Method for generating multi-modal based on deep target penetration blockbuster material

Abstract

The application discloses a best-selling material multi-modal generation method based on deep target penetration, and particularly relates to the technical field of material multi-modal generation. Through unified mapping, alignment, completion and labeling processing of multi-source data, the accumulation risk of feedback distortion is reduced from the source. Through sensitivity mapping and strategy conversion of distortion diagnosis features, the influence of delay and pollution is directly constrained as grade allocation, exploration ratio and sample screening rules in material generation and delivery strategy, so that the generation link forms a dynamic balance between quality control and diversity exploration. At the same time, after the material is verified by cross-modal consistency and trust score, it is written into the candidate library and bound with strategy meta information, ensuring the content credibility and execution traceability of the generated results. In the delivery link, the design of additional tracking identifier and back event link is added.

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 they interact, they will produce an amplified 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 returned 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 indicators 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. And because 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 above background.

[0004] To achieve the above object, the present application provides the following technical scheme:

[0005] The best-selling material multi-modal generation method based on deep targeted penetration comprises the following steps:

[0006] 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;

[0007] Perform end-to-end latency calculation and session fingerprint comparison on the normalized event stream, combine frequency mutation detection to label events with delay attribute and pollution confidence, and generate labeled event records;

[0008] Batch aggregate and sliding statistics on labeled event records in short and long windows, extract delay contribution and pollution proportion combined with causal discrimination, form distortion diagnosis feature table and version storage;

[0009] 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;

[0010] 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, generate materials after cross-modal consistency verification and trust score evaluation, and write into candidate material library with strategy meta information;

[0011] 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;

[0012] 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 it is combined with the normalized event stream to form a self-adaptive closed loop management.

[0013] In a preferred embodiment, multiple source original records are collected and time stamp alignment and field standardization are performed, delivery platform logs, user behavior feedback, channel feedback and external anti-fraud detection are unified to form a normalized event stream and written into time sequence storage, as follows:

[0014] 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 the original batch stream into original event rows one by one; add crawling meta information and idempotent identifier to the original event rows, and submit the event rows with idempotent identifier to the time field completion set;

[0015] 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 the original time sequence to each event;

[0016] Estimate and align the clock bias of the original time sequence in the sliding window; adjust the unified time reference field based on the alignment with the system unified reference time;

[0017] The uniform time reference event line is standardized according to a registerable semantic mapping table; and the standardized event line is subjected to duplicate detection;

[0018] A derivative tag is added to the standardized event line after deduplication at an event level, and the derivative tag is attached to the event line as a structured field;

[0019] The event line with the derivative tag is serialized according to a partition key sequence and written into time series storage in a columnar compression format.

[0020] 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:

[0021] The normalized event stream is read from the time series storage, and source context fields are supplemented for each event to form an extended event view;

[0022] The extended event view is used to estimate the clock bias of each source by referring to heartbeat events and link acknowledgments, and to correct timestamps; the corrected timestamps are used to calculate and report 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;

[0023] The extended event view with delay attributes is used to construct a session fingerprint vector through synthetic device fingerprint and behavior sequence, and compared in a historical fingerprint library using multi-scale hashing and similarity retrieval to generate log-level fingerprint similarity and session association chain, and the fingerprint similarity and association chain are written back to the event record;

[0024] The event record with delay and fingerprint information is used to calculate event density by fingerprint, IP and channel within a sliding window, and compared with an adaptive baseline to identify frequency mutations; after identifying the mutation window, the events in the window are labeled with a mutation batch identifier;

[0025] The events with the mutation batch identifier are used to calculate single abnormality scores according to delay anomaly degree, fingerprint repetition rate, mutation intensity and behavior consistency, and to synthesize the single abnormality scores into a pollution confidence label using a hierarchical confidence fusion strategy according to historical channel reliability, and to write multi-factor abnormality evidence into the event label table as a structured field.

[0026] In a preferred embodiment, the labeled event records are batch aggregated and slidingly counted in short and long windows, combined with causal discrimination to extract delay contribution and pollution proportion, to form a distortion diagnosis feature table and versioned storage, as follows:

[0027] The labeled event records are 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;

[0028] Short window and long window slicing on the original batch flow from the temporary batch table, and forming short window batch segment and long window batch segment tables from the slicing results;

[0029] Calculating statistics for each batch segment and writing into the statistical view table;

[0030] Comparing the statistical view table with the historical "clean baseline" by stratification, detecting deviations by quantile difference and sequence trend, labeling the deviations as delay abnormality and pollution exposure, and writing the mutation window into the mutation index table as Boolean identification;

[0031] Using the mutation index table and the event snapshot of the corresponding batch segment as causal evidence to execute the causal discrimination process, and recording the index difference before and after the intervention and the test conclusion as the preliminary estimation evidence set of delay contribution and pollution proportion;

[0032] Fusing the preliminary estimation evidence set by event confidence, channel risk score, and window consistency to form the final delay contribution and pollution proportion with confidence intervals and evidence references;

[0033] Writing the final delay contribution, pollution proportion, related statistical items, evidence references, and operation suggestions into the immutable distortion diagnosis feature table as structured row records.

[0034] In a preferred embodiment, the distortion diagnosis feature table is mapped to obtain dynamic weight constraints, the delay and pollution contributions are converted into strategy levels, exploration ratios, and sample screening rules, and the strategy configuration file is pushed to the strategy warehouse, as follows:

[0035] Reading the diagnosis records from the distortion diagnosis feature table and normalizing the weights of each record to form weighted diagnosis rows with business weights and confidence intervals;

[0036] Aggregating the weighted diagnosis rows by time window and channel dimension to obtain a dynamic weight constraint set;

[0037] Mapping the weight intervals to strategy level boundaries, corresponding exploration ratio ranges, and sample usage rules according to the preset strategy mapping rules to form a strategy level mapping rule;

[0038] Serializing the strategy level mapping rule into a structured key-value format strategy configuration file.

