Explosive material multi-modal generation method based on deep targeted penetration

Abstract

The invention discloses an explosive material multi-modal generation method based on deep targeted penetration, and particularly relates to the technical field of material multi-modal generation, and the method comprises the steps: carrying out the unified mapping, alignment, complementation and tagging processing of multi-source data, and reducing the accumulation risk of feedback distortion from the source; through sensitivity mapping and strategic conversion of distortion diagnosis features, the influence of delay and pollution is directly constrained as grade distribution, exploration proportion and sample screening rules in a material generation and delivery strategy, so that a generation link forms dynamic balance between quality control and diversity exploration, and meanwhile, the generation efficiency is improved. The materials are written into a candidate library after cross-modal consistency verification and trust scoring, strategy meta-information is bound, the content credibility and the execution traceability of a generated result are guaranteed, and in a delivery link, a tracking identifier and a return event link are additionally designed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material multi-modal generation, and more particularly, to a best-selling material multi-modal generation method based on deep targeted penetration. BACKGROUND

[0002] In the era of short videos, video materials have become the core determinant of advertising delivery effectiveness. How to produce high-conversion materials at low cost and high efficiency has become an increasingly prominent industry problem. Against this background, the best-selling material multi-modal generation system based on deep targeted penetration has attracted widespread attention. In the best-selling material multi-modal generation system based on deep targeted penetration, the timeliness of the closed-loop feedback link and the credibility of the feedback signal constitute the core of the system's adaptive ability, but when the two interact, they will produce an amplifying and self-reinforcing distortion effect. Specifically, if the feedback from the delivery end is delayed or partially ineffective, the system cannot correct the estimate of user preferences with real behavior in a short period. At the same time, once the feedback signal is contaminated by abnormal data such as false clicks, volume brushing, or non-target group interactions, these erroneous information not only directly distorts the direction of demand vector, but also is delayed and absorbed as "effective" samples by the model in a longer time window, causing the generation model to continue to adjust along the wrong gradient. The two superimposed form a vicious cycle in the closed loop: the contaminated historical feedback prompts the model to generate biased materials, and the biased materials further produce pseudo-high performance indicators (or non-target interactions) in delivery. These pseudo-performance are written back to the training and demand database, but due to the time delay of feedback processing, this kind of error signal has deeply affected multiple iterations before being identified and removed. The result is a series of quantifiable negative consequences such as demand vector distribution drift, systematic overestimation of predicted CTR / CVR, decline in material targeting, and budget waste. Moreover, since pollution and time delay are mutually causal, the correction cost is high and the recovery period is long. SUMMARY

[0003] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a best-selling material multi-modal generation method based on deep targeted penetration to solve the problems raised in the background art.

[0004] To achieve the above object, the present application provides the following technical scheme: The best-selling material multi-modal generation method based on deep targeted penetration comprises the following steps: Collecting multi-source original records and performing timestamp alignment and field standardization, mapping the delivery platform log, user behavior reply, channel feedback and external anti-fraud detection into a standardized event stream and writing into a time series storage; Performing end-to-end delay calculation and session fingerprint comparison on the standardized event stream, combining frequency mutation detection to label the delay attribute and pollution confidence for the event, and generating labeled event records; Record labeled events in short and long windows for batch aggregation and sliding statistics, combined with causal discrimination to extract delay contribution and pollution proportion, form a distortion diagnosis feature table and version storage; Map the distortion diagnosis feature table to get dynamic weight constraints, convert delay and pollution contribution to strategy level, exploration ratio and sample screening rules, and push to strategy warehouse in the form of strategy configuration file; Take the strategy configuration file and high-confidence event set as constraint input to drive the batch generation of script template, visual parameter and shot script, and generate materials after cross-modal consistency verification and trust score evaluation, and write into the candidate material library with strategy meta information; The candidate material library is attached with tracking identifier and distributed by channel according to the strategy when triggered, and the delivery request with identifier enters the user end to form a tracking back event stream and write into the time sequence storage; The tracking back event stream is aligned, completed and detected again, the delay and pollution labels are updated and written back to the database, and the standardized event stream is entered into the next cycle together to form a self-adaptive closed-loop management.

[0005] In a preferred embodiment, collect multi-source original records and perform timestamp alignment and field standardization, map delivery platform logs, user behavior feedback, channel feedback and external anti-fraud detection into standardized event stream and write into time sequence storage, as follows: Batch pull original records from delivery platform logs, user behavior feedback, channel feedback and third-party anti-fraud interfaces, and form original batch stream according to the crawling time; parse each original batch stream into an original event row; add crawling meta information and idempotent identifier to the original event row, and submit the event row with idempotent identifier to the time field completion set; The time field completion set performs three-order time priority selection at the event row level, records all original time fields and calculates the event arrival time difference, and assigns an original time sequence to each event; Estimate and align the original time sequence in the sliding window; adjust the unified time reference field according to the system unified reference time; Standardize the unified time reference event row according to the registrable semantic mapping table; perform duplicate detection on the standardized event row; Add derived labels to the de-duplicated standardized event row at the event level, and attach the derived labels to the event row as structured fields; Serialize the event row with derived labels according to the partition key sequence, and write it into the time sequence storage in columnar compression format.

[0006] In a preferred embodiment, end-to-end latency calculation and session fingerprint comparison are performed on the normalized event stream, combined with frequency mutation detection to label events with delay attributes and pollution confidence, and generate labeled event records, as follows: The normalized event stream is read from the time series storage, and the source context field is supplemented for each event to form an extended event view; The extended event view is extended to estimate the clock bias of each source by referring to heartbeat events and link acknowledgement, and the timestamp is corrected. The corrected timestamp is used to calculate the reporting delay, link transmission delay, and platform processing delay for each event, and the delay metrics are attached to the extended event view as delay attributes; The extended event view with delay attributes is used to construct session fingerprint vectors through synthetic device fingerprints and behavior sequences. Multi-scale hashing and similarity retrieval are used for comparison in the historical fingerprint library to generate log-level fingerprint similarity and session association chain, and the fingerprint similarity and association chain are written back to the event record; The event record with delay and fingerprint information is used to identify frequency mutations by fingerprint, IP, and channel event density statistics in a sliding window and comparison with an adaptive baseline. After identifying the mutation window, the events in the window are labeled with a mutation batch identifier; The events with mutation batch identifier are used to calculate single abnormal score according to delay abnormality degree, fingerprint repetition rate, mutation intensity, and behavior consistency, and to synthesize pollution confidence label by hierarchical confidence fusion strategy according to historical channel reliability, while multi-factor abnormal evidence is written into the event label table as a structured field.

