is completed through scene signature; scene signature includes at least the following fields: page type, main content theme tag, ad placement level, terminal type, time period tag, and interaction status level. If necessary, fields such as network level and preceding content tag can also be added. Page type, ad placement level, and terminal type reflect the basic media environment for ad display; the main content theme tag and preceding content tag reflect the content context of the ad; the time period tag reflects the access time interval; the interaction status level reflects the current browsing behavior stage; and the network level reflects changes in network status. The time period tag is generated according to preset time segments, which can be divided into 1-hour, 2-hour, or 4-hour segments. Adjacent time periods refer to two consecutive time segments within the same natural day. The interaction status level can be selected as five levels: fast swiping, continuous browsing, pause reading, active interaction, and pre-departure state. The network level can be selected as three levels: weak, medium, and strong, or from level 1 to level 5. Dividing the interaction status level and network level into limited levels is to ensure that subsequent continuity determination and scene segmentation can judge the continuity of the scene according to the range of level changes, avoiding inconsistent change boundaries.

[0081] The input for scene continuation determination includes at least the scene signature at the end of the previous ad optimization session, the scene signature at the beginning of the current ad optimization session, a list of adjacent scene snapshots, and stable scene markers from previous sessions; the output includes at least a continuation marker, a non-continuation marker, a pending continuation marker, and a new scene start marker; the scene signature at the end of the previous ad optimization session and the scene signature at the beginning of the current ad optimization session must be consistent in page type, ad placement level, and terminal type, identical in content theme main tag, belong to the same or adjacent time period in time period tag, and the interaction state level must not change from continuous browsing to the pre-departure state or from active interaction to the pre-departure state. When the status changes across levels, it is judged as continuation; when any key field in the content theme main tag, page type, or ad placement level changes between the two consecutive scene signatures, it is judged as non-continuation; when the page type and ad placement level remain consistent between the two consecutive scene signatures, but the content theme main tag only changes its sub-tag within the same theme category, and the network level fluctuation does not exceed one level, it is judged as continuation pending confirmation; the reason for setting the continuation pending confirmation status is that changes in sub-tags within the same theme category may correspond to either continuous evolution of the scene or a change in the scene. If it is directly judged as continuation or non-continuation, it may easily affect the accuracy of subsequent scene segmentation;

[0082] For the pending acceptance status, a short observation interval is entered to continue reading scene signal records. The short observation interval is 3 to 15 seconds, preferably 5 seconds. This duration is determined based on the scene signal sampling step size and the stabilization time after page switching. If the observation duration is less than 3 seconds, under the condition of the 1 to 5 second sampling step size set in the previous processing, it is usually insufficient to cover at least one complete scene confirmation cycle, and it is easy to misjudge short-term jitter as stable acceptance. If the observation duration is more than 15 seconds, it will significantly delay the subsequent scene segmentation processing and reduce real-time performance. 5 seconds is preferred because in common scenarios such as information flow, short video, and interactive page, this duration can usually cover the recovery interval after a brief fluctuation in the page state, and will not significantly lengthen the scene acceptance judgment delay. After the short observation interval ends, if the scene signature stabilizes again and meets the acceptance conditions, the pending acceptance mark is updated to acceptance. If the key fields deviate further, it is updated to non-acceptance and a new scene start mark is generated.

[0083] After determining the continuity of a scenario, the continuous scenarios within the current ad optimization session are segmented. Scenario segmentation uses a rolling division method, rather than fixed-time division. This is because scenario changes during ad display do not always align with fixed-time boundaries. Directly segmenting by fixed time could easily break up scenarios that should be continuous or merge scenarios that have already changed significantly. Scenario segmentation is based on scenario stability rules. These rules include at least the following conditions: unchanged page type, unchanged ad placement level, unchanged content theme main tag, unchanged terminal type, interaction state level change not exceeding one level, network level change not exceeding one level, and page browsing depth change rate not exceeding a preset range. Among these, page type, ad placement level, content theme main tag, and terminal type are primary stability rules because these fields directly determine the ad's display environment and content context. Interaction state level, network level, and page browsing depth change rate are secondary stability rules used to reflect the degree of behavioral fluctuation within the same display environment.

[0084] The threshold for the page browsing depth change rate is set at 5% to 30% per second, preferably 10% to 20%. The page browsing depth is determined according to the proportion of the currently browsed area to the browsable area of ​​the target page. This threshold is determined based on the page advancement speed in information feed browsing and short video scrolling scenarios. If it is lower than 5% per second, even slight browsing actions may cause the scene to be prematurely segmented, which is not conducive to forming a continuous observation interval. If it is higher than 30% per second, significant page advancement in fast scrolling scenarios may be misjudged as a stable scenario. The preferred value is 10% to 20% because this range can better distinguish between two common behaviors: continuous browsing and fast scrolling. It preserves the continuity in the normal browsing process while also being able to identify obvious changes in page advancement.

[0085] Continuous time periods that satisfy the scene stability rules are divided into the same scene segment. When any key field exceeds the corresponding stability range, the current scene segment ends and the next scene segment begins. The output of each scene segment includes at least the following fields: segment identifier, segment start time, segment end time, main scene signature, stability level, fluctuation flag, and comparability flag. The stability level is divided into three levels: high, medium, and low. A high level indicates that two or more consecutive sampling windows satisfy all stability rules. A medium level indicates that all major stability rules are satisfied, and there are only slight fluctuations in the auxiliary stability rules. A low level indicates that the major stability rules are violated, or the auxiliary stability rules fluctuate frequently. Using two or more consecutive sampling windows as the high-level judgment condition is to avoid instantaneous stability within a single sampling window being misjudged as continuous stability. Dividing the stability level into three levels is necessary because it is necessary to distinguish between three types of states: obvious stability, relative stability, and obvious fluctuation, and it is also necessary to control the number of levels for easy subsequent quick judgment. When there are too few levels, it is difficult to distinguish between different stability states; when there are too many levels, it will increase the complexity of configuration and judgment.

[0086] After scene segmentation, creative content is divided into creative unit groups according to independent adjustment boundaries. These independent adjustment boundaries are determined based on content function, display location, the impact of changes on other creative content, and the separability of rollback. The creative unit group approach is used because if the entire creative is always treated as the smallest processing object, it becomes difficult to distinguish which part of the creative content caused the change when ad performance changes, and it also hinders subsequent partial adjustments and rollbacks. Creative unit group division is based on the creative configuration state at the start of the session, combined with the ad creative resource table, brand restriction table, and ad review cache table. The inputs for creative unit group division include at least the creative configuration state at the start of the session, the ad creative resource table, the brand restriction table, and the ad review cache table. The outputs include at least the unit group identifier, unit group type, member version set, adjustable flag, dependency flag, conflict flag, and priority field.

[0087] Creative unit groups should include at least the following types: information expression group, visual presentation group, benefits description group, action guidance group, and jump configuration group. The information expression group should include at least the following fields: title, short description, and main selling points. The visual presentation group should include at least the following fields: main image, cover, background style, and color scheme. The benefits description group should include at least the following fields: promotional copy, price hints, and time-sensitive tags. The action guidance group should include at least the following fields: button copy and action prompts. The jump configuration group should include at least the following fields: landing page anchor point and jump entry configuration. Grouping creative content with similar functions into the same unit group aims to create clear boundaries for local adjustments while maintaining the overall stability of the creative structure. This facilitates subsequent judgment and control of changes based on functional boundaries.