[0039] In a preferred embodiment, the strategy configuration file and the high-confidence event set are used as constraint inputs to drive the batch generation of script templates, visual parameters, and shot scripts, and the generated materials are written into the candidate material library after cross-modal consistency verification and trust score evaluation with strategy meta-information, as follows:

[0040] Receiving a policy configuration file and a high-confidence event set and parsing into a structured constraint set and a list of event fields;

[0041] Driving template retrieval and priority ranking according to the structured constraint set and the high-confidence event set; calculating a priority score for each matching candidate according to similarity and historical performance prediction, and writing the score and candidate ID into a candidate template table;

[0042] The selected template set is batch parameterized and variant expanded, and the parameter set and proportion strategy of each variant are written into a variant plan table;

[0043] Triggering low-cost rendering and rapid synthesis of copy variants, visual parameter sets and shot scripts according to the variant plan in parallel;

[0044] Performing cross-modal consistency verification, rule compliance check and historical similarity comparison on the temporary generation pool of copy, picture and video samples; and obtaining the similarity of the nearest neighbor history blockbuster using semantic vector distance, and writing the verification evidence and sub-item scores into a verification result table;

[0045] The verification result table is fused with generator confidence, event source confidence and historical performance prediction to calculate a trust score, and the processing action is determined according to the trust score and policy constraints; the trust score, improvement suggestion, processing action and traceable evidence package are written into a candidate material record;

[0046] The candidate material record is stored in the candidate material library in the form of a structured entry.

[0047] In a preferred embodiment, the candidate material library is attached with a tracking identifier when triggered for delivery, and is distributed according to the strategy. The delivery request with the identifier enters the user end, forms a tracking return event stream and is written into a time sequence storage, as follows:

[0048] Reading the to-be-delivered material and its complete meta information from the candidate material library, and generating a delivery task list according to the channel distribution rules defined in the strategy configuration file;

[0049] Generating a unique tracking identifier for each delivery task, which is composed of material ID, channel ID, strategy version number, timestamp and random hash to form a non-collision identifier, and writing it into the task list and material meta information;

[0050] Sorting the delivery task list according to channel, priority and delivery window, and delivering it to each channel delivery portal through the scheduler interface;

[0051] The channel receives the delivery request, renders the material to the user end, and embeds the tracking identifier and policy meta information, to ensure that the user end interaction event carries the tracking identifier and is synchronously reported to the central back transmission bus;

[0052] When the user end generates an interaction event, the event data is bound with the tracking identifier and is sent to the central back transmission bus according to the channel back transmission protocol;

[0053] The central back transmission bus receives the event, performs data cleaning and standardization processing, and forms a tracking back transmission event stream.

[0054] The technical effects and advantages of the present application are as follows:

[0055] 1. The present application introduces the dynamic identification, quantitative diagnosis and policy constraint of the delay attribute and pollution confidence in the whole process of material generation and delivery, so that the closed-loop feedback link has the self-adaptive resistance ability to distortion effect, thereby realizing the whole controllable and iterative enhancement of the generation and optimization of hit material, ensuring the consistency of input data in timeliness and reliability through unified mapping, alignment, completion and labeling processing of multi-source data, reducing the accumulation risk of feedback distortion from the source; on the other hand, through the sensitivity mapping and policy transformation of distortion diagnosis features, the influence of delay and pollution is directly constrained as the grade allocation, exploration ratio and sample screening rules in the material generation and delivery strategy, so that the generation link forms a dynamic balance between quality control and diversity exploration, at the same time, the material is written into the candidate library after cross-modal consistency verification and trust score binding, and the policy meta information is guaranteed, ensuring the content credibility and execution traceability of the generation results, in the delivery link, the design of additional tracking identifier and back transmission event link makes the real performance of each material can be completely mapped into the feedback cycle, and forms the adaptive closed-loop management under the iterative update of delay and pollution labels.

[0056] 2. The present application not only significantly improves the precision of material targeting and the controllability of hit generation, but also effectively avoids the demand vector drift and budget waste caused by feedback distortion in traditional methods, realizes the stable optimization and continuous iteration in high dynamic environment, and has obvious technical innovation and practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings;

[0058] Figure 1 The flowchart of the embodiment method of the present application. DETAILED DESCRIPTION

[0059] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0060] Embodiments: Figure 1 The present application provides a hit material multi-modal generation method based on deep targeted penetration, including the following steps:

[0061] Collect multi-source original records and perform timestamp alignment and field standardization. Map the delivery platform log, user behavior reply, channel feedback and external anti-fraud detection into a standardized event stream and write into time series storage.

[0062] Perform end-to-end latency calculation and session fingerprint comparison on the standardized event stream. Combine frequency mutation detection to label events with delay attribute and pollution confidence, and generate labeled event records.

[0063] Batch aggregate and sliding statistics of labeled event records in short and long windows. Combine causal discrimination to extract delay contribution and pollution proportion, form distortion diagnosis feature table and version storage.

[0064] Map the distortion diagnosis feature table to get dynamic weight constraints. Convert delay and pollution contribution into strategy level, exploration ratio and sample screening rules, and push to strategy storage in the form of strategy configuration file.