[0007] In a preferred embodiment, the labeled event record is batch aggregated and sliding statistics in short and long windows, combined with causal discrimination to extract delay contribution and pollution proportion, to form distortion diagnosis feature table and version storage, as follows: The labeled event record is read from the time series storage, and the events are organized into an original batch stream according to batch identifier and timestamp order, and the original batch stream is written into a temporary batch table; The original batch stream is sliced into short and long window slices from the temporary batch table, and the slice results form short window batch segments and long window batch segment tables; Statistical quantities are calculated for each batch segment and written into a statistical view table; The statistical view table and historical "clean baseline" are compared in a hierarchical manner, and the deviation is marked as delay abnormality degree and pollution exposure degree by quantile difference and sequence trend detection, and the mutation window is written into a mutation index table as a Boolean identifier; The mutation index table and event snapshot of the corresponding batch segment are used as causal evidence to execute the causal discrimination process. The difference in indicators before and after intervention and the test conclusion are recorded as the preliminary estimate evidence set of delay contribution and pollution proportion. The preliminary estimated evidence set is stratified and weighted according to event confidence, channel risk score and window consistency to form the final delay contribution and pollution ratio, along with confidence intervals and evidence citations. The final delay contribution, pollution percentage, related statistics, evidence citations, and operational recommendations are written into the immutable distortion diagnostic feature table as structured row records.

[0008] In a preferred embodiment, the distortion diagnostic feature table is mapped with sensitivity to obtain dynamic weight constraints. Latency and contamination contributions are transformed into strategy levels, exploration ratios, and sample selection rules, and then pushed to the strategy repository in the form of a strategy configuration file, as follows: The diagnostic records are read from the distortion diagnostic feature table and each record is weighted and benchmarked to form a weighted diagnostic row with business weight and confidence interval; The weighted diagnostic rows are aggregated according to the time window and channel dimension to obtain a dynamic weight constraint set; The dynamic weight constraint set is mapped to the weight interval as a strategy level fault, the corresponding exploration ratio range and sample usage rules according to the preset strategy mapping rules, thus forming the strategy level mapping rules. Serialize the policy level mapping rules into a structured key-value format policy configuration file.

[0009] In a preferred embodiment, the strategy configuration file and the set of high-confidence events are used as constraint inputs to drive the batch generation of copy templates, visual parameters, and storyboards. The generated materials are written into the candidate material library after cross-modal consistency verification and trust scoring evaluation, along with strategy metadata, as follows: Receive the policy configuration file and the set of high-confidence events and parse them into a structured set of constraints and a list of event fields; Template retrieval and priority ranking are driven by a set of structured constraints and a set of high-confidence events; priority scores are estimated for matching candidates based on similarity and historical performance, and the scores and candidate IDs are written into the candidate template table; The selected template set is batch parameterized and variant expanded, and the parameter set and proportion strategy of each variant are written into the variant plan table. Trigger low-cost rendering and rapid compositing in parallel according to the variant plan, based on the text variants, visual parameter sets and storyboards; The text, images, and video samples in the temporary generation pool are subjected to cross-modal consistency verification, rule compliance checks, and historical similarity comparison; the semantic vector distance is used to obtain the similarity with the nearest historical best-selling products, and the verification evidence is written into the verification result table in a structured record along with the verification sub-item scores; The verification result table is combined with the generator confidence, event source confidence, and historical performance prediction to perform a confidence-weighted fusion calculation to calculate the trust score. The processing action is determined based on the trust score and policy constraints. The trust score, improvement suggestions, processing action, and traceable evidence package are written into the candidate material record. Candidate materials are persistently stored in the candidate material library as structured entries.

[0010] In a preferred embodiment, the candidate material library is marked with a tracking identifier when the delivery is triggered and allocated to channels according to a strategy. The delivery request with the identifier enters the user terminal, forming a tracking-enabled return event stream and writing it to the timing storage, as follows: Read the materials to be deployed and their complete metadata from the candidate material library, and generate a deployment task list according to the channel allocation rules defined in the strategy configuration file; A unique tracking identifier is generated for each campaign. The identifier combines the creative ID, channel ID, strategy version number, timestamp, and random hash to form an uncollision-free identifier, which is then written into the campaign list and creative metadata. The task list is sorted by channel, priority and delivery window, and then sent to each channel's delivery entry point through the scheduler interface. After receiving the delivery request, the channel renders the creative to the user's device and embeds tracking identifiers and strategy metadata to ensure that user interaction events are reported synchronously with tracking identifiers. When an interactive event occurs on the user end, the event data is bound to the tracking identifier and sent to the central backhaul bus according to the channel backhaul protocol. After receiving events, the central backhaul bus performs data cleaning and normalization processing to form a backhaul event stream with tracking.

[0011] The technical effects and advantages of this invention are as follows: 1. This invention introduces dynamic identification, quantitative diagnosis, and strategic constraints of delay attributes and contamination confidence levels throughout the entire process of content generation and delivery. This enables the closed-loop feedback chain to adaptively resist distortion effects, thereby achieving full controllability and iterative enhancement in the generation and optimization of viral content. By uniformly mapping, aligning, completing, and labeling multi-source data, the consistency of input data in terms of timeliness and credibility is ensured, reducing the risk of accumulated feedback distortion from the source. On the other hand, through the sensitivity mapping and strategic transformation of distortion diagnostic features, the impact of delay and contamination is directly constrained into the level allocation, exploration ratio, and sample selection rules in the content generation and delivery strategy. This creates a dynamic balance between quality control and diversity exploration in the generation stage. At the same time, after cross-modal consistency verification and trust scoring, the content is written into the candidate library and bound to strategy meta-information to ensure the credibility of the generated results and the traceability of execution. In the delivery stage, the design of additional tracking tags and feedback event links ensures that the true performance of each piece of content can be fully mapped to the feedback loop, forming an adaptive closed-loop governance under the iterative update of delay and contamination tags.

[0012] 2. This invention not only significantly improves the accuracy of material targeting and the controllability of generating best-selling products, but also effectively avoids the demand vector drift and budget waste caused by feedback distortion in traditional methods. It achieves robust optimization and continuous iteration in a highly dynamic environment, and has obvious technological innovation and practical application value. Attached Figure Description

[0013] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0014] 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.