[0088] The priority value for creative unit groups ranges from 1 to 5. This range is used because it distinguishes between low-impact, medium-impact, and high-impact unit groups while controlling the number of levels, facilitating rapid judgment during real-time processing. When the priority level is less than 3, it becomes difficult to differentiate the processing order between different unit groups. When the priority level is more than 5, the complexity of configuration and judgment increases significantly. A higher priority value indicates a greater impact of changes to the corresponding unit group on user perception, click behavior, or subsequent paths, and also indicates a higher adjustment risk. Without special business constraints, the priority of the action guidance group is preferably 2 to 3 because this type of unit group typically affects the click triggering method but has relatively less interference with the overall visual presentation and redirection path. The priority of the information expression group is preferably 3 to 4 because this type of unit group directly affects the user's understanding of the ad content. The priority of the visual presentation group and the redirection configuration group is preferably 4 to 5 because this type of unit group has a greater impact on the user's initial perception and subsequent access path, and the fluctuations after adjustments are more pronounced.

[0089] Dependency flags are used to constrain whether a creative unit group depends on the availability of another creative unit group or external business status. A dependency flag is set when the availability of a unit group is contingent upon the status of a previous unit group, external business status, or the validity of a target resource. For example, the benefits description group depends on the promotion's validity period; after the current promotion ends, the corresponding member version is no longer a candidate for availability. The jump configuration group depends on the availability of the landing page; when the target landing page is inaccessible or the corresponding content is offline, the jump configuration group does not enter the adjustable state. Conflict flags are used to indicate whether two creative unit groups are allowed to be adjusted simultaneously within the same observation window. A conflict flag is set when simultaneous changes to two unit groups within the same observation window affect the subsequent determination of the source of change. For example, when the visual presentation group and layout style-related content are adjusted simultaneously, it is difficult to determine the source of change later, so they should not be applied simultaneously within the same observation window. When the action guidance group and the jump configuration group are adjusted simultaneously, click behavior and access paths will change simultaneously, which is also not conducive to distinguishing the source of influence later.

[0090] When a member version of a creative unit group has no available candidate version during the current ad optimization session, or when the ad review cache table shows that the corresponding candidate version has expired, the adjustability flag for that creative unit group is set to disabled. This rule is adopted because subsequent creative adjustments must be based on the availability of candidate versions and the fulfillment of the corresponding review conditions; otherwise, even if the scene segmentation and change source judgment results are clear, no executable adjustment action can be formed. When the scene segmentation stability level in the current ad optimization session is low, only the division and recording of creative unit groups are allowed, and no executable adjustment qualification is generated for any creative unit group. The reason for this setting is that when the scene itself is in a state of significant fluctuation, even if the boundaries of the creative unit group are clear, it is not advisable to directly enter the subsequent adjustment chain; otherwise, scene fluctuations may be mistaken for creative problems.

[0091] Specifically, such as Figure 3 As shown: Within the same scene segment formed in the current ad optimization session, the change records of creative units are compared and correlated with feedback events to form a result for judging the source of ad performance changes; the input of the change comparison relationship includes at least the scene segment identifier, creative unit group status sequence, feedback event sequence, comparison status of the previous similar scene segment, and adjustment record table in the same ad optimization session; the output includes at least the change source judgment result, judgment confidence level, pending confirmation flag, and conflict flag; wherein, the creative unit group status sequence includes at least the current version status of the creative unit group, change time, adjustable flag, and conflict flag; the comparison status of the previous similar scene segment includes at least the corresponding key feedback status, creative unit group version status, and historical change direction record, which are used to judge the correspondence between the current feedback change and the creative change and scene change;

[0092] The determination of the source of change is accomplished using a hierarchical association approach. First, it checks whether a change has occurred within the current scene segment. If no change has occurred, it compares the current feedback event with the feedback status in the previous similar scene segment. When a significant change occurs in key feedback, the current source of change is prioritized as a scene change candidate. Key feedback is selected according to the advertising objective type. Click-through ads include at least post-exposure click-through rate and dwell time; conversion-through ads include at least post-click conversion rate and effective dwell time; interaction-through ads include at least interaction trigger rate and close rate. Effective dwell time can be determined by reaching a preset duration threshold or by the occurrence of a preset browsing behavior during the dwell time. The dwell time threshold is set between 3 and 30 seconds. For feed ads, a preferred threshold is 3 to 10 seconds; for short video ads, a preferred threshold is 3 to 8 seconds; and for interactive page ads, a preferred threshold is 10 to 30 seconds. This setting is because a dwell time that is too short is usually insufficient to reflect the true reading or interaction intent, while a dwell time within the appropriate range better reflects the user's actual attention to the advertising content.

[0093] The threshold for significant changes in key feedback is set at a relative change of 5% to 30%, preferably 8% to 15%. The relative change range is determined with reference to the baseline value of the corresponding key feedback in the previous segment of the same scenario. The reason for using a relative change threshold is that the feedback base for different advertising target types is different, and using a relative change range makes it easier to make consistent comparisons in different scenarios such as clicks, conversions, and interactions. When the threshold is too low, normal fluctuations may be misjudged as significant changes. When the threshold is too high, it is not conducive to timely identification of the source of change. The preferred range of 8% to 15% is because this range can usually cover the common range where advertising performance changes in a real way but does not reach extreme anomalies.

[0094] When a creative unit group changes, the arrival windows of feedback events before and after the creative change are correlated. These feedback event arrival windows include at least an immediate feedback window, a short-delay feedback window, and a delayed feedback window. The immediate feedback window is defined as 0 to 2 seconds, the short-delay feedback window as 2 to 30 seconds, and the delayed feedback window as 30 to 600 seconds. These three windows represent a general range; different ad types can select specific windows within this range that better align with their business rhythm. The reason for this window classification is that ad feedback events typically exhibit a combination of short, concentrated returns and subsequent delayed returns in their arrival time. The evaluation process requires differentiating feedback at different stages through layered windows. If the upper limit of the instant feedback window is too short, effective clicks, dwell times, and visible feedback within a short period after the creative change may not be fully included in the relevant scope. If the upper limit of the delayed feedback window is too long, too many subsequent behaviors will be mixed into the real-time evaluation process. The upper limit of the delayed feedback window is set to 600 seconds because most feedback such as short dwell times after clicks, short-link conversions, and page closures are usually completed within a few minutes after the display. If the upper limit is too low, some effective delayed feedback will not be included in the current evaluation. If the upper limit is too high, subsequent behaviors that are not directly related to the current change may be included in the real-time evaluation.

[0095] For feed click-based ads, it is preferable to use two windows: one for 0 seconds to 2 seconds and the other for 2 seconds to 20 seconds. For short video ads, it is preferable to use two windows: one for 0 seconds to 1 second and the other for 1 second to 10 seconds. For interactive page ads, it is preferable to use two windows: one for 0 seconds to 5 seconds and the other for 5 seconds to 60 seconds. The reason for this setting is that the click and dwell feedback of feed click-based ads is usually completed within a few seconds to tens of seconds after display, the feedback return speed of short video ads is even faster, and the dwell and interaction feedback duration of interactive page ads is relatively longer. If the window is set too short, it will be excluded before the feedback is fully received; if the window is set too long, it will reduce the targeting of real-time judgment.