[0065] Take the strategy configuration file and high-confidence event set as constraint input to drive batch generation of copy templates, visual parameters and shot scripts. After cross-modal consistency verification and trust score evaluation, the generated materials are written into the candidate material library with strategy meta information.

[0066] The candidate material library is attached with tracking identifiers and distributed to channels according to the strategy when triggered. The delivery request with identifier enters the user end, forms a tracking return event stream and writes into time series storage.

[0067] The tracking return event stream is aligned, completed and detected again. Update the delay and pollution labels and rewrite the database. Together with the standardized event stream, it enters the next cycle of adaptive closed-loop management.

[0068] Collect multi-source original records and perform timestamp alignment and field standardization. Map the delivery platform log, user behavior reply, channel feedback and external anti-fraud detection into a standardized event stream and write into time series storage, as follows:

[0069] Pull raw records from delivery platform logs, user behavior receipts, channel feedback, and third-party anti-fraud interface in batches and form raw batch streams according to the time of extraction;

[0070] Parse the raw batch stream into raw event rows one by one, including expanding nested structures, parsing encoding formats, extracting raw time fields and key fields (event generation time, reporting time, user ID, material ID, session ID, device fingerprint, etc.), and writing failed rows to an error queue for backtracking;

[0071] Add extraction meta-information and idempotent identifier to the raw event row, generate the idempotent identifier through (event ID / session ID / event content hash) combination, and use the idempotent identifier as the unique key for subsequent deduplication and retry processing, and submit the event row with the idempotent identifier to the time field completion set;

[0072] The time field completion set performs three-order time priority selection at the event row level: preferentially using event generation time, then using reporting time, and then using extraction time; record all original time fields and calculate the event arrival time difference (event reception time minus event generation time), and assign the original time sequence (generation time, reporting time, reception time) to each event;

[0073] Perform clock bias estimation and alignment on the original time sequence within a sliding window, and the clock bias estimation is achieved through cross-source verification (comparison of events across channels in the same session, heartbeat / benchmark event alignment) and sliding median deviation filtering; adjust the unified time reference field to the system unified benchmark time, and mark the "delay level" according to the configurable strategy for late events;

[0074] In an optional example, the clock bias estimation is achieved through cross-source verification (comparison of events across channels in the same session, heartbeat / benchmark event alignment) and sliding median deviation filtering, as follows:

[0075] Suppose there is an ad delivery session Session-12345, and the same user has generated click behavior in three different channels (APP, H5, and third-party alliance). Each channel reports the time of the click event:

[0076] APP reporting time: 2025-09-08 10:00:02;

[0077] H5 reporting time: 2025-09-08 10:00:06;

[0078] Third-party alliance reporting time: 2025-09-08 10:00:11;

[0079] At the same time, there is a heartbeat / reference event inside the system, recording the real reference time that should be triggered at that moment: 2025-09-08 10:00:00.

[0080] Step 1: Cross-source verification

[0081] The system first collects events from the same session (Session-12345) in three channels, forming an event group.

[0082] Compare the timestamps within the event group: APP (+2s), H5 (+6s), Alliance (+11s), which have differences relative to the reference time (10:00:00).

[0083] Through the heartbeat / reference event, the system confirms that the real reference point is at 10:00:00, so the "deviation" of each source is: APP deviation = +2 seconds, H5 deviation = +6 seconds, Alliance deviation = +11 seconds.

[0084] The system records these deviations, forming a deviation sequence: 2, 6, 11.

[0085] Step 2: Sliding median deviation filtering

[0086] The system maintains a sliding window (assuming window size = 5 sessions) to collect cross-source deviations of recent sessions.

[0087] The current window deviation set may be: 3, 2, 6, 11, 4 seconds.

[0088] Take the median in this window: after sorting, it is 2, 3, 4, 6, 11, median = 4 seconds.

[0089] The system takes 4 seconds as the corrected reference for the current time alignment.

[0090] For the three events of Session-12345:

[0091] APP original deviation +2 seconds → corrected difference = -2 seconds (close to the median, fine-tuned to around 10:00:00).

[0092] H5 original deviation +6 seconds → corrected difference = +2 seconds (close to the median, corrected to be closer to the reference).

[0093] Alliance original deviation +11 seconds → corrected difference = +7 seconds (exceeds the reasonable threshold, marked as a late event, while retaining the original value for tracing).

[0094] Final result

[0095] After cross-source validation and sliding median deviation filtering, the system generates the aligned reference timestamp: 10:00:00 ± 4 seconds.

[0096] The APP event is aligned to 10:00:02 (high confidence).

[0097] The H5 event is aligned to 10:00:04 (medium confidence).

[0098] The alliance event is marked as late and adjusted to 10:00:07 (low confidence, may trigger compensation mechanism).

[0099] The unified time reference event row is standardized according to the registrable semantic mapping table: field name mapping, enumeration value mapping, numerical unit normalization, and missing value strategy application (interpolation, marking, or retaining null values), while retaining the original field as a meta-field to meet traceability;

[0100] The standardized event row is subjected to duplicate detection and idempotent write preparation. Duplicate detection is determined by both idempotent identification and content similarity threshold: for duplicates, the merge strategy is executed (retain the most complete time sequence and the highest confidence value field), and idempotent write logic is implemented for concurrent retry scenarios.

[0101] In an optional example, the duplicate detection is determined by both idempotent identification and content similarity threshold, as follows:

[0102] Obtain two event records from different channels:

[0103] 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 = Jump Click.

[0104] 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.

[0105] Determination process:

[0106] Idempotent identification determination

[0107] The system first calculates the idempotent identification:

[0108] For Event A, the idempotent identification = Hash (evt_123 | sess_001 | mat_45 | u_789).