[0015] Example: Figure 1 This invention presents a method for generating multimodal content for viral products based on deep targeting penetration, comprising the following steps: Collect original records from multiple sources, align them with timestamps and standardize fields, and map platform logs, user behavior receipts, channel feedback and external anti-fraud detection into a standardized event stream and write it into time series storage; Perform end-to-end latency calculation and session fingerprint comparison on the normalized event stream, combine frequency mutation detection to label the event with latency attributes and pollution confidence, and generate tagged event records; Labeled events are recorded in short and long windows for batch aggregation and sliding statistics. Delay contribution and pollution ratio are extracted by combining causal discrimination to form a distortion diagnosis feature table and store it in a versioned manner. The distortion diagnostic feature table is used to obtain dynamic weight constraints through sensitivity mapping. The latency and contamination contribution are transformed into strategy level, exploration ratio and sample screening rules, and pushed to the strategy repository in the form of a strategy configuration file. The strategy configuration file and the set of high-confidence events are used as constraint inputs to drive the batch generation of copy templates, visual parameters and storyboards. The generated materials are written into the candidate material library after cross-modal consistency verification and trust score evaluation, along with strategy meta-information. When the candidate material library is triggered, a tracking identifier is attached and the materials are allocated to channels according to the strategy. The delivery request enters the user terminal with the identifier, forming a tracking-enabled return event stream and writing it into the time sequence storage. The tracked event flow undergoes alignment, completion, and re-inspection, updates delay and contamination labels, and writes back to the database. Together with the normalized event flow, it enters the next cycle, forming an adaptive closed-loop governance.

[0016] We collect raw records from multiple sources, align them with timestamps, and standardize fields. We then map platform logs, user behavior receipts, channel feedback, and external anti-fraud detection data into a standardized event stream and write it into time-series storage, as detailed below: The system retrieves raw records in batches from platform logs, user behavior receipts, channel feedback, and third-party anti-fraud interfaces, and forms raw batch streams in batches according to the retrieval time. The raw batch streams also carry retrieval metadata (retrieval time, source identifier, batch ID, retrieval node identifier). The original batch stream is parsed into original event lines one by one. The parsing includes expanding the nested structure, parsing the encoding format, extracting the original time field and key fields (event generation time, reporting time, user ID, material ID, session ID, device fingerprint, etc.), and writing the lines that fail to be parsed into the error queue for backtracking. Add fetched metadata and an idempotent flag to the original event line. The idempotent flag is generated by combining (event ID / session ID / event content hash). The idempotent flag also serves as the unique key for subsequent deduplication and retry processing. The event line with the idempotent flag is then submitted to the time field completion set. The time field completion set performs a three-level time priority selection at the event row level: priority is given to the event generation time, followed by the reporting time, and then the capture time; at the same time, all raw time fields are recorded and the event arrival time difference (event reception time minus event generation time) is calculated, and each event is assigned a raw time sequence (generation time, reporting time, reception time). The original time series is clock skew estimated and aligned within a sliding window. Clock skew estimation is achieved through cross-source verification (comparison of cross-channel events in the same session, heartbeat / baseline event alignment) and sliding median skew filtering. Alignment is achieved by adjusting the system's unified base time to generate a unified time base field, and late events are labeled with "lateness level" according to a configurable strategy. In an optional example, the clock skew estimation is achieved through cross-source verification (comparison of events across channels within the same session, heartbeat / baseline event alignment) and moving median skew filtering, as follows: Suppose there is an ad campaign session (Session 12345) where the same user clicked on ads across three different channels (APP, H5, and third-party affiliate program). Each channel reported the time of the click event: APP reporting time: 2025-09-08 10:00:02; H5 submission time: 2025-09-08 10:00:06; Third-party alliance reporting time: 2025-09-08 10:00:11; Meanwhile, there is a heartbeat / baseline event inside the system, which records the actual baseline time that should have been triggered at that moment: 2025-09-08 10:00:00.

[0017] Step 1: Cross-source verification The system first aggregates events from the same session (Session-12345) across the three channels to form an event group.

[0018] Comparing the timestamps within the event group: APP (+2s), H5 (+6s), and Alliance (+11s), they differ from the base time (10:00:00).

[0019] Based on the heartbeat / baseline event, the system confirms that the true reference point is at 10:00:00. Therefore, the "deviation" for each source is as follows: APP deviation = +2 seconds, H5 deviation = +6 seconds, and Alliance deviation = +11 seconds.

[0020] The system records these deviations, forming a deviation sequence: 2, 6, 11.

[0021] Step 2: Moving median deviation filtering The system maintains a sliding window (assuming window size = 5 sessions) to collect cross-source biases from the most recent sessions.

[0022] The current window deviation set could be: 3, 2, 6, 11, 4 seconds.

[0023] Take the median within this window: after sorting, the numbers are 2, 3, 4, 6, 11, and the median is 4 seconds.

[0024] The system uses 4 seconds as a correction reference for the current time alignment.

[0025] Regarding the three events of Session-12345: APP original deviation +2 seconds → Corrected difference = -2 seconds (close to the median, fine-tuned to around 10:00:00).

[0026] H5 original deviation +6 seconds → Corrected difference = +2 seconds (closer to the median, corrected to be closer to the baseline).

[0027] Alliance original deviation +11 seconds → Corrected deviation = +7 seconds (exceeding a reasonable threshold, marked as a late event, while retaining the original value for retrospective purposes).

[0028] Final result After cross-source verification and moving median deviation filtering, the system uniformly generated an aligned baseline timestamp: 10:00:00 ± 4 seconds.

[0029] The APP event was aligned to 10:00:02 (high credibility).

[0030] The H5 event is aligned to 10:00:04 (with medium credibility).

[0031] The alliance event was marked as late and adjusted to 10:00:07 (low credibility, may trigger compensation mechanism).

[0032] The unified time base event row performs field standardization based on a registerable semantic mapping table: field name mapping, enumeration value mapping, numerical unit normalization, and missing value strategy application (imputation, marking, or retaining null values), while retaining the original fields as meta fields to meet traceability requirements. The standardized event rows are subjected to duplicate detection and idempotent writing preparation. Duplicate detection is determined by two criteria: idempotent flag and content similarity threshold. If duplicates are identified, a merging strategy is executed (the most complete time series and the highest confidence value field are retained). Idempotent writing logic is implemented for concurrent retry scenarios. In an optional example, the duplicate detection is determined through a dual method of idempotency flag and content similarity threshold, as follows: Retrieve event logs from two different sources: Event A: Event ID = evt_123, Session ID = sess_001, Material ID = mat_45, User ID = u_789, Event Generation Time = 2025-09-08 12:01:02, Click Type = Redirect Click.