[0096] After a change in the creative unit group, if the scene segmentation remains unchanged, and if the key feedback falls consecutively in the instant feedback window and the short-delay feedback window after the creative change, and the direction of change is consistent with the upward or downward direction of the confirmed record in the previous ad optimization session, then the change in ad performance is determined to be primarily caused by the creative adjustment; the aforementioned confirmed record can be the scene transition record or creative adjustment record in the previous ad optimization session; when the creative unit group change and the scene segmentation switch occur simultaneously within the same time window, the source of the change is not directly determined, but the current situation is recorded as a conflict marker and enters the conflict observation state;

[0097] The observation duration for the conflict observation state is set between 30 and 300 seconds, preferably 60 to 180 seconds. This is because if the observation duration is too short, the feedback from the superposition of scene and creative changes will not be fully apparent; if the observation duration is too long, it will significantly delay subsequent adjustment steps. The preferred duration of 60 to 180 seconds is because this range can adequately cover the main feedback return intervals after superimposed changes in information feed, short video, and interactive page scenarios. During the conflict observation state, subsequent feedback events are collected, and adjustments at the same or higher level as the current target unit group are frozen. The level here is determined by the priority of the creative unit group.

[0098] To differentiate between stable judgments that can be followed up and temporary judgments that still require further observation, a judgment confidence level is set. The judgment confidence level is divided into three levels: Level 1, Level 2, and Level 3. Level 1 indicates only a single feedback change with a small number of valid feedback samples, ranging from 5 to 20. Level 2 indicates two or more key feedbacks changing together, with 20 to 100 valid feedback samples. Level 3 indicates repeated consistent changes occurring in similar scenario segments, with more than 100 valid feedback samples, or repeated confirmation in two or more ad optimization sessions. Valid feedback samples must include at least feedback records that have completed session attribution binding, are not marked as mismatched feedback, and correspond to the key feedback type. The reason for using the above quantity range is that… When the sample size is less than 5, the randomness is strong, making it unsuitable for source identification. Once the sample size reaches 20 or more, the impact of a single abnormal feedback on the overall assessment begins to decrease. When the sample size reaches 100 or more, or after repeated confirmation across sessions, the trend usually exhibits stronger stability. Level 2 uses two or more key feedbacks changing together to avoid prematurely raising the assessment level due to random fluctuations in a single feedback. Level 3 uses two or more advertising optimization sessions for repeated confirmation to ensure confirmation stability while avoiding excessively long confirmation cycles. When the assessment confidence level is Level 1, only candidate sources of change are generated; no executable adjustment triggers are directly generated. When the assessment confidence level is Level 2 or 3, the source of change assessment results can be used for subsequent adjustment control.

[0099] To ensure comparability, the concept of segmentation based on similar scenarios is introduced. Segmentation based on similar scenarios requires that two consecutive scenario segments maintain consistency or high similarity in page type, ad placement level, terminal type, and main content theme tag. High similarity must meet at least the following conditions: the main content theme tag belongs to the same higher-level category, the time period tag differs by no more than one level, and the network level changes by no more than one level. Using no more than one level as the similarity condition allows for slight fluctuations while maintaining basic environmental consistency. If the similar scenario condition is not met, the feedback changes are only used as auxiliary observation data and not as the basis for determining the source of change. The reason for using segmentation based on similar scenarios is that judging the source of changes in advertising effectiveness needs to be based on the premise of a basically consistent display environment. Directly comparing scenarios with significant differences can easily lead to misjudging scenario differences as creative changes.

[0100] The system also handles the attribution of asynchronous feedback. If a feedback event arrives after the current ad optimization session has ended, but carries an impression tracking code, or can be bound to the current scenario segment through ad object identifier, impression identifier, and time window, then the feedback event is allowed to be added back to the corresponding creative change record. If it cannot be explicitly bound, it is marked as mismatched feedback and will not participate in real-time judgment. The reason for setting up the add-back rules is that feedback such as click-through dwell, short-link conversion, and page closing may arrive late after the session ends. If all of them are directly excluded, it will reduce the completeness of the change source judgment. If add-back is done directly without binding conditions, it is easy to incorrectly classify irrelevant feedback into the current scenario segment.

[0101] When two creative unit group changes occur consecutively within the same scene segment, and the interval between the two changes is less than the minimum separation threshold for the current ad type, the change source determination result of that scene segment is marked as a mixed influence state. The minimum separation threshold is used to limit the minimum interval between two creative unit group changes within the same scene segment, and it ranges from 3 seconds to 60 seconds. The reason for this setting is that when the interval is less than 3 seconds, the feedback caused by the previous change has not yet fully arrived, and the subsequent change has already occurred, making it difficult to effectively distinguish at the feedback level; when the interval is greater than 60 seconds, it will reduce the efficiency of real-time adjustment. For feed ads, the minimum separation threshold is preferably 10 seconds; for short video ads, the minimum separation threshold is preferably 5 seconds; for interactive page ads, the minimum separation threshold is preferably 20 seconds. Different ad types have differences in user dwell time and feedback return rhythm, so different preferred minimum separation thresholds are used. Scene segments marked as mixed influence states are only retained as observation records and are not directly used as the basis for subsequent adjustment triggers.

[0102] Specifically, such as Figure 4As shown: Based on the change source determination results, scene segment stability, creative unit group adjustment status, and feedback confirmation level generated in the previous processing, an adjustment control decision for the target creative unit group is generated. The inputs to the adjustment control include at least the current scene segment stability level, change source determination results, determination confidence level, creative unit group adjustability flag, creative unit group conflict flag, current session cumulative adjustment count, last adjustment time, feedback confirmation level, media dynamic replacement permission flag, and advertiser restriction rules. The outputs include at least the target unit group identifier, target adjustment timing, adjustment path identifier, execution permission flag, and fallback pointer. The current session cumulative adjustment count and last adjustment time are from the Ad Optimization Meeting. The adjustment record table in the text; the feedback arrival completeness is used to characterize whether the key feedback on which the current adjustment control depends has reached a identifiable state. It can be determined based on the key feedback set corresponding to the current advertising target type, the arrival status of each key feedback in the corresponding feedback window, the number of effective feedback samples, and the stability of the feedback direction; the key feedback set is consistent with the previous change source judgment stage. Among them, the key feedback for click-oriented advertising includes at least the click-through rate and dwell time after exposure, the key feedback for conversion-oriented advertising includes at least the conversion rate after click and effective dwell time, and the key feedback for interaction-oriented advertising includes at least the interaction trigger rate and close rate; the feedback confirmation level is divided into three levels: Level 1, Level 2, and Level 3;

[0103] Level 1 indicates that at least one key feedback has arrived within the corresponding feedback window, but the set of key feedback has not yet reached the minimum coverage condition corresponding to the current advertising goal, or the number of valid feedback samples has only reached the initial judgment lower limit; Level 2 indicates that the main key feedback corresponding to the current advertising goal has reached the minimum coverage condition, and the number of valid feedback samples has reached the execution judgment threshold; Level 3 indicates that the main key feedback has arrived sufficiently, the feedback direction remains consistent within the continuous observation window, and the number of valid feedback samples has reached the stable confirmation threshold; the minimum coverage condition can be judged based on the arrival of no less than two key feedback types, or the simultaneous arrival of a single key feedback and auxiliary confirmation feedback; if key feedback has arrived but the feedback directions conflict, or only returns sporadically within the delayed feedback window without forming continuous confirmation, the feedback confirmation level is no higher than Level 1; if key feedback has reached the minimum coverage condition but some delayed feedback has not yet returned, the feedback confirmation level is no higher than Level 2; if the key feedback set is complete, the number of samples reaches the stable confirmation threshold, and the feedback direction does not show reverse fluctuations within the continuous observation window, the feedback confirmation level is determined to be Level 3; after adopting the above classification method, the feedback confirmation level can form a corresponding relationship with the judgment confidence level, feedback arrival status, and execution threshold, which facilitates subsequent state flow control calls;