[0109] For event B, the idempotent identifier = hash(evt_123 | sess_001 | mat_45 | u_789).

[0110] The idempotent identifiers of the two events are exactly the same, so the system considers them as potentially duplicate events.

[0111] Content similarity threshold discrimination

[0112] The system further compares the field content:

[0113] Time difference = 2 seconds (within the allowed time jitter threshold of 5 seconds).

[0114] Click type field similarity (based on edit distance): jump click and jump similarity = 0.85, greater than the set threshold 0.8. In summary, the two records are highly consistent in time and semantics, so they are confirmed as duplicate events.

[0115] Merge strategy system

[0116] Leave one final event row:

[0117] The time field adopts the earliest generation time 2025-09-08 12:01:02.

[0118] The click type field adopts the semantic jump click.

[0119] The event ID and user information remain unchanged. The final merged result is written to the time series storage, and the other record is marked as "duplicate merge".

[0120] For the deduplicated standardized event row, add derived labels at the event level: channel inference (path backtracking), device type analysis (UserAgent analysis), geographic attribution estimation (IP lookup), and anti-fraud score merging (third-party score and internal rule weighting), and attach the derived labels to the event row as structured fields;

[0121] The event row with derived labels is serialized according to the partition key (date / hour, channel, batch ID) and written to the time series storage in columnar compression format, while the partition information, start and end offset, and event count are published to the downstream consumption topic through the message queue during the write transaction commit phase, so as to facilitate downstream real-time consumption and correlation query;

[0122] Write confirmation is written to the audit log and change log through atomic transaction, the audit log records the write time, partition, offset, idempotent identifier and processing result; the audit log triggers the monitor to evaluate the write delay and packet loss rate and sends an operation and maintenance notification through the alarm channel when the threshold is abnormal to facilitate manual intervention and backtracking investigation.

[0123] Perform end-to-end latency calculation and session fingerprint comparison on the normalized event stream, combined with frequency anomaly detection to label events with delay attribute and pollution confidence, and generate labeled event records as follows:

[0124] Read the normalized event stream from the time series store, and supplement each event with source context fields (event number, event occurrence time, platform backhaul time, batch identifier, device parameters, network information, session identifier, behavior type), form an extended event view and write it to a temporary table;

[0125] Correct the time stamps by estimating the clock bias of each source with reference to heartbeat events and link acknowledgement, and correct the time stamps. Calculate the reporting delay, link transmission delay and platform processing delay for each event, and add the delay metrics as delay attributes to the extended event view for indexed queries;

[0126] Construct session fingerprint vectors by synthesizing device fingerprints (device parameters + network features) and behavior sequences from the extended event view with delay attributes, and compare them in the historical fingerprint library using multi-scale hashing and similarity retrieval to produce log-level fingerprint similarity and session association chain, and write back the fingerprint similarity and association chain to the event record to support cross-event tracing;

[0127] Statistical event density by fingerprint, IP, and channel within a sliding window, and compare it with the adaptive baseline to identify frequency anomalies. After identifying the anomaly window, label the events in the window with the anomaly batch identifier and record the anomaly strength, anomaly duration and anomaly source;

[0128] Calculate single-item anomaly scores based on delay anomaly degree, fingerprint repetition rate, anomaly strength and behavior consistency, and use a hierarchical confidence fusion strategy based on historical channel trustworthiness to combine single-item scores into a pollution confidence label (pollution confidence is obtained by weighted sum of each single-item score), while writing multi-factor anomaly evidence into the event label table for auditing;

[0129] The single-item anomaly scores based on delay anomaly degree, fingerprint repetition rate, anomaly strength and behavior consistency are calculated as follows:

[0130] Extract delay features: For events with anomaly batch identifier, call their calculated reporting delay, link transmission delay and processing delay, respectively, and compare them with the historical mean of the same source channel, and weight the comparison results to obtain the delay deviation ratio as the delay anomaly measure;

[0131] Extracting fingerprint features: By performing clustering on the session fingerprint vectors of events within the same mutation batch, calculating the proportion of repetitions within the cluster, and comparing the similarity across batches, the fingerprint repetition rate is obtained as a measure of fingerprint abnormality;

[0132] Extracting mutation features: By differentiating the event frequency within the batch from the baseline distribution, and calculating the mutation amplitude, mutation duration, and mutation growth rate, the mutation amplitude, mutation duration, and mutation growth rate are weighted and summed as the mutation intensity, which is used as a measure of mutation abnormality;

[0133] Extracting behavior features: By performing sequence consistency detection on the behavior type sequence in the event, measuring the sequence difference from the typical user behavior pattern, combining the behavior time interval and the proportion of operation sequence anomalies, and weighting and summing to obtain the behavior consistency score, which is used as a measure of behavior abnormality;

[0134] Generating individual scores: Normalize the delay abnormality measure, fingerprint abnormality measure, mutation abnormality measure, and behavior abnormality measure respectively, and record them as delay individual score, fingerprint individual score, mutation individual score, and behavior individual score respectively, and write them into the abnormality score field of the event record.

[0135] According to the pollution confidence in the event label table, drive sample governance: For events with ultra-high confidence, append the event ID to the isolation list and issue isolation instructions to the training and delivery pipeline through the message queue. For events with medium confidence, write a low-weight identifier in the training pipeline and reduce the bid or limit exposure during delivery. For low-confidence events, maintain regular weight and channel eligibility;

[0136] Pack the isolation list and low-weight label into an auditable evidence package (including original trace, time delay sequence, fingerprint link, and mutation timing diagram) by batch, and the evidence package enters the manual review queue. The manual review conclusion is written back to the event label table and triggers an incremental training trigger to synchronize the verified cleaning samples to the offline retraining set to continuously calibrate the anomaly detector and pollution confidence.