[0033] Event B: Event ID = evt_123, Session ID = sess_001, Material ID = mat_45, User ID = u_789, Event Generation Time = 2025-09-08 12:01:04, Click Type = Jump.

[0034] Judgment process: Idempotency identification The system first calculates the idempotency indicator: For event A, the idempotency flag = hash(evt_123|sess_001|mat_45|u_789).

[0035] For event B, the idempotency flag = hash(evt_123|sess_001|mat_45|u_789).

[0036] The two events have identical idempotency identifiers, so the system considers them to be potential duplicate events.

[0037] Content similarity threshold discrimination The system further compares the field contents: Time difference = 2 seconds (within the allowed time jitter threshold of 5 seconds).

[0038] Click type field similarity (based on edit distance): The similarity between jump click and jump is 0.85, which is greater than the set threshold of 0.8. Based on this, the two records are highly overlapping in time and semantics, and are therefore identified as duplicate events.

[0039] Merge strategy system Leave one final event line: The time field uses the earliest generation time, 2025-09-08 12:01:02.

[0040] The click type field uses a more semantically complete jump click.

[0041] The event ID and user information are retained unchanged. The final merge result is written to the time series storage, and another record is marked as "duplicate merge".

[0042] Add derived tags to the event-level of the deduplicated standardized event rows: channel inference (path backtracking), device type resolution (UserAgent resolution), geolocation estimation (IP reverse lookup), and anti-fraud score merging (weighting of third-party scores and internal rules), and append the derived tags to the event rows as structured fields; Event rows with derived tags are serialized by partition key (date / hour, channel, batch ID) and written to time series storage in columnar compression format. At the same time, during the write transaction commit phase, partition information, start and end offsets and event counts are published to downstream consumption topics through message queues so that downstream consumption and related queries can be performed in real time. Write confirmation is written to the audit log and change log via atomic transactions. The audit log records the write time, partition, offset, idempotency flag, and processing result. The audit log triggers the monitor to evaluate write latency and packet loss rate, and sends an operation and maintenance notification through the alarm channel when the threshold is abnormal, so that manual intervention and backtracking can be performed.

[0043] End-to-end latency calculation and session fingerprint comparison are performed on the normalized event stream. Frequency mutation detection is then used to label events with latency attributes and contamination confidence, generating tagged event records, as detailed below: Read the normalized event stream from the time-series storage, and complete the source context fields (event number, event occurrence time, platform feedback time, batch identifier, device parameters, network information, session identifier, behavior type) for each event to form an extended event view and write it to a temporary table; The extended event view is used to estimate the clock deviation of each source and correct the timestamps by referring to heartbeat events and link receipts. The corrected timestamps are used to calculate the reporting delay, link transmission delay and platform processing delay for each record, and the delay metric is attached to the extended event view as a delay attribute for index query. Extended event views with delayed attributes construct session fingerprint vectors by synthesizing device fingerprints (device parameters + network features) and behavioral sequences. Multi-scale hashing and similarity retrieval are used to compare them in the historical fingerprint database to generate log-level fingerprint similarity and session association chains. The fingerprint similarity and association chains are then written back to the event log to support cross-event tracing. Events with delay and fingerprint information are recorded in a sliding window. Event density is statistically analyzed by fingerprint, IP, and channel, and compared with an adaptive baseline to identify frequency mutations. After identifying the mutation window, events in the window are labeled with mutation batches and the mutation intensity, mutation duration, and mutation source are recorded. Events with mutation batch identifiers will have individual anomaly scores calculated based on the degree of delay anomaly, fingerprint repetition rate, mutation intensity, and behavioral consistency. Based on the credibility of historical channels, a hierarchical confidence fusion strategy will be used to synthesize the individual scores into a contamination confidence label (the contamination confidence is obtained by weighted summation of the individual scores). At the same time, multi-factor anomaly evidence will be written into the event label table as structured fields for auditing purposes. The events with mutation batch identifiers will be calculated with individual anomaly scores based on the degree of delay anomaly, fingerprint repetition rate, mutation intensity, and behavioral consistency, as detailed below: Extracting delay features: For events with mutation batch identifiers, call their calculated reporting delay, link transmission delay, and processing delay, compare them with the historical average of the same source channel, and sum the comparison results by weight to obtain the delay deviation ratio, which is used as a measure of delay anomaly. Fingerprint feature extraction: By performing clustering on the session fingerprint vectors of events within the same mutation batch, the repetition rate within the cluster is calculated and compared with the similarity across batches to obtain the fingerprint repetition rate, which serves as a measure of fingerprint anomalies. Extracting mutation features: By differentiating the frequency of events within a batch from the baseline distribution, and calculating the mutation amplitude, duration of mutation, and growth rate, the mutation amplitude, duration of mutation, and growth rate are weighted and summed to obtain the mutation intensity, which is used as a measure of mutation anomaly. Extracting behavioral features: By performing sequence consistency detection on the sequence of behavioral types in the event, the sequence difference between the sequence and typical user behavior patterns is measured. The behavioral consistency score is obtained by weighted summing the abnormality ratio of behavioral time interval and operation sequence, which serves as a measure of behavioral anomaly. Generate individual scores: Normalize the delay anomaly metric, fingerprint anomaly metric, mutation anomaly metric, and behavior anomaly metric respectively, and record them as delay individual scores, fingerprint individual scores, mutation individual scores, and behavior individual scores respectively, and write them to the anomaly score field of the event record.

[0044] Sample governance is driven by the pollution confidence level in the event label table: for events with extremely high confidence, the event ID is appended to the isolation list and the isolation instruction is sent to the training and delivery pipeline through the message queue; for events with medium confidence, a low-weight identifier is written into the training pipeline and the bid is lowered or the exposure is limited during delivery; for events with low confidence, the normal weight and channel qualification are maintained. The isolation list and low-weight annotations are packaged into auditable evidence packages in batches (including original traces, time-delay sequences, fingerprint links, and mutation time series diagrams). The evidence packages enter the manual review queue. The manual review conclusions are written back to the event label table with review labels and the validated cleaned samples are synchronized to the offline retraining set through incremental training triggers to continuously calibrate the anomaly detector and contamination confidence.