[0104] Adjustment control is implemented using a state machine approach. The adjustment states include at least four states: observation, preparation, execution, delay, freeze, and recovery. The observation state indicates that the source of change is unstable and current conditions are insufficient for execution. The preparation state indicates that candidate unit groups can be selected, but the actual execution conditions have not yet been met. The execution state indicates that the target unit group is executed once. The delay state indicates that although a judgment result has been generated, execution is still suspended due to constraints such as the cooling window, feedback confirmation level, media availability, and other conditions. The freeze state indicates that adjustment is suspended due to conflicts, anomalies, excessive frequency, or resource unavailability. The recovery state indicates that a rollback or partial repair is still needed after the previous adjustment. This state division is adopted because real-time optimization of advertising creatives is not executed immediately after a single trigger, but requires repeated switching between observation, preparation, execution, delay, freeze, and recovery states to meet the requirements of adjustment timeliness, distinguishability of change sources, and execution stability.

[0105] When the source of change is determined to be caused by a change in the scene, the creative idea is not adjusted directly, but instead enters an observation or delayed state. Only when the scene change occurs more than twice consecutively in the same type of scene, and the current creative unit group has a clear negative inheritance record in that type of scene, is it allowed to enter the preparatory state. Using more than two consecutive occurrences as the condition for entering the preparatory state is to avoid prematurely initiating creative adjustment preparation due to a single scene fluctuation. When the source of change is determined to be caused by a creative adjustment, and the confidence level reaches level two or above, a target unit group can be selected from the creative unit groups associated with the current negative change. Using level two or above as the threshold for selecting the target unit group is to avoid triggering actual adjustments based on a single feedback change or a small number of samples.

[0106] The selection of target unit groups is determined in the following order: low coupling priority, low risk priority, rollback priority, and minimum single change priority. Low coupling priority means prioritizing creative unit groups with fewer dependencies to reduce the scope of impact. Low risk priority means prioritizing creative unit groups with lower priority but proven to affect the effect to reduce overall display fluctuations. Rollback priority means prioritizing creative unit groups with stable old versions that can be rolled back to, so as to quickly recover if the adjustment fails. Minimum single change priority means prioritizing the adjustment of only one creative unit group in the same execution cycle, starting with the smallest change. The reason for this setting is that the selection of target unit groups should not only consider whether there is room for adjustment, but also whether it is easy to distinguish the source of change after adjustment and whether it is easy to recover in abnormal situations.

[0107] The timing of adjustments is constrained by the scene segment stability level, feedback confirmation level, and cooldown window. The cooldown window limits the minimum waiting time between two adjacent creative adjustments, ranging from 5 to 300 seconds. If the cooldown window is less than 5 seconds, the immediate feedback from the previous adjustment has not yet fully arrived, making it easy for it to overlap with the subsequent adjustment. If the cooldown window is greater than 300 seconds, the real-time adjustment capability will be significantly reduced. The preferred cooldown window for feed ads is 15 to 60 seconds, for short video ads it's 5 to 30 seconds, and for interactive page ads it's 30 to 180 seconds. This is because feed ads have a medium single-view duration and a relatively fast feedback return rate; short video ads have a shorter viewing window and need to complete local judgments more quickly; and interactive pages... Ad dwell time is relatively long, and the duration of post-click dwell and interactive feedback is even longer, thus requiring a longer waiting period. The cooldown window should also be coordinated with the preceding feedback window, at least covering the immediate feedback window, and prioritizing the coverage of short-delayed feedback windows, to avoid triggering adjustments again before the previous feedback has fully manifested. When the time since the last adjustment is less than the cooldown window, even if the source of the change is clear, it only enters a delayed state. When the feedback confirmation level is lower than level two, it also enters a delayed state. When the scene segment stability level is low, it enters a frozen state. The above constraints are adopted because creative adjustments must be based on feedback that has reached a level that can be judged, the scene itself is relatively stable, and the impact of the previous adjustment has been basically manifested; otherwise, it is easy to cause continuous trial and error and confusion in attribution.

[0108] The adjustment path specifies the order in which target unit groups switch from the current version to the candidate version. The adjustment path is set according to the principle of increasing adjustment magnitude. The adjustment path for the information expression group can be: first switch the short selling point text, then switch the title, and then switch the long description. The adjustment path for the action guidance group can be: first switch the button text, then switch the prompt, and then switch the button position. The adjustment path for the visual presentation group can be: first switch the cover text, then switch the partial color scheme, and then switch the main image. The principle of increasing adjustment magnitude is adopted because small adjustments have less interference with the overall display and are more conducive to subsequent judgment of the specific source of change. Only when the previous level of adjustment has been proven insufficient to improve the effect can a higher level of adjustment path be entered.

[0109] The data structure for adjusting the path should include at least the following fields: path identifier, unit group type, starting version, candidate version sequence, applicable scenario signature for candidate versions, minimum observation time for each level of switching, and whether skipping levels is allowed. The minimum observation time is set according to the risk level of the adjusted path: 10 to 60 seconds for low-risk paths, 30 to 180 seconds for medium-risk paths, and 60 to 300 seconds for high-risk paths. This is because low-risk paths usually only involve changes to local text or prompts, and feedback occurs quickly; medium-risk paths usually involve changes to titles, covers, or local visual elements, requiring a longer observation time; and high-risk paths involve changes to the main image, jump configuration, or overall perception, affecting a wider range and requiring more time for feedback confirmation. The aforementioned observation time should at least cover the main feedback return cycle corresponding to the target path. If the observation time is too short, a judgment may be made before sufficient feedback has arrived; if the observation time is too long, it will affect the subsequent real-time adjustment rhythm.

[0110] Skipping levels is not allowed under normal circumstances; skipping the previous level and directly entering the next level is only allowed when the previous level version on the same path produces negative results three or more times in the current similar scenario; using three or more consecutive negative results as the condition for skipping levels is to avoid prematurely skipping the low intervention level due to one or two local fluctuations, and to promptly enter the next adjustment level when the previous level version has been repeatedly verified as incompatible; the aforementioned three or more consecutive negative results include at least three confirmed negative verification records in the same scenario segment, or a cumulative total of three negative confirmations in two or more ad optimization sessions without any positive confirmations in between;

[0111] When the cumulative number of adjustments in a current session reaches the single-session adjustment limit, it will directly enter a frozen state. The single-session adjustment limit is set to 1 to 8 times according to the ad type. This range matches the session validity period, ad dwell time, and attribution stability. Too few adjustments make it difficult to complete the necessary testing, while too many adjustments can easily cause display fluctuations and attribution distortion. For feed ads, 2 to 4 adjustments are preferred because the duration of a single view is limited. For short video ads, 1 to 3 adjustments are preferred because the viewing time is shorter. For interactive page ads, 3 to 6 adjustments are preferred because the dwell time is longer and the feedback loop is richer. When the media dynamic replacement license mark is disabled, or the candidate version's review status expires, the execution license mark is directly set to disabled, and the corresponding target unit group is transferred to the observation state or frozen state. The media dynamic replacement license mark can be determined based on the media interface capabilities, ad slot replacement restrictions, and the current display stage. Advertiser restriction rules include at least non-replaceable fields, brand consistency restrictions, activity validity period restrictions, and redirection target restrictions.