[0137] Aggregate and slide statistics on labeled event records in short and long windows by batch, extract delay contribution and pollution proportion based on causal discrimination, form a distortion diagnosis feature table, and versionize storage, as follows:

[0138] Read labeled event records from time series storage, and organize events into original batch streams (fields: event ID, batch ID, occurrence time, reporting time, device fingerprint, channel, behavior type, pollution confidence, delay label) in order of batch ID and timestamp, and write the original batch stream into the temporary batch table.

[0139] 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.

[0140] 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.

[0141] 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.

[0142] 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).

[0143] The deviation is labeled as delayed anomaly degree and pollution exposure degree, as detailed below:

[0144] 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.

[0145] 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.

[0146] Causal inference procedure is performed by taking the mutation index table and the event snapshot of the corresponding batch segment as causal evidence: first, time sequence test is performed, then matching control (taking homogeneous control group) is performed, followed by intervention simulation (removing or re-drawing suspicious events in the event snapshot and recalculating key indicators), and finally, robustness test (lag test, placebo test) is performed; the difference in indicators before and after the intervention is recorded as the preliminary estimation evidence set of the delay contribution and the pollution proportion.

[0147] The preliminary estimation evidence set is layered confidence weighted fusion according to event confidence, channel risk score and window consistency: higher weight is given to high confidence sources, median aggregation is taken for consistent estimation of long and short windows, and low confidence is given to inconsistent estimation and marked for manual review, thereby forming the final delay contribution and pollution proportion with confidence interval and evidence reference.

[0148] The delay contribution refers to the proportion of the deviation of the indicator caused by the feedback link delay in the overall distortion effect. Its essence is: if the system runs under ideal no-delay conditions, what proportion of the difference between the reference value and the actual value of the indicator can be attributed to delay rather than other factors; by subtracting the no-delay control indicator value from the actual observed indicator value, the delay deviation is obtained, and by dividing the delay deviation by the actual observed indicator value minus the baseline indicator value, the delay contribution is obtained;

[0149] The pollution proportion refers to the proportion of the deviation caused by the pollution of the feedback signal by abnormal data (false interaction, brushing, non-target group) in the overall distortion effect. Its essence is: what proportion of the information increment in the event set is caused by pollution, causing the system to learn or place incorrectly; by calculating the indicator deviation caused by the pollution event, the pollution proportion is obtained by dividing the indicator deviation caused by all events;

[0150] The final delay contribution, pollution proportion, correlation statistics, evidence reference and operation suggestion are written into the immutable distortion diagnosis feature table (field examples: diagnosis ID, window ID, batch ID, short / long window statistics snapshot pointer, delay contribution, pollution proportion, confidence interval, evidence package pointer, suggestion action, version number, generation time, calculator version) in a structured row record, and the corresponding evidence package (event snapshot, intervention simulation result, robustness test log) is stored in the audit object storage, and finally the versioned writing of the diagnosis snapshot and the downstream notification action are recorded in the change log to complete the auditable link.

[0151] The distortion diagnosis feature table is mapped to obtain dynamic weight constraints, and the delay and pollution contribution is converted into strategy level, exploration ratio and sample screening rules, and is pushed to the strategy warehouse in the form of strategy configuration file, as follows:

[0152] Read the diagnosis record from the distortion diagnosis feature table and weight each record according to the business sensitivity table, activity priority and channel risk parameters to form a weighted diagnosis row with business weight and confidence interval.

[0153] Aggregate the weighted diagnosis row by time window and channel dimension, and cross-calibrate with the resource cost table, real-time capacity constraints and historical rectification efficiency table to obtain a dynamic weight constraint set considering business cost and executability.

[0154] Map the weight interval to the policy level fault (e.g. conservative, watchful, exploratory) according to the preset strategy mapping rule, and the corresponding exploration proportion range and sample usage rule (e.g. pool in / weight down / isolation) to form a strategy level mapping rule.

[0155] Serializes the strategy level mapping rule into a structured key-value format strategy configuration file, and embeds the diagnosis version number, evidence package pointer, generation timestamp, confidence interval and signature verification field in the configuration file to ensure traceable and auditable policy meta-information.

[0156] The strategy configuration file is checked for syntax and constraint consistency by the configuration mode verifier, and automatically verified in a sandbox environment through historical playback and small-scale simulation. The verification result is attached to the change record of the configuration file as a verification report and simulation index snapshot.

[0157] The configuration file and verification report are written to the policy repository and create version labels and immutable change logs in the repository. The repository triggers a message queue to notify the downstream distribution system for gray release testing when writing.

[0158] The test indicators generated by the gray release test are written back to the monitoring bus and differenced with the decomposition indicators of the distortion diagnosis feature table. The difference analysis determines whether to automatically roll back the configuration file in the repository to the latest stable snapshot or promote this version to the full policy, while writing the rollback or promotion event to the audit log to complete the closed-loop governance.

[0159] Take the strategy configuration file and high-confidence event set as constraint input to drive the batch generation of script templates, visual parameters and shot scripts. The generated materials are written to the candidate material library after cross-modal consistency verification and trust score evaluation, and are attached with policy meta-information, as follows:

[0160] Receive the strategy configuration file and high-confidence event set and parse it into a structured constraint set and event field list. According to the conservative level, exploration proportion, diversity upper limit, channel preference, sensitive word blacklist and attention time length, etc. Key-value constraints, and write the constraint set to a temporary constraint table.