[0045] Labeled events are recorded in short and long windows for batch aggregation and sliding statistics. Delay contribution and contamination percentage are extracted using causal discrimination to form a distortion diagnostic feature table, which is then versioned and stored, as follows: The event records are read from the time-series storage and organized into a raw batch stream (fields: event ID, batch ID, occurrence time, reporting time, device fingerprint, channel, behavior type, contamination confidence, delay tag) according to the batch identifier and timestamp order. The raw batch stream is then written to a temporary batch table.

[0046] Slice the original batch stream into short-window and long-window segments from the temporary batch table (e.g., short window 5–15 minutes, long window 1–6 hours), and form short-window batch fragment and long-window batch fragment tables (fields: window ID, start time, end time, list of included events) from the sliced ​​results.

[0047] For each batch of segments, calculate the statistics and write them into the statistics view table. The statistics are specified as follows: total number of events, number of credible events, effective signal rate, median delay, percentile delay, delay over-threshold ratio, fingerprint repetition rate, fingerprint entropy, abnormal interaction density ratio, conversion rate deviation (compared with historical baseline), and channel distribution entropy. Write the source snapshot pointer for each statistic to facilitate traceability.

[0048] It's important to note that fingerprint entropy refers to the information entropy calculated within a statistical window based on the distribution of device fingerprints (such as device parameter combinations, network feature vectors, session context, etc.). It measures the diversity and balance of fingerprints. A higher value indicates a more uniform distribution of device fingerprints, and a greater probability of diverse genuine users in the system; a lower value indicates a high concentration of fingerprints, potentially indicating the risk of repeated occurrences of a single or a small number of forged fingerprints. Channel distribution entropy, on the other hand, refers to the information entropy calculated within a statistical window based on the distribution of the number of events across delivery channels (such as ad placements, media sources, traffic entry points, etc.). It reflects the balance of channel contribution. A higher value indicates a more balanced distribution of events across multiple channels; a lower value indicates that events are concentrated in a few channels, potentially suggesting traffic hijacking, channel fraud, or single-point dependence in delivery.

[0049] The statistical view table was compared with the historical "clean baseline" in a stratified manner (same channel, same time period, same activity label). The deviation was identified by quantile difference and sequence trend detection. The deviation was labeled as delayed anomaly degree and pollution exposure degree. The mutation window was written into the mutation index table with Boolean identifier (including mutation intensity, duration, and first time point).

[0050] The deviation is labeled as delayed anomaly degree and pollution exposure degree, as detailed below: In the comparison between the statistical view table and the clean baseline, the deviation of delay-related indicators (such as delay median, delay percentile, and delay over-threshold ratio) is used as input. Hierarchical normalization is performed and weighted summation is performed to obtain the relative offset intensity. The relative offset intensity is labeled as the delay anomaly field to ensure that its value reflects the magnitude and stability of the delay deviation.

[0051] The deviation of pollution-related indicators (such as fingerprint repetition rate, fingerprint entropy, abnormal interaction density, conversion rate deviation, and channel distribution entropy) is used as input. Hierarchical normalization is performed and weighted summation is performed to obtain the pollution anomaly intensity. The pollution anomaly intensity is then labeled as the pollution exposure field to characterize the significance of potential feedback pollution.

[0052] Using the mutation index table and corresponding batch fragment event snapshots as causal evidence, a causal discrimination process was executed: first, a chronological test was performed; then, a matching control was performed (using a homogeneous control group); next, an intervention simulation was performed (removing or re-extracting suspicious events from the event snapshots and recalculating key indicators); finally, robustness tests (lag tests and placebo tests) were performed. The differences in indicators before and after the intervention and the test conclusions were recorded as a preliminary estimate of the contribution of delay and the proportion of pollution.

[0053] The preliminary estimated evidence set is stratified and weighted according to event confidence, channel risk score and window consistency: higher confidence sources are given higher weight, median aggregation is taken for consistent estimates of long and short windows, and inconsistent estimates are given lower confidence and marked as requiring manual review, thus forming the final delay contribution and pollution ratio, with confidence intervals and evidence citations.

[0054] The delay contribution refers to the proportion of the index deviation caused by feedback link delay in the overall distortion effect. Essentially, it is: if the system operates under ideal, delay-free conditions, what proportion of the difference between the reference value and the actual value of the index can be attributed to delay rather than other factors? The delay contribution is obtained by subtracting the control index value under delay-free conditions from the actual observed index value, and then dividing the delay deviation by the actual observed index value minus the baseline index value. The pollution percentage refers to the proportion of deviation caused by abnormal data (fake interactions, inflated traffic, non-target groups) in the overall distortion effect. Essentially, it represents the percentage of incremental information in the event set that leads to incorrect learning or delivery by the system due to pollution. The pollution percentage is obtained by statistically analyzing the deviation of indicators caused by pollution events and dividing it by the total deviation of indicators caused by all events. The final delay contribution, contamination percentage, related statistics, evidence citations, and operational recommendations are written into an immutable distortion diagnostic feature table as structured row records (field examples: diagnosis ID, window ID, batch ID, short / long window statistical snapshot pointer, delay contribution, contamination percentage, confidence interval, evidence package pointer, recommended action, version number, generation time, calculator version). The corresponding evidence package (event snapshot, intervention simulation results, robustness test log) is stored in the audit object storage. Finally, the versioning of the diagnostic snapshot and downstream notification actions are recorded in the change log to complete the auditable link.

[0055] The distortion diagnostic feature table is dynamically weighted through sensitivity mapping to obtain dynamic weight constraints. Latency and contamination contributions are transformed into policy levels, exploration ratios, and sample selection rules, and then pushed to the policy repository in the form of policy configuration files, as follows: The diagnostic records are read from the distortion diagnostic feature table and each record is weighted and benchmarked according to the business sensitivity table, activity priority and channel risk parameters to form a weighted diagnostic row with business weight and confidence interval.

[0056] The weighted diagnostic rows are aggregated by time window and channel dimension and cross-calibrated with the resource cost table, real-time capacity constraints and historical correction efficiency table to obtain a dynamic weight constraint set that takes into account business costs and executability.

[0057] The dynamic weight constraint set is mapped to the weight range as a strategy level fault (e.g., conservative, wait and see, exploratory) according to the preset strategy mapping rules, the corresponding exploration ratio range and sample usage rules (e.g., pooling / reducing weight / isolation), thus forming the strategy level mapping rules.

[0058] The policy level mapping rules are serialized into a structured key-value format policy configuration file, and diagnostic version number, evidence package pointer, generation timestamp, confidence interval and signature verification fields are embedded in the configuration file to ensure traceable and auditable policy metadata.