[0112] When a target unit group has a conflict marker with another unit group that is already in the preparation state, only one target unit group is retained for subsequent execution judgment; the processing order prioritizes the comparison of creative unit group priorities; if the priorities are the same, the rollback cost is compared, and the unit group with the lower rollback cost is retained first, while the other unit group is put into the observation state; the rollback cost is determined at least based on factors such as whether a rollback version exists, the scope of the impact on the display structure after rollback, and the amount of content to be replaced after rollback; the reason for this setting is that if two unit groups with conflict markers enter the execution preparation state at the same time, they are likely to interfere with each other in the subsequent observation window, which is not conducive to distinguishing the source of change and recovering from anomalies;

[0113] The recovery state is used to handle rollback or partial repair needs after the previous adjustment. When the current adjustment has been executed, but subsequent observation records show that the key feedback corresponding to the target unit group continues to deteriorate, or the media returns abnormally, the candidate version fails, or the jump target is unavailable, the current unit group enters the recovery state. Continued deterioration can be judged by the key feedback decreasing in more than two consecutive observation windows, and the decrease reaches a preset negative threshold. In the recovery state, the rollback pointer is called first to restore to the previous stable version. If the previous stable version is unavailable, the most recently available version is selected for partial repair based on the path record. Partial repair includes at least restoring partial copy, restoring partial visual elements, or restoring the original jump configuration. The reason for this setting is that the real-time adjustment link should not only be able to execute new adjustments, but also have the ability to reversibly handle adjustment failures, thereby ensuring the continuity and stability of the ad display process.

[0114] Specifically, such as Figure 5 As shown: After the adjustment of the target creative unit group takes effect, the post-adjustment observation phase begins. During this phase, feedback association, result verification, anomaly rollback, and acceptance record updates are completed. The inputs for the post-adjustment observation phase include at least the following fields: adjustment action identifier, creative version before adjustment, creative version after adjustment, adjustment effective time, scene segment where the adjustment occurred, comparison records of similar scene segments, subsequent feedback event sequence, and historical acceptance records. The outputs include at least the following fields: feedback association results, verification conclusions, rollback conclusions, stable acceptance records, observed acceptance records, frozen records, anomaly records, and long-term effect archives.

[0115] In the post-adjustment observation phase, feedback events are first assigned and bound. Feedback attribution is prioritized according to the following order: display tracking code, session identifier, scene segment identifier, and adjustment effective time window. The display tracking code directly identifies the display source; the session identifier and scene segment identifier limit the time and context of the feedback; and the adjustment effective time window distinguishes the sequential relationship of multiple adjustment actions within the same ad optimization session. The adjustment effective time window can be determined from 0 seconds after the adjustment takes effect until the end of the corresponding main feedback window. When multiple adjustment actions exist within the same ad optimization session, feedback is prioritized to the action with the closest adjustment effective time and consistent scene segment. If it is still impossible to determine, the feedback event is marked as conflicting feedback, written only to the long-term performance archive, and not included in the current review. The reason for using the above binding order is that the display tracking code has the strongest directionality, the session identifier and scene segment identifier can further narrow the attribution range, and the adjustment effective time window can distinguish the feedback sources corresponding to multiple actions within the same session, thereby reducing attribution confusion.

[0116] After completing the feedback binding, the adjusted scene segment is compared with the same scene segment before the adjustment; the judgment criteria for the same scene segment follow the same scene rules formed in the previous processing; if the current scene segment does not meet the same scene condition as the comparison scene segment, the current adjustment enters the confirmation queue and is not subject to final verification; the records in the confirmation queue will re-enter the verification after meeting the same scene condition or after the delayed feedback is supplemented; the reason for this setting is that the result verification needs to be based on the premise that the display environment is basically the same. If the scene difference is too large, the change in key feedback may mainly come from the scene change, rather than being directly attributed to the creative adjustment;

[0117] The results verification adopted four categories of conclusions: positive, negative, neutral, and pending confirmation. A positive conclusion indicates that the key feedback has reached the preset improvement threshold after adjustment; a negative conclusion indicates that the key feedback has decreased or that there are obvious side effects; a neutral conclusion indicates that the change in key feedback has not exceeded the improvement threshold; a pending confirmation conclusion indicates that the number of feedbacks is insufficient, the scenario comparability is insufficient, or the delayed feedback has not been fully reached; obvious side effects include at least an increase in the close rate, an increase in the bounce rate, a significant decrease in dwell time, or an increase in invalid clicks; the aforementioned obvious side effects can be judged according to the corresponding negative indicator reaching the preset negative threshold; insufficient feedback can be judged according to the number of effective feedback samples being lower than the first-level judgment lower limit in the previous processing, or although the aforementioned lower limit is reached, the key feedback type does not cover the minimum feedback set required for the current advertising goal; the reason for adopting the above verification conclusion types is that the change in feedback after adjustment is not always simply manifested as an increase or decrease, but may also be due to insufficient change, insufficient samples, or unmet conditions, requiring the use of multi-level conclusions to distinguish the subsequent follow-up and rollback methods;

[0118] The improvement threshold is set at a relative change of 5% to 20% based on the advertising goal type; 8% to 12% is preferred for feed click scenarios, 5% to 10% for conversion scenarios, and 10% to 15% for short video dwell scenarios. The relative change range is determined with reference to the benchmark value of the corresponding key feedback in the same scenario segment before the adjustment. The reason for using the above range is that feed click behavior fluctuates moderately, and if the threshold is too low, normal fluctuations may be misjudged as improvements; the conversion feedback base is usually small, so a lower threshold is needed to identify effective changes; short video dwell behavior is more affected by the content rhythm, so a relatively higher threshold is needed to exclude the influence of content fluctuations; if the improvement threshold is set too low, normal fluctuations may also be judged as effective adjustments; if the improvement threshold is set too high, effective improvements that already have carrying value may be ignored.

[0119] When the verification conclusion is positive, the adjustment record is written into the stable acceptance record. The stable acceptance record should include at least the following fields: ad target identifier, scene signature, target unit group identifier, adjustment path level, version before adjustment, version after adjustment, first confirmation time, number of confirmations, stability level, and applicable expiration time. The applicable expiration time can be calculated based on the activity validity period, material validity period, brand restriction period, or the most recent confirmation time. The initial number of confirmations is recorded as 1. When a positive conclusion is obtained for the first time but the number of confirmations has not yet reached the formal stability threshold, it is recorded as preparatory stability. In subsequent ad optimization sessions, when a positive result in the same direction appears again under the same or highly similar scene signature, the number of confirmations increases. When the number of confirmations reaches 2 to 5, the stability level is upgraded from preparatory stability to formal stability, preferably 3. The reason for this setting is that a single positive result may still be affected by local fluctuations. Only after the number of confirmations reaches more than 2 can it be initially shown that the path has reuse value. Upgrading after the number of confirmations exceeds 5 will delay the formation of the stable record. 3 is preferred because this number can achieve a balance between confirmation stability and acceptance timeliness.

[0120] When the verification conclusion is neutral, it is not written into the formal stable entry, but into the observation and acceptance record. The observation and acceptance record is used to save the state of the current path in the target scenario that has been executed but has not yet formed a clear positive or negative conclusion. When it reappears in the same or highly similar scenario, it can be used as a reference for the second-best candidate path. The reason for using the observation and acceptance record is that some adjustments will not immediately show a significant improvement or deterioration after the first execution, but they still have the value of continuing to observe and verify. If they are discarded directly, the acceptance utilization rate will be reduced.