[0161] According to the structured constraint set and the high-confidence event set, the template retrieval and priority ranking are driven: event keywords, sentiment tendency, product attributes, and channel preference are used as retrieval conditions for multi-scale matching in the copy template library, visual style library, and storyboard fragment library; priority scores are calculated for the matching candidates according to similarity and historical performance prediction, and the scores and candidate IDs are written into the candidate template table.

[0162] The priority scores of the matching candidates are calculated according to similarity and historical performance prediction, as follows:

[0163] The meta-information field of the candidate material (including generation source, keyword weight, sentiment intensity, visual element distribution, historical delivery label, etc.) is read, and a similarity retriever is called to perform multi-modal similarity comparison between each candidate and the historical high-quality material library, to obtain semantic similarity scores and visual structure similarity scores.

[0164] The semantic similarity scores and visual structure similarity scores are weighted and fused with the historical performance prediction indicators (CTR, CVR, interaction depth, retention time, etc. given by the predictor) of the candidate material to obtain priority scores;

[0165] The selected template set is subjected to batch parameterization filling and variant expansion: event field list is used to fill the copy placeholder, event visual preference is used to generate the color palette and composition weight, and attention time is used to splice the storyboard fragment; according to the exploration ratio, controlled combination expansion (exhaustive + sampling) is performed on each template, and the parameter set and proportion strategy of each variant are written into the variant plan table.

[0166] The copy variants, visual parameter sets, and storyboard scripts are triggered for low-cost rendering and rapid synthesis in parallel according to the variant plan: low-resolution preview text, multiple picture drafts, and short-time axis samples are generated first to save resources; during the rendering process, the generator version, random seed, time consumption, and resource indicators are recorded, and the rendering products are written into the temporary generation pool in an indexed manner.

[0167] The copy, picture, and video samples in the temporary generation pool are subjected to cross-modal consistency verification, rule compliance check, and historical similarity comparison: keyword-visual element mapping check (checking whether the copy keywords appear in the picture / video through OCR or object detection), timing rhythm and copy emphasis point alignment check, brand element and sensitive word check; the semantic vector distance is used to obtain the similarity of the nearest neighbor historical hit, and the verification evidence is recorded in a structured manner, and the verification sub-scores are written into the verification result table.

[0168] The check result table is combined with the generator confidence, event source confidence, and historical performance prediction to calculate a trust score by confidence-weighted fusion, and the processing action is determined according to the trust score and the policy constraint: 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, and processing action are written into the candidate material record together with the traceable evidence package (event ID, template ID, parameter set, generation log, and check record).

[0169] The candidate material record is stored in the candidate material library in a structured entry form, and the entry includes a thumbnail / preview link, complete generation parameters, a trust score, a policy version, an evidence package pointer, channel adaptation tags, and priority scores; and the entry is injected into a candidate priority queue according to priority, and the priority queue simultaneously maintains gray delivery qualifications, grouping suggestions, and rollback conditions for delivery scheduling to directly read and execute.

[0170] The candidate material library is attached with a tracking identifier when delivery is triggered and is distributed according to the policy, and the delivery request with the identifier enters the user end, forming a traceable return event stream and writing into the time sequence storage, as follows:

[0171] The candidate material library is attached with a tracking identifier when delivery is triggered and is distributed according to the policy, and the delivery request with the identifier enters the user end, forming a traceable return event stream and writing into the time sequence storage, as follows:

[0172] A unique tracking identifier (tracking ID) is generated for each delivery task, which is composed of the material ID, channel ID, policy version number, timestamp, and random hash to form a non-collision identifier, and is written into the task list and material meta-information to ensure traceability throughout the process from delivery to return.

[0173] The delivery task list is sorted according to the channel, priority, and delivery window, and is sent to the delivery entrance of each channel through the scheduler interface; during the delivery process, the task fields are checked for format, legality, and exposure time window conflict, and a delivery execution log is generated and written into the scheduling audit table for subsequent review.

[0174] After receiving the delivery request, the channel renders the material to the user end, while embedding the tracking identifier and policy meta-information (such as URL parameters, hidden IDs, or front-end event binding), ensuring that user interaction events (click, stay, slide, interaction, etc.) can carry the tracking identifier for synchronous reporting.

[0175] When the user terminal interacts with the event, the event data is bound with the tracking identifier, and the channel is returned to the central backhaul bus according to the protocol, and the environment information (network status, device type, geographic location, event timestamp) is attached, ensuring that the event can be mapped to the corresponding material and strategy version.

[0176] After the central backhaul bus receives the event, it performs data cleaning and standardization processing: verifies the legality of the tracking identifier, fills in the missing fields, corrects the time delay and time zone, and merges the repeated events to form a complete and traceable backhaul event stream, and writes it into the time sequence storage for real-time analysis and historical archiving.

[0177] The backhaul event stream generates an index table (according to tracking ID, material ID, strategy version, channel, time window) during the writing process in the time sequence storage, and establishes an event link mapping, ensuring that each backhaul event can be directly traced back to the candidate material, delivery strategy and evidence package, providing a reliable data foundation for subsequent closed-loop distortion analysis and strategy optimization.

[0178] The backhaul event stream with tracking 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 jointly entered into the next round of circulation, forming an adaptive closed-loop management, as follows:

[0179] Read the backhaul event stream with tracking from the time sequence storage, and sort it according to the event ID, tracking identifier and channel dimension, match the event with the original candidate material record, strategy version and user portrait, and form an event-material mapping table.