[0059] The strategy configuration file undergoes syntax and constraint consistency checks by the configuration mode validator, and automated verification is performed in the sandbox environment using historical playback and small-scale simulations. The verification results are attached to the configuration file's change history as a verification report and simulation metric snapshots.

[0060] Configuration files and verification reports are written to the policy repository through a controlled submission process, and version tags and immutable change logs are created in the repository. When the repository is written, a message queue is triggered to notify the downstream distribution system to conduct a canary deployment test.

[0061] The test metrics generated by the gray-scale deployment test are written back to the monitoring bus and compared with the decomposed metrics in the distortion diagnostic feature table. The comparison analysis determines whether to automatically roll back the configuration files in the repository to the most recent stable snapshot or promote the version to the full deployment strategy. At the same time, the rollback or promotion event is written to the audit log to complete the closed-loop governance.

[0062] Using strategy configuration files and a set of high-confidence events as constraint inputs, the system drives the batch generation of copy templates, visual parameters, and storyboards. The generated materials are then written into the candidate material library along with strategy metadata after cross-modal consistency verification and trust scoring evaluation, as detailed below: The system receives the policy configuration file and the set of high-confidence events, parses them into a structured set of constraints and a list of event fields, sets key-value constraints such as conservatism level, exploration ratio, diversity limit, channel preference, sensitive word blacklist and attention duration, and writes the constraint set into a temporary constraint table.

[0063] Template retrieval and priority ranking are driven by a set of structured constraints and a set of high-confidence events: using event keywords, sentiment, product attributes and channel preferences as search criteria, multi-scale matching is performed in the copywriting template library, visual style library and storyboard fragment library; priority scores are estimated for matching candidates based on similarity and historical performance, and the scores and candidate IDs are written into the candidate template table.

[0064] The priority score for matching candidates is calculated based on similarity and historical performance, as follows: Read the metadata fields of the candidate materials (including generation source, keyword weight, sentiment intensity, visual element distribution, historical delivery tags, etc.), and call the similarity retrieval tool to perform multimodal similarity comparison between each candidate and the historical high-quality material library to obtain semantic similarity score and visual structure similarity score.

[0065] The priority score is obtained by weighting and fusing the semantic similarity score and visual structure similarity score with the historical performance prediction indicators of candidate materials (CTR, CVR, interaction depth, retention time, etc., provided by the predictor). The selected template set is batch parameterized and expanded with variations: the text placeholders are filled with the event field list, the color palette and composition weights are generated with the event visual preferences, and the storyboard fragments are spliced ​​with the attention duration; controlled combination expansion (exhaustive search + sampling) is performed on each template according to the exploration ratio, and the parameter set and proportion strategy of each variation are written into the variation plan table.

[0066] The text variants, visual parameter sets, and storyboards are triggered in parallel according to the variant plan for low-cost rendering and rapid compositing: first, low-resolution preview text, multiple image drafts, and short timeline samples are generated to save resources; during the rendering process, the generator version, random seed, time consumption, and resource indicators are recorded and written to a temporary generation pool along with the rendering artifacts in an indexed manner.

[0067] The text, images, and video samples in the temporary generation pool are subjected to cross-modal consistency verification, rule compliance checks, and historical similarity comparisons: keyword-visual element mapping checks (verifying that the keywords in the text appear in the images / videos via OCR or object detection), temporal rhythm and text emphasis alignment checks, and brand elements and sensitive words checks; and semantic vector distance is used to obtain the similarity with neighboring historical best-selling products, and the verification evidence is written into the verification result table in a structured record along with the verification sub-item scores.

[0068] The verification result table, together with the generator confidence, event source confidence, and historical performance prediction, is weighted and fused to calculate the trust score. Based on the trust score and policy constraints, the following actions are determined: high trust is directly marked as a candidate; medium trust is accompanied by improvement suggestions and queued for secondary small-scale generation; low trust is marked and isolated or submitted for manual review. The trust score, improvement suggestions, actions, and traceable evidence package (event ID, template ID, parameter set, generation log, and verification record) are written into the candidate material record.

[0069] Candidate creatives are persistently stored in the candidate creative library as structured entries. Each entry includes a thumbnail / preview link, complete generation parameters, trust score, strategy version, evidence package pointer, channel adaptation tag, and priority score. The entries are then injected into the candidate priority queue according to priority. The priority queue also maintains gray-scale deployment eligibility, grouping suggestions, and rollback conditions, which can be directly read and executed by the deployment scheduler.

[0070] When a campaign is triggered, the candidate creative library attaches a tracking identifier and allocates it to channels according to the strategy. The campaign request, carrying the identifier, enters the user's client, forming a tracking-enabled event stream and writing it to the time-series storage, as detailed below: The system reads the creative materials to be deployed and their complete metadata (including strategy level, trust score, tracking identifier, channel compatibility tag and audit evidence) from the candidate creative material library, and generates a deployment task list according to the channel allocation rules defined in the strategy configuration file. The fields include creative ID, target channel, exposure time window, audience profile, tracking identifier and priority level.

[0071] A unique tracking identifier (tracking ID) is generated for each campaign. This identifier, combined with the creative ID, channel ID, strategy version number, timestamp, and random hash, forms an uncollision-free identifier and is written into the campaign list and creative metadata to ensure full traceability from campaign to feedback.

[0072] The task list is sorted by channel, priority and delivery window, and then sent to each channel's delivery entry point through the scheduler interface. During the delivery process, the task fields are validated for format, legality, and exposure time window conflict detection. A delivery execution log is generated and written to the scheduling audit table for subsequent review.

[0073] After receiving the delivery request, the channel renders the creative to the user's device and embeds tracking identifiers and strategy metadata (such as URL parameters, hidden IDs, or front-end event bindings) to ensure that user interaction events (clicks, pauses, swipes, interactions, etc.) can be reported synchronously with tracking identifiers.

[0074] When an interactive event occurs on the user's end, the event data is bound to the tracking identifier and sent to the central backhaul bus according to the channel backhaul protocol. At the same time, environmental information (network status, device type, geographical location, event timestamp) is attached to ensure that the event can be mapped to the corresponding material and policy version.

[0075] After receiving events, the central backhaul bus performs data cleaning and normalization processing: verifying the legality of tracking identifiers, filling in missing fields, correcting delays and time zones, merging duplicate events, forming a complete and comparable backhaul event stream, and writing it into time-series storage for real-time analysis and historical archiving.