[0121] When the verification conclusion is negative, an abnormal rollback is executed according to the intervention level of the adjustment path. The rollback method is determined according to the intervention level of the adjustment path. For low-intervention paths, immediate rollback is preferred, restoring to the previous stable version within 1 to 30 seconds after receiving a negative conclusion. The reason for this setting is that low-intervention paths usually only involve local text, local prompts, or other low-impact changes, making rollback simple and suitable for rapid recovery after a negative conclusion appears. If the rollback time limit is too short, it may trigger repeated operations before the media response is completed; if the rollback time limit is too long, it will prolong the duration of the negative display. Intervention paths can first enter a short observation recovery state, and then observe for an additional 10 to 60 seconds. If the result is still negative after a few seconds, a rollback is executed. The reason for this setting is that intervention paths usually involve changes in the title, cover, or local visual elements, and some feedback has a short delay, so extra observation time needs to be reserved to avoid premature rollback. If the observation time is too short, the trend cannot be fully confirmed; if the observation time is too long, the recovery action will be delayed. Once a negative conclusion is reached in a high-intervention path, a rollback is executed immediately, and the current path is written to the freeze record under the scene signature. The reason for this setting is that high-intervention paths usually involve changes in the main image, jump configuration, or overall perception, and their impact range is larger. Continuing to retain a negative path can easily significantly amplify the adverse effects, so the processing is no longer delayed.

[0122] The freeze record includes at least the following fields: ad object identifier, scene signature, target unit group identifier, freeze start time, freeze end time, and freeze reason level. The freeze reason level is used to characterize the severity of the reason why the current path is prohibited from continuing to enter the processing stage, and can be classified according to the source of the anomaly, the recoverability of the anomaly, and the scope of impact on the reuse of subsequent paths. The freeze reason level can be divided into three levels: Level 1, Level 2, and Level 3.

[0123] Level 1 corresponds to short-term recoverable reasons, including at least a single media execution anomaly, a single negative feedback verification, and short-term resource unavailability. These can be re-entered for observation after a short freeze period. Level 2 corresponds to moderately restricted reasons, including at least candidate version failure, repeated negative verification, negative feedback after short-term observation recovery, and unresolved path-level conflicts. The corresponding path enters a medium freeze period and is restricted from further direct execution. Level 3 corresponds to severely restricted reasons, including at least rollback failure, version mismatch, persistent media activation failure, unavailable jump target, and repeated freeze triggering of the same path in similar scenarios. The corresponding path enters a longer freeze period and is prohibited from reuse until the anomaly is cleared.

[0124] The determination of the freeze reason level is based primarily on the initial judgment of the anomaly source category, followed by adjustments based on whether the anomaly has been eliminated, whether it recurs across sessions, and whether it affects the integrity of the display structure. When the same path meets multiple freeze reasons simultaneously, the highest level is recorded. This grading method establishes a correspondence between the freeze reason level and the freeze period setting, subsequent path selection, and anomaly clearing processing. The freeze time is set from 30 seconds to 7 days. This range is chosen because the freeze processing must cover both temporary disabling in short-term anomaly scenarios and the disabling of obviously incompatible paths over a longer period. The preferred freeze time for feed ads is 10 minutes to 12 hours, for short video ads it's 5 minutes to 6 hours, and for interactive page ads it's 30 minutes to 24 hours. This setting is because the user dwell time, feedback return rhythm, and path reuse frequency differ among the three types of ads. Feed ads and short videos typically require faster recovery testing capabilities, while interactive page paths have a longer duration of impact and are more suitable for longer freeze periods. The maximum freeze time is no more than 7 days to prevent short-term path failures from being mistakenly extended to long-term disabling.

[0125] If media return failure, page display anomalies, version mismatch, or rollback failure occurs after this adjustment is implemented, it will be directly recorded as an anomaly record. An anomaly record should include at least the following fields: anomaly type, anomaly occurrence time, anomaly corresponding action identifier, whether manual review is required, anomaly clearance status, and subsequent recovery status. The anomaly clearance status can be updated based on the results of manual review, media recovery, or successful version replacement. Whether manual review is required can be determined based on the anomaly type. When the anomaly type is rollback failure, version mismatch, persistent media return failure, or repeated triggering of the frozen path, the manual review requirement flag can be set to "yes". For rollback failure anomalies, subsequent ad optimization sessions will check the anomaly flag before reading the acceptance record. If the anomaly is not cleared, the same path should not be used again. The reason for this setting is that rollback failure, version mismatch, and persistent return failure usually indicate that there is strong uncertainty in the path execution chain. If it is reused, it is easy to amplify the impact of the anomaly.

[0126] The system also handles the re-verification of delayed feedback; when conversion feedback arrives within the re-verification window after the current ad optimization session has been closed, and can be clearly attributed to a specific adjustment action, the original verification conclusion can be adjusted upwards or downwards based on the adjacent conclusion levels; for example, an original neutral conclusion can be upgraded to a positive conclusion, or an original positive conclusion can be downgraded to a neutral conclusion; the re-verification window uses 0 seconds to 300 seconds for click-related feedback, 0 seconds to 48 hours for conversion-related feedback, and 0 seconds to 24 hours for lead generation-related feedback; click-related... Feedback is typically reported within minutes of the display. If it is not reported within 300 seconds, the direct correlation with the current adjustment action is significantly reduced. Conversion-related and lead generation-related feedback may span longer behavioral chains, so a longer feedback window needs to be set. To avoid long-term delayed feedback disrupting real-time reception, feedback exceeding the feedback window is only written to the long-term performance file and does not modify the stable reception record or the observation reception record. The long-term performance file should include at least the following fields: feedback type, feedback time, corresponding advertising target identifier, and original verification conclusion.

[0127] After verification, the adjusted records will be used for subsequent ad optimization sessions. Once a subsequent ad optimization session is established, the official stable record that matches the current scenario signature will be read first. When multiple candidate records exist, they will be selected in the order of high confirmation count, recent confirmation time, and low freeze risk. Freeze risk is used to characterize the degree of restriction on the current path when it is used as a candidate in subsequent ad optimization sessions, and is comprehensively judged based on the number of historical freezes, the recent freeze time, and the level of freeze reason. Freeze risk can be divided into three levels: low, medium, and high. Low risk corresponds to paths with few historical freeze records, a long time since the last freeze in the current session, and a low level of freeze reason. Medium risk corresponds to paths with many freeze records, a recent freeze time since the last freeze in the current session, or a medium level of freeze reason. High risk corresponds to paths with repeated freezes, a short interval between the last freeze time and the last freeze in the current session, or a high level of freeze reason.

[0128] The risk of account freezing is determined in the following order: the level of the reason for freezing, the interval between the most recent freezing time and the current session, and the number of historical freezings. When two or more of the above factors reach a high level of restriction, the risk of account freezing is determined to be high. When only one factor reaches a high level of restriction, or multiple factors are at a medium level of restriction, the risk of account freezing is determined to be medium. All other cases are determined to be low. The above determination method is adopted to ensure that the formation process of account freezing risk is consistent with the severity of the anomaly, the proximity of the time, and the recurrence.