[0180] Perform alignment operation on the event-material mapping table: correct the timestamp, unify the time zone, correct the network transmission delay, use heartbeat and link feedback as the delay benchmark, generate the corrected event stream, to ensure that the backhaul event corresponds to the delivery material and strategy version.

[0181] Perform data completion on the corrected event stream: fill in the missing fields (such as event type, dwell time, interaction status, device information) according to the historical event sequence and user behavior pattern, and label the completion source and confidence, to ensure data integrity for subsequent analysis.

[0182] Detect the completed event stream again: based on delay distribution, fingerprint repetition rate, frequency mutation and behavior consistency, recalculate the delay label and pollution label, and combine the historical cleaning baseline to do layered confidence fusion, generate updated delay contribution and pollution proportion, and add traceable evidence field.

[0183] Write the updated delay and pollution labels back to the event database, and merge them with the original standardized event stream to form a unified event data set with real-time feedback labels, providing standardized input for the next round of generation, strategy adjustment and candidate material selection.

[0184] In the write-back process, versioned snapshots and change logs are established, recording event ID, tracking identifier, pre-update and post-update labels, confidence interval, evidence package pointer and processing time, ensuring that the evolution of each event is traceable and auditable.

[0185] The merged event dataset automatically triggers the next round of closed-loop generation and strategy optimization: the generated new material, updated strategy configuration and high-confidence event set will again drive copywriting, visual and storyboard generation, and the return event continues to enter the alignment, completion and detection cycle again, achieving adaptive closed-loop management, so that material generation and delivery strategy dynamically match real user behavior and channel feedback.

[0186] The present application introduces the dynamic identification, quantitative diagnosis and strategic constraint of delay attribute and pollution confidence in the whole process of material generation and delivery, so that the closed-loop feedback link has the ability to adaptively resist distortion effects, thereby realizing the whole-process controllability and iterative enhancement of blockbuster material generation and optimization. Through unified mapping, alignment, completion and labeling of multi-source data, the consistency of input data in timeliness and reliability is ensured, and the accumulation risk of feedback distortion is reduced from the source. On the other hand, through the sensitivity mapping and strategic transformation of distortion diagnosis features, the impact of delay and pollution is directly constrained as grade allocation, exploration ratio and sample screening rules in material generation and delivery strategy, so that the generation link forms a dynamic balance between quality control and diversity exploration. At the same time, the material is written into the candidate library after cross-modal consistency verification and trust scoring and is bound to the strategy meta-information, ensuring the content credibility and execution traceability of the generated results. In the delivery link, the design of additional tracking identifier and return event link enables the real performance of each material to be completely mapped into the feedback loop, and forms adaptive closed-loop management under the iterative update of delay and pollution labels.

[0187] The present application not only significantly improves the precision of material targeting and the controllability of blockbuster generation, but also effectively avoids the demand vector drift and budget waste caused by feedback distortion in traditional methods, realizes stable optimization and continuous iteration in a high dynamic environment, and has obvious technical innovation and practical application value.

[0188] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0189] It should be understood that the size of the sequence number of each process in the various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0190] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for generating a multi-modal based on a hit material with deep target penetration, characterized in that: Comprise the following steps: Collecting multi-source original records and performing timestamp alignment and field standardization, mapping the delivery platform log, user behavior feedback, channel feedback and external anti-fraud detection into a standardized event stream and writing into time series storage; Performing end-to-end delay calculation and session fingerprint comparison on the standardized event stream, combining frequency mutation detection to label delay attribute and pollution confidence for events, and generating labeled event records; Batch aggregation and sliding statistics are performed on the labeled event records in short and long windows, combined with causal discrimination to extract delay contribution and pollution proportion, forming a distortion diagnosis feature table and version storage; The distortion diagnosis feature table is mapped to get dynamic weight constraints, and the delay and pollution contribution is converted into strategy level, exploration ratio and sample screening rules, and pushed to the strategy warehouse in the form of strategy configuration file; The strategy configuration file and the high-confidence event set are used as constraint inputs to drive the batch generation of script templates, visual parameters and shot scripts, and the generated materials are written into the candidate material library after cross-modal consistency verification and trust score evaluation, with strategy meta information; The candidate material library is attached with tracking identifiers and distributed by channel according to the strategy when triggered, and the delivery request with identifier enters the user end, forming a tracking return event stream and writing into time series storage; The tracking return event stream is aligned, completed and detected again, updating the delay and pollution labels and rewriting the database, and entering the next cycle with the standardized event stream to form a self-adaptive closed-loop management.

2. The method of claim 1, wherein: Collecting multi-source original records and performing timestamp alignment and field standardization, mapping the delivery platform log, user behavior feedback, channel feedback and external anti-fraud detection into a standardized event stream and writing into time series storage, as follows: Batch pulling original records from delivery platform log, user behavior feedback, channel feedback and third-party anti-fraud interface and forming original batch stream according to grabbing time; parsing each original batch stream into an original event row; adding grabbing meta information and idempotent identifier to the original event row, and submitting 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; Performing clock bias estimation and alignment on the original time sequence in a sliding window; Adjusting the aligned event row to generate a unified time reference field; Standardizing the unified time reference event row according to the registrable semantic mapping table; and performing duplicate detection on the standardized event row; Adding a derived label to the event row at the event level, and attaching the derived label to the event row as a structured field; Serializing the event row with derived label according to the partition key and writing it into time series storage in columnar compression format.