[0076] During the writing of the backhaul event stream to the time-series storage, an index table is generated (by tracking ID, material ID, strategy version, channel, and time window), and an event link mapping is established to ensure that each backhaul event can be directly traced back to the candidate material, the delivery strategy, and the generated evidence package, providing a reliable data foundation for subsequent closed-loop distortion analysis and strategy optimization.

[0077] The event stream with tracking feedback undergoes alignment, completion, and re-inspection, updates latency and contamination labels, and is written back to the database. It then enters the next cycle together with the normalized event stream, forming an adaptive closed-loop governance process, as detailed below: Read the tracked event stream from the time-series storage, sort it by event ID, tracking identifier and channel dimension, and match the events with the original candidate material records, strategy version and user profile to form an event-material mapping table.

[0078] Perform alignment operations on the event-material mapping table: correct timestamps, unify time zones, correct network transmission delays, use heartbeats and link receipts as latency benchmarks, and generate corrected event streams to ensure that the returned events strictly correspond to the delivery materials and strategy versions.

[0079] Perform data completion on the corrected event stream: fill in missing fields (such as event type, dwell time, interaction status, device information) based on historical event sequences and user behavior patterns, and mark the source and confidence level of the completion to ensure data integrity for subsequent analysis.

[0080] The completed event stream is then re-examined: based on delay distribution, fingerprint repetition rate, frequency mutation and behavioral consistency, delay labels and contamination labels are recalculated, and combined with historical clean baselines for hierarchical confidence fusion to generate updated delay contribution and contamination ratio, while adding traceable evidence fields.

[0081] The updated delay and contamination labels are written back to the event database and merged with the original normalized event stream to form a unified event dataset with real-time feedback labels, providing standardized input for the next round of generation, strategy adjustment, and candidate material selection.

[0082] During the write-back process, versioned snapshots and change logs are created to record event IDs, tracking identifiers, tags before and after updates, confidence intervals, evidence package pointers, and processing times, ensuring that the evolution of each event is traceable and auditable.

[0083] The merged event dataset automatically triggers the next round of closed-loop generation and strategy optimization: the newly generated materials, updated strategy configurations, and high-confidence event sets will drive the generation of copy, visuals, and storyboards again. The returned events continuously enter the alignment, completion, and re-inspection cycle to achieve adaptive closed-loop governance, enabling the material generation and delivery strategy to dynamically match real user behavior and channel feedback.

[0084] This invention introduces dynamic identification, quantitative diagnosis, and strategic constraints of delay attributes and contamination confidence levels throughout the entire process of content generation and delivery. This enables the closed-loop feedback chain to adaptively resist distortion effects, thereby achieving full control and iterative enhancement of the generation and optimization of viral content. By uniformly mapping, aligning, completing, and labeling multi-source data, the consistency of input data in terms of timeliness and credibility is ensured, reducing the risk of accumulated feedback distortion from the source. On the other hand, through the sensitivity mapping and strategic transformation of distortion diagnostic features, the impact of delay and contamination is directly constrained into the level allocation, exploration ratio, and sample selection rules in the content generation and delivery strategy. This creates a dynamic balance between quality control and diversity exploration in the generation stage. At the same time, after cross-modal consistency verification and trust scoring, the content is written into the candidate library and bound to strategy meta-information to ensure the credibility of the generated results and the traceability of execution. In the delivery stage, the design of additional tracking tags and feedback event links ensures that the true performance of each piece of content can be fully mapped to the feedback loop, forming an adaptive closed-loop governance under the iterative update of delay and contamination tags.

[0085] This invention not only significantly improves the accuracy of material targeting and the controllability of generating best-selling products, but also effectively avoids the demand vector drift and budget waste caused by feedback distortion in traditional methods. It achieves robust optimization and continuous iteration in a highly dynamic environment, and has obvious technological innovation and practical application value.

[0086] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0087] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0088] 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.

Claims

1. A method for generating multimodal viral content based on deep targeted penetration, characterized by: Includes the following steps: Collect raw records from multiple sources, align them with timestamps and standardize fields, and map platform logs, user behavior receipts, channel feedback and external anti-fraud detection into a standardized event stream and write it into time series storage; Perform end-to-end latency calculation and session fingerprint comparison on the normalized event stream, combine frequency mutation detection to label the event with latency attributes and pollution confidence, and generate tagged event records; Labeled events are recorded in short and long windows for batch aggregation and sliding statistics. Delay contribution and pollution ratio are extracted by combining causal discrimination to form a distortion diagnosis feature table and store it in a versioned manner. The distortion diagnostic feature table is used to obtain dynamic weight constraints through sensitivity mapping. The latency and contamination contribution are transformed into strategy level, exploration ratio and sample screening rules, and pushed to the strategy repository in the form of a strategy configuration file. The strategy configuration file and the set of high-confidence events are used as constraint inputs to drive the batch generation of copy templates, visual parameters and storyboards. The generated materials are written into the candidate material library after cross-modal consistency verification and trust score evaluation, along with strategy meta-information. When the candidate material library is triggered, a tracking identifier is attached and the materials are allocated to channels according to the strategy. The delivery request enters the user terminal with the identifier, forming a tracking-enabled return event stream and writing it into the time sequence storage. The tracked event flow undergoes alignment, completion, and re-inspection, updates delay and contamination labels, and writes back to the database. Together with the normalized event flow, it enters the next cycle, forming an adaptive closed-loop governance.

2. The method for generating multimodal viral content based on deep targeted penetration according to claim 1, characterized in that: We collect raw records from multiple sources, align them with timestamps, and standardize fields. We then map platform logs, user behavior receipts, channel feedback, and external anti-fraud detection data into a standardized event stream and write it into time-series storage, as detailed below: Raw records are batch-fetched from platform logs, user behavior receipts, channel feedback and third-party anti-fraud interfaces and formed into raw batch streams according to the fetch time; the raw batch streams are parsed into raw event lines one by one; fetched metadata and idempotency flags are added to the raw event lines, and the event lines with idempotency flags are submitted to the time field completion set; The time field completion set performs third-order time priority selection at the event row level, while recording all raw time fields and calculating the event arrival time difference, assigning an original time series to each event; The original time series is estimated and aligned within a sliding window; Alignment is performed to generate a unified time base field based on the system's unified base time. The unified time base event rows are standardized based on the registerable semantic mapping table; duplicate detection is performed on the standardized event rows; Add derived tags to the deduplicated, standardized event rows at the event level, and append the derived tags as structured fields to the event rows; Event rows with derived tags are serialized by partition key and written to time series storage in columnar compression format.