[0129] When multiple candidate paths are read in subsequent ad optimization sessions, paths with lower freezing risk are prioritized for sorting; paths with medium freezing risk are considered as supplementary candidates when there are insufficient formal stable records; paths with high freezing risk are not considered as direct candidates for the current session; if the anomaly corresponding to the current path has been cleared and new positive confirmations appear in subsequent observation records, it can re-enter the candidate path range; through the above processing, a unified priority relationship is formed among formal stable records, observational acceptance records, and frozen records; if there are no formal stable records in the current scenario, observational acceptance records are read as a preliminary reference; if there are neither formal stable records nor observational acceptance records, it is treated as a path without acceptance; if there are frozen records, the corresponding path is directly excluded.

[0130] Example 2: Based on Example 1, the specific application process of a real-time optimization method for advertising creatives based on dynamic scene perception is further explained:

[0131] Taking a news feed ad for a sports shoe product as an example, the ad is placed on a mobile content feed page, with the ad placement located in the middle of the page's recommended section. The current ad creative includes at least a title, short selling point text, main image, button text, and landing page entry configuration. When the target ad first enters the target ad placement, a corresponding ad optimization session is established, and fields such as ad task identifier, ad target identifier, ad plan identifier, ad placement identifier, media page type, terminal type, display start time, current creative version, strategy version, and parameter version are recorded. When the current ad is a news feed ad, the session validity period can be 60 seconds, and the scene signal sampling step size can be 2 seconds. After the ad optimization session is established, scene signals such as page type, content theme main tag, ad placement level, terminal type, network level, time period tag, user real-time interaction status, and page browsing depth are continuously received. At the same time, the current creative configuration status and feedback events such as exposure, click, dwell, and bounce are recorded and uniformly written into the session cache. This ensures that the scene information, creative status, and feedback data during the ad display process are aligned on the same time basis, serving as a unified input basis for subsequent processing.

[0132] During the ad optimization session, scenario continuation determination is first performed based on continuously collected scenario signals. In the current time period, the preceding content of the ad slot is sports and fitness content, the page type is an information feed list page, the terminal type is a mobile portrait screen, and the user is in a continuous browsing state. In the current time period, the page type, ad slot level, and terminal type remain unchanged, the main content theme tag remains "sports and fitness," only the sub-tag changes from "running training" to "outdoor hiking," and the network level fluctuation does not exceed one level. In this case, the current state can be recorded as pending confirmation of continuation, and a short observation period can be entered. The short observation period can be approximately 5 seconds. If, within this observation period, the user's interaction state remains continuous browsing or paused reading, and the page browsing depth change rate remains within the range of 10% to 20% per second, then... The current scenario is updated to a continuation, and the corresponding continuous time period is divided into the same scenario segment. If the main theme tag of the subsequent content changes to travel information unrelated to sports and fitness, or if the ad placement level changes significantly, the current scenario segment ends and a new scenario segment begins. After the scenario segmentation is formed, the creative content is further divided according to independent adjustment boundaries to obtain creative unit groups such as information expression group, visual presentation group, action guidance group, and jump configuration group. For example, the title and short selling point text can be divided into the information expression group, the main image into the visual presentation group, and the button copy into the action guidance group. After adopting the above method, subsequent processing no longer treats the entire creative as the only processing object, but uses the creative unit group as the smallest adjustment granularity to continue to judge the source of change and adjust and control it.

[0133] Within the same scenario segment, continue to compare and correlate changes and feedback events in the creative unit group. When the current ad is displayed for 20 seconds, if the ad title and main image remain unchanged, but the button text changes from "Buy Now" to "View New Products," then this change is recorded in the action guidance group's change sequence. Subsequently, the click-through rate and dwell time in the instant feedback window and short-delay feedback window after the 20th second are compared with the corresponding feedback in the same scenario segment before the adjustment. If the click-through rate increases within the window from 0 to 2 seconds, and the dwell time increases synchronously within the window from 2 to 20 seconds, and the direction of this change is consistent with the change in the button text after the previous ad optimization session, then... If the positive feedback record is consistent, the change in the current ad performance is primarily caused by creative adjustments. Conversely, if no creative unit group changes occur within the current scenario segment, but the click-through rate and dwell time significantly decrease compared to the previous similar scenario segment, and the decrease reaches the reverse range of the information flow improvement threshold, such as 8% to 15%, then the source of the change is prioritized as a scenario change candidate. If the button copy changes while the main content theme tag also changes significantly, the current situation is marked as a conflict and enters conflict observation mode. In conflict observation mode, feedback events can continue to be collected for the next 60 to 180 seconds, but further adjustments at the same or higher level as the current target unit group are no longer allowed.

[0134] After determining the source of change, adjustment control decisions are generated based on the scene segment stability level, judgment confidence level, and feedback confirmation level. When the judgment result indicates that the change was caused by a creative adjustment, the judgment confidence level reaches level two, the scene segment stability level is high, the feedback confirmation level reaches level two, and the time since the last adjustment exceeds the cooldown window, the execution state is allowed. At this time, target unit groups are selected in the following order: low coupling priority, low risk priority, rollback priority, and minimum single change priority. If both the information expression group and the action guidance group are identified as being related to the current negative change, and the action guidance... If a group has a lower priority and a lower rollback cost, the action guidance group should be prioritized as the target unit group. The adjustment path should be set according to the principle of increasing adjustment magnitude, and instead of directly entering the high intervention level, the lower-level path switch should be executed first. For example, the button text can be changed from "View New Products" to "View Now". If the subsequent observation results are still negative, the prompt or button position can be changed again. If the previous level version of the same path produces negative verification records more than three times in the current similar scenario, it is allowed to skip this level and directly enter the next level path, thereby reducing invalid attempts and preserving the step-by-step verification relationship.

[0135] After the target creative unit group is adjusted, the post-adjustment observation phase begins. Feedback is linked and results are verified for similar scenarios before and after the adjustment. After the button copy adjustment takes effect, subsequent clicks, dwell times, and close feedback are bound by display tracking code, session identifier, scenario segment identifier, and adjustment effective time window. Then, the current scenario is compared with similar scenarios before the adjustment. If the click-through rate increases by 10%, dwell time increases by 8%, and there is no abnormal increase in close rate or significant increase in bounce rate, a positive conclusion is reached, and this adjustment is recorded in the stable acceptance record. The stable acceptance record includes at least the following fields: ad target identifier, scenario signature, target unit group identifier, version before adjustment, version after adjustment, number of confirmations, and stability level. Upon confirmation, the stability level is recorded as pre-stable, and the number of confirmations is recorded as 1. In the subsequent two ad optimization sessions, if a positive result in the same direction occurs again in the same or highly similar scenarios, the number of confirmations increases to 3, and the stability level is upgraded to formally stable. If the close rate increases or the dwell time decreases significantly after the button copy is adjusted, a negative conclusion is formed. When the button copy belongs to a low-intervention path, the rollback pointer can be called within 1 to 30 seconds to restore to the previous stable version. When the title or cover adjustment belongs to an intervention path, a short observation and recovery phase of 10 to 60 seconds can be entered first before deciding whether to rollback. When the main image or jump configuration adjustment belongs to a high-intervention path, once a negative conclusion occurs, a rollback is immediately executed, and the path is written into the freeze record under the current scenario signature.