3. The method of claim 2, wherein: Performing end-to-end delay calculation and session fingerprint comparison on the standardized event stream, combining frequency mutation detection to label delay attribute and pollution confidence for events, and generating labeled event records, as follows: Read the standardized 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 is expanded to refer to the heartbeat event, estimate the clock bias of each source, and correct the timestamp. The corrected timestamp is used to calculate and report the delay, link transmission delay, and platform processing delay. The delay metric is added to the extended event view as a delay attribute; The extended event view with delay attribute is used to construct a session fingerprint vector by combining device fingerprints and behavior sequences. Multi-scale hashing and similarity retrieval are used to compare the vector with historical fingerprints in a library to generate log-level fingerprint similarity and session association chain. The fingerprint similarity and association chain are written back to the event record; The event record with delay and fingerprint information is used to calculate the event density by fingerprint, IP, and channel in a sliding window. The density is compared with an adaptive baseline to identify frequency mutations. After identifying the mutation window, the events in the window are labeled with a mutation batch identifier; The event record with mutation batch identifier is used to calculate single abnormal score based on delay anomaly degree, fingerprint repetition rate, mutation intensity, and behavior consistency. The single score is combined into a pollution confidence label using a hierarchical confidence fusion strategy based on historical channel reliability. The multi-factor abnormal evidence is written into the event label table as a structured field.

4. The method of claim 3, wherein: The labeled event record is aggregated and slidingly calculated in short and long windows. The delay contribution and pollution proportion are extracted based on causal discrimination to form a distortion diagnosis feature table and versioned storage. The specific steps are as follows: The labeled event record is read from the time series storage and organized into an original batch stream according to the batch identifier and timestamp order. 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. The slice results form short and long window batch segment tables; The statistical quantities of each batch segment are calculated and written into a statistical view table; The statistical view table is compared with the historical "clean baseline" in a hierarchical manner. The deviation is detected and labeled as delay anomaly degree and pollution exposure degree. The mutation window is written into a mutation index table as a Boolean identifier; The mutation index table and the event snapshot of the corresponding batch segment are used as causal evidence to execute a causal discrimination process; The index difference and test conclusion before and after the intervention are recorded as the preliminary estimation evidence set of delay contribution and pollution proportion; The preliminary estimation evidence set is weighted and fused based on event confidence, channel risk score, and window consistency to form the final delay contribution and pollution proportion with confidence interval and evidence reference; The final delay contribution, pollution proportion, associated statistical items, evidence reference, and operation suggestion are written into an immutable distortion diagnosis feature table as a structured row record.

5. The method of claim 4, wherein: The distortion diagnosis feature table is mapped to obtain dynamic weight constraints. The delay and pollution contribution is converted into strategy level, exploration ratio, and sample screening rule. The strategy configuration file is pushed to the strategy warehouse. The specific steps are as follows: The diagnosis record is read from the distortion diagnosis feature table and weighted based on the baseline. A weighted diagnosis row with business weight and confidence interval is formed; The weighted diagnosis row is aggregated according to the time window and channel dimension to obtain a dynamic weight constraint set; The dynamic weight constraint set is mapped to a policy level fault according to a preset strategy mapping rule, a corresponding exploration ratio range and a sample use rule, and a policy level mapping rule is formed; The policy level mapping rule is serialized into a structured key-value format policy configuration file.

6. The method of claim 5, wherein: The policy configuration file and the high-confidence event set are used as constraint inputs to drive the batch generation of script templates, visual parameters and shot scripts. The generated materials are written into the candidate material library after cross-modal consistency verification and trust score evaluation, and are attached with policy meta information. The specific process is as follows: Receive the policy configuration file and the high-confidence event set and parse them into a structured constraint set and an event field list; According to the structured constraint set and the high-confidence event set, drive template retrieval and priority sorting; Calculate the priority score of the matched candidates according to similarity and historical performance prediction, and write the score and candidate ID 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; The script variants, visual parameter sets and shot scripts are triggered in parallel according to the variant plan for low-cost rendering and rapid synthesis; Perform cross-modal consistency verification, rule compliance check and historical similarity comparison on the script, picture and video samples in the temporary generation pool; and use semantic vector distance to obtain the similarity of the nearest neighbor historical blockbuster, and record the verification evidence in a structured manner together with the verification sub-score into the verification result table; Calculate the trust score by weighting and fusing the generator confidence, event source confidence and historical performance prediction, and determine the processing action according to the trust score and policy constraint; write the trust score, improvement suggestion, processing action and traceable evidence package into the candidate material record; The candidate material record is stored in the candidate material library in a structured entry form.

7. The method of claim 6, wherein: When the candidate material library is triggered for delivery, a tracking identifier is added and the channels are distributed according to the strategy. The delivery request with the identifier enters the user end, forming a tracking back event stream and writing it into the time sequence storage. The specific process is as follows: Read the to-be-delivered material and its complete meta information from the candidate material library, and generate a delivery task list according to the channel distribution rules defined in the strategy configuration file; Generate a unique tracking identifier for each delivery task. The identifier is composed of material ID, channel ID, strategy version number, timestamp and random hash to form a non-collision identifier, and is written into the task list and material meta information; Sort the delivery task list according to the channel, priority and delivery window in batches, and send it to the channel delivery entrance through the scheduler interface; After receiving the delivery request, the channel renders the material to the user end, and embeds the tracking identifier and policy meta information to ensure that the user end interaction event carries the tracking identifier and is reported synchronously; When the user end generates an interaction event, bind the event data and the tracking identifier, and send it to the central back bus according to the channel back protocol; The central back bus receives the event and performs data cleaning and standardization processing to form a tracking back event stream.

Images