3. The method for generating multimodal viral content based on deep targeted penetration according to claim 2, characterized in that: End-to-end latency calculation and session fingerprint comparison are performed on the normalized event stream. Frequency mutation detection is then used to label events with latency attributes and contamination confidence, generating tagged event records, as detailed below: Read the normalized event stream from the time-series storage and complete the source context field for each event to form an extended event view; The extended event view estimates the clock offset of each source and corrects the timestamps by referencing heartbeat events and link receipts. The corrected timestamps are used to calculate the reporting delay, link transmission delay, and platform processing delay for each item, and the delay metric is attached to the extended event view as a delay attribute. The extended event view with delayed attributes is used to construct a session fingerprint vector by synthesizing device fingerprints and behavioral sequences. Multi-scale hashing and similarity retrieval are used to compare the fingerprints in the historical fingerprint database to generate log-level fingerprint similarity and session association chains. The fingerprint similarity and association chains are then written back to the event record. Events with delay and fingerprint information are recorded in a sliding window. Event density is calculated by fingerprint, IP, and channel and compared with an adaptive baseline to identify frequency mutations. After identifying the mutation window, the events in the window are labeled with mutation batch identifiers. Events with mutation batch identifiers are assigned individual anomaly scores based on the degree of delay anomaly, fingerprint repetition rate, mutation intensity, and behavioral consistency. These individual scores are then combined into contamination confidence labels using a hierarchical confidence fusion strategy based on historical channel reliability. Simultaneously, multi-factor anomaly evidence is written into the event label table as structured fields.

4. The method for generating multimodal viral content based on deep targeted penetration according to claim 3, characterized in that: Labeled events are recorded in short and long windows for batch aggregation and sliding statistics. Delay contribution and contamination percentage are extracted using causal discrimination to form a distortion diagnostic feature table, which is then versioned and stored, as follows: Read tagged event records from time-series storage, organize the events into a raw batch stream according to batch identifier and timestamp order, and write the raw batch stream into a temporary batch table; The original batch stream is sliced ​​into short-window and long-window segments from the temporary batch table, and the slicing results are used to form short-window batch segment and long-window batch segment tables. Calculate the statistics for each segment in each batch and write them into the statistical view table; The statistical view table was compared with the historical "clean baseline" in a stratified manner. The deviation was calibrated by quantile difference and sequence trend detection. The deviation was labeled as delayed abnormality and pollution exposure. The mutation window was written into the mutation index table with Boolean identifier. The mutation index table and the event snapshots of the corresponding batch fragments are used as causal evidence to perform a causal discrimination process; The differences in indicators before and after the intervention, along with the test results, were recorded as a preliminary evidence set for the contribution of delay and the proportion of pollution. The preliminary estimated evidence set is stratified and weighted according to event confidence, channel risk score and window consistency to form the final delay contribution and pollution ratio, along with confidence intervals and evidence citations. The final delay contribution, pollution percentage, related statistics, evidence citations, and operational recommendations are written into the immutable distortion diagnostic feature table as structured row records.

5. The method for generating multimodal viral content based on deep targeted penetration according to claim 4, characterized in that: The distortion diagnostic feature table is dynamically weighted through sensitivity mapping to obtain dynamic weight constraints. Latency and contamination contributions are transformed into policy levels, exploration ratios, and sample selection rules, and then pushed to the policy repository in the form of policy configuration files, as follows: The diagnostic records are read from the distortion diagnostic feature table and each record is weighted and benchmarked to form a weighted diagnostic row with business weight and confidence interval; The weighted diagnostic rows are aggregated according to the time window and channel dimension to obtain a dynamic weight constraint set; The dynamic weight constraint set is mapped to the weight interval as a strategy level fault, the corresponding exploration ratio range and sample usage rules according to the preset strategy mapping rules, thus forming the strategy level mapping rules. Serialize the policy level mapping rules into a structured key-value format policy configuration file.

6. The method for generating multimodal viral content based on deep targeted penetration according to claim 5, characterized in that: Using strategy configuration files and a set of high-confidence events as constraint inputs, the system drives the batch generation of copy templates, visual parameters, and storyboards. The generated materials are then written into the candidate material library along with strategy metadata after cross-modal consistency verification and trust scoring evaluation, as detailed below: Receive the policy configuration file and the set of high-confidence events and parse them into a structured set of constraints and a list of event fields; Template retrieval and priority ranking are driven by a set of structured constraints and a set of high-confidence events. Calculate priority scores for matching candidates based on similarity and historical performance, and write the scores and candidate IDs into the candidate template table; The selected template set is batch parameterized and variant expanded, and the parameter set and proportion strategy of each variant are written into the variant plan table. Trigger low-cost rendering and rapid compositing in parallel according to the variant plan, based on the text variants, visual parameter sets and storyboards; The text, images, and video samples in the temporary generation pool are subjected to cross-modal consistency verification, rule compliance checks, and historical similarity comparison; the semantic vector distance is used to obtain the similarity with the nearest historical best-selling products, and the verification evidence is written into the verification result table in a structured record along with the verification sub-item scores; The verification result table is combined with the generator confidence, event source confidence, and historical performance prediction to perform a confidence-weighted fusion calculation to calculate the trust score. The processing action is determined based on the trust score and policy constraints. The trust score, improvement suggestions, processing action, and traceable evidence package are written into the candidate material record. Candidate materials are persistently stored in the candidate material library as structured entries.

7. The method for generating multimodal viral content based on deep targeted penetration according to claim 6, characterized in that: When a campaign is triggered, the candidate creative library attaches a tracking identifier and allocates it to channels according to the strategy. The campaign request, carrying the identifier, enters the user's client, forming a tracking-enabled event stream and writing it to the time-series storage, as detailed below: Read the materials to be deployed and their complete metadata from the candidate material library, and generate a deployment task list according to the channel allocation rules defined in the strategy configuration file; A unique tracking identifier is generated for each campaign. The identifier combines the creative ID, channel ID, strategy version number, timestamp, and random hash to form an uncollision-free identifier, which is then written into the campaign list and creative metadata. The task list is sorted by channel, priority and delivery window, and then sent to each channel's delivery entry point through the scheduler interface. After receiving the delivery request, the channel renders the creative to the user's device and embeds tracking identifiers and strategy metadata to ensure that user interaction events are reported synchronously with tracking identifiers. When an interactive event occurs on the user end, the event data is bound to the tracking identifier and sent to the central backhaul bus according to the channel backhaul protocol. After receiving events, the central backhaul bus performs data cleaning and normalization processing to form a backhaul event stream with tracking.

Images