[0136] If media return fails to take effect, version mismatch occurs, or rollback fails after adjustment execution, an exception record is generated synchronously, and the exception clearance status is checked before reading the acceptance record in subsequent ad optimization sessions. If the exception is not cleared, the same path is prohibited from being used again. If delayed feedback arrives after the session is closed and can be clearly attributed to this button copy adjustment, the original verification conclusion can be adjusted up or down according to the adjacent conclusion level. For example, the original neutral conclusion can be upgraded to a positive conclusion, or the original positive conclusion can be downgraded to a neutral conclusion. Feedback exceeding the rollback window is only written to the long-term performance archive and will not be modified in the official stable record. When a new ad optimization session is started, the official stable record that matches the current scenario signature is read first. When multiple official stable records exist, they are selected in the order of high confirmation count, recent confirmation time, and low risk of freezing. When no official stable record exists, the observation acceptance record is read as a preliminary reference. When the current path is in the frozen record, the corresponding path is directly excluded.

[0137] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0138] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented in software, the above embodiments can be implemented in whole or in part by a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions of the embodiments of this application are implemented in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted wirelessly or wiredly from one website, computer, server, or data center to another website, computer, server, or data center. Wired methods include optical fiber, twisted pair, coaxial cable, etc. Wireless methods include infrared, microwave, etc. Available media include any available media that can be accessed by a computer or data storage devices such as servers and data centers that contain one or more sets of available media. Available media can be magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0139] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0140] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time optimization of advertising creatives based on dynamic scene perception, characterized in that, include: S1. Establish an ad optimization session corresponding to the target ad display task, and continuously acquire scene signals, creative configuration status and feedback events in the ad display environment during the ad optimization session; S2. Based on the continuously acquired scene signals, determine the scene continuity and segment the current advertising display process, and divide the creative content into independently adjustable creative unit groups. S3. Within the same scene segment, associate the change records of the creative unit group with the feedback event to determine whether the change in advertising effect is caused by scene change or creative adjustment. S4. Based on the judgment results, combined with the stable state of the current scene segment, the adjustment state of the creative unit group, and the confirmation state of the feedback event, determine the adjustment timing and adjustment path of the target creative unit group. S5. After completing the adjustment of the target creative unit group, perform feedback correlation, result verification and anomaly rollback for similar scenarios before and after the adjustment, and use the verified adjustment records for subsequent ad optimization sessions.

2. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, Establish an ad optimization session corresponding to the target ad display task, and continuously acquire scene signals, creative configuration status, and feedback events in the ad display environment during the ad optimization session, including: Segment ad optimization sessions by target audience, ad placement, media page type, and session duration. Write scene signals, creative configuration status, and feedback events under a unified time base; Perform initial validation and supplementary data collection on key fields; Perform a reliability determination on the return sequence of the initial creative configuration; Receive short-delay feedback before the session ends, and perform cache limit control and session state transition management.

3. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, Based on continuously acquired scene signals, the current ad display process is judged for scene continuity and segmented, including: A scene signature is generated based on continuous scene signal records in the session buffer. The connection between the end scene of the previous session and the beginning scene of the current session is determined, and a short-term observation is performed when the connection is pending confirmation. The scenarios are segmented in a scrolling manner based on changes in page type, ad placement level, main content theme tag, terminal type, interaction status, network status, and page browsing depth, forming segment identifiers, main scenario signatures, stability levels, fluctuation markers, and comparability markers.

4. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, The creative content is divided into independently adjustable creative unit groups, including: Based on the creative configuration status at the start of the session, the ad creative resource table, the brand restriction table, and the ad review cache table; Creative unit groups are divided according to the functional boundaries of information expression, visual presentation, rights and interests explanation, action guidance, and jump configuration; Configure member version sets, adjustable tags, dependency tags, conflict tags, and priorities; Eligibility for adjustment is determined based on candidate version availability, external business status, target resource availability, and scenario stability level.

5. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, Within the same scene segment, the change records of creative unit groups are linked with feedback events, including: Based on scene segmentation identifiers, creative unit group status sequences, feedback event sequences, comparison statuses of similar scenes, and adjustment record tables; Within the same scene segment, feedback events before and after the creative change are hierarchically associated according to the instant feedback window, short-delay feedback window, and delayed feedback window. The attribution of delayed feedback is completed based on the display tracking code, session identifier, scene segment identifier, and adjustment effective time window.

6. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, Determining whether changes in advertising effectiveness are caused by changes in the scenario or by adjustments to the creative content includes: First, determine whether any changes have occurred to the creative unit groups within the current scene segment; Key feedback is selected based on the type of advertising objective, and confidence level is determined by combining the feedback benchmark, change direction record and number of valid feedback samples from the previous segment of the same scenario. Execute conflict observation when creative changes and scene transitions occur within the same time window; When the interval between multiple creative changes within the same scene segment is less than a preset threshold, a mixed influence flag is executed. When the conditions of the same type of scenario are met, generate scenario change candidates or creative adjustment candidates.

7. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, Based on the judgment results, combined with the stable state of the current scene segmentation, the adjustment state of the creative unit group, and the confirmation state of the feedback event, including: Adjustment control decisions are generated based on the results of the change source determination, the stability level of the scene segment, the determination confidence level, the feedback confirmation level, the adjustable mark of the creative unit group, the conflict mark of the creative unit group, the cumulative number of adjustments in the current session, the last adjustment time, the media dynamic replacement permission mark, and the advertiser restriction rules. Based on the scene change judgment result or the creative change judgment result, the target unit group is controlled to switch between the observation state, the preparation state, the execution state, the delayed state, the frozen state and the recovery state.

8. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, Determine the timing and path for adjustments to the target creative unit group, including: The target creative unit groups are determined based on the principles of low coupling priority, low risk priority, rollback priority, and minimum single change priority. The timing of adjustments is determined based on the cooling window, the time of the last adjustment, the stability level of the scene segment, and the feedback confirmation level. Configure candidate version switching paths, switching observation durations at each level, and conditions for skipping levels according to the principle of increasing adjustment magnitude; When the cumulative number of adjustments exceeds the limit, there are unit group conflicts, negative path accumulation, media is disabled, review fails, or feedback continues to worsen, restrictions, freezes, rollbacks, or restorations will be implemented.

9. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, After adjusting the target creative unit groups, feedback correlation, result verification, and anomaly rollback are performed on similar scene segments before and after the adjustment, including: Based on the action adjustment identifier, creative versions before and after the adjustment, adjustment effective time, scene segment where the adjustment is located, comparison records of similar scene segments, and subsequent feedback event sequence, the attribution binding is performed according to the display tracking code, session identifier, scene segment identifier, and adjustment effective time window. When the conditions of similar scenarios are met, the key feedback improvement threshold and side effect judgment rules are used to verify and classify the data, and the verification conclusions are recorded in the stable acceptance record, the observation acceptance record, the frozen record or the abnormal record. When the verification result is negative, the intervention level of the adjusted path will be adjusted to implement immediate rollback, short observation recovery, or freeze.

10. The method for real-time optimization of advertising creatives based on dynamic scene perception according to claim 1, characterized in that, The verified adjustment records will be used for handling subsequent ad optimization sessions, including: Read the official stable records, observe the acceptance records and freeze records according to the scenario signature matching relationship, and determine the acceptance path of the subsequent advertising optimization session based on the number of confirmations, the most recent confirmation time and the risk of freezing. For delayed feedback within the supplementary window that can be clearly attributed, the verification conclusion will be adjusted according to the adjacent conclusion levels. Feedback exceeding the recovery window is written into the long-term performance archive; Perform exclusion processing or empty path processing for abnormal and uncleared paths, frozen paths, and paths without successors.

Images