Big data advertisement tag classification system based on ai analysis

Abstract

The application discloses a big data advertisement label classification system based on AI analysis, and particularly relates to the field of advertisement data processing and intelligent label classification, and is used for solving the interest recognition distortion and the delivery label drift problem caused by the coarse granularity of the end side theme abstract, is that the end side coarse grain theme is connected into a coherent behavior context, and then the curvature amplitude index and the path stability potential double parameters are used for judging and merging the credibility, then the pseudo-continuous track is remodeled, and the fine-grained interest label is derived under the mixed attention modulation; the overall architecture breaks through the information fault after the abstract compression, so that the audience portrait closely follows the real-time interest and does not drift with the data reduction, the delivery end accurately bids according to the audience portrait, the distorted label is no longer repeatedly corrected, the budget waste is greatly reduced, and the exposure frequency distribution tends to be balanced; the progressive resonance graph and the label tree mapping library are kept synchronous, the real-time bidding engine is always anchored to the latest interest coordinate, the advertisement experience is close to the user demand, and the brand safety risk is inhibited with the decrease of the mislabeling.

Description

Technical Field

[0001] This invention relates to the field of advertising data processing and intelligent tag classification, and more specifically, to a big data advertising tag classification system based on AI analysis. Background Technology

[0002] Mobile advertising is undergoing a privacy sandbox transformation. The mobile operating environment encapsulates user browsing history into on-device topic summaries, exposing only a coarse-grained overview of interests to the outside world. This encapsulation is completed within the device, and external data pipelines no longer access individual behavior logs. High-dimensional features such as audience interests, material semantics, and page context are compressed into limited summaries, and native links between cross-application touchpoints are consequently interrupted, resulting in a sharp reduction in the dimensionality of streaming features.

[0003] In this environment, the AI-driven ad tag classification architecture reveals non-obvious shortcomings. As training corpora lose their contextual details, the classification decision area shifts over time, resulting in tag outputs appearing to float across the same audience. In the online clustering stage, the lack of cross-session alignment cues leads to the fragmentation of the same interest into multiple branches, causing the tag tree to expand rapidly and its branches to become mutually exclusive. The ad delivery engine then participates in bidding based on distorted interest profiles; when users' true interests have shifted, the budget repeatedly cycles through ineffective exposures, amplifying resource waste as the bidding chain continues. The problem stems from the granular impact triggered by privacy encapsulation, manifesting in feature compression and broken correlation links, with consequences concentrated in the decline of tag accuracy and ad delivery imbalance.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, embodiments of the present invention provide a big data advertising tag classification system based on AI analysis. This system connects coarse-grained topics from the client-side into a coherent behavioral framework, then uses a dual-parameter decision of curvature amplitude exponent and path stability potential to determine the credibility of the merged system, thereby reshaping pseudo-continuous trajectories and deriving fine-grained interest tags under hybrid attention modulation. The overall architecture breaks down information gaps after summary compression, ensuring that audience profiles closely align with real-time interests without drifting due to data reduction. This allows for precise bidding on the ad delivery end, eliminating repeated corrections based on distorted tags, significantly reducing budget waste, and achieving a more balanced distribution of exposure frequency. The progressive resonance graph and tag tree mapping library remain synchronized, ensuring the real-time bidding engine is always anchored to the latest interest coordinates. This results in an advertising experience that meets user needs, and brand security risks are mitigated as mislabeling decreases, thus addressing the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Topic documentation module: The topic summary and the context identifier of the delivery material on the receiving end are written into the topic temporary storage table in sequence and a window timestamp is attached;

[0008] The network weaving module reads the topic temporary table, uses the cross-window semantic resonance retrieval algorithm to filter out continuous topic pairs, generates a progressive resonance graph and stores it in the resonance buffer;

[0009] Trajectory reconstruction module: First, calculate the semantic drift curvature for each topic path within the resonance buffer, extract key features to evaluate the trajectory merging credibility for each topic path, merge nodes to construct pseudo-continuous trajectories when the criteria are met, and label candidate attribution labels at the same time;

[0010] The tag inference module feeds pseudo-continuous trajectories into the tag inference engine, uses a hybrid attention parsing algorithm to generate fine interest tags, and then writes them into the tag tree mapping library.

[0011] Mapping distribution module: Synchronizes the latest entries in the tag tree mapping library to the bidding decision engine by index, along with a precision tag.

[0012] In a preferred embodiment, the subject documentation module processes the following:

[0013] The device receives a topic summary and a delivery material context identifier from the receiving end. A window timestamp is generated for each received data record. The window timestamp is based on the current time when the data is received and aligned to a preset time window length. The device-side topic summary, delivery material context identifier, and window timestamp are written as a record into a topic temporary storage table. The topic temporary storage table stores records in the order of receipt to preserve the temporal characteristics of the data. The topic temporary storage table is maintained by cleaning up records whose window timestamps are earlier than the current time minus a preset time period to ensure data freshness and storage efficiency.

[0014] In a preferred embodiment, the processing content of the vein weaving module is as follows:

[0015] Retrieves records containing edge-side topic summaries, delivery material context identifiers, and window timestamps from the topic temporary storage table; converts edge-side topic summaries into vector representations using a pre-trained semantic embedding model; calculates the cosine similarity between edge-side topic summary vectors in adjacent time windows; filters out consecutive topic pairs based on a preset similarity threshold; constructs a progressive resonance graph, where nodes are edge-side topic summaries, directed edges represent consecutive topic pairs, and edge attributes include cosine similarity and time difference; stores the progressive resonance graph in the resonance buffer in the form of an adjacency list.

[0016] In a preferred embodiment, the trajectory reconstruction module processes the following:

[0017] The topic paths in the progressive resonance graph are obtained from the resonance buffer. Semantic drift curvature is calculated for each topic path, determined by measuring the change in the direction of the embedding vector between adjacent nodes. Curvature amplitude exponent and path stability potential are extracted as key features. The curvature amplitude exponent measures the fluctuation amplitude of semantic drift by calculating the interquartile range of the open angle, while the path stability potential comprehensively evaluates the topic similarity and the stability of time jumps. A Bayesian sparse regressor is used to evaluate the credibility of trajectory merging based on the curvature amplitude exponent and path stability potential, and a merging credibility coefficient is output. When the merging credibility coefficient is higher than a preset threshold, pseudo-continuous trajectories are constructed from the merged path nodes, and candidate attribution labels are generated for the pseudo-continuous trajectories.

[0018] In a preferred embodiment, the trajectory reconstruction module further includes the following processing:

[0019] The curvature amplitude index quantifies the fluctuation of semantic drift by calculating the interquartile range of the open angles. The specific steps are as follows: First, collect the open angles between all adjacent nodes on the topic path. These open angles represent the degree of change in semantic direction. Then, sort these open angles by size and find the 25th quantile Q1 and the 75th quantile Q3. Finally, the curvature amplitude index is defined as the difference between Q3 and Q1, that is, the interquartile range IQR = Q3 - Q1.

[0020] In a preferred embodiment, the trajectory reconstruction module further includes the following processing:

[0021] The path stability potential measures the overall stability of a path by comprehensively evaluating the stability of topic similarity and time jumps. Its calculation process includes: First, using a sliding window technique, calculating local topic similarity segment by segment along the topic path, and then taking the mean of all similarities within the sliding window to reflect the semantic consistency of the path; Second, analyzing the stability of each time jump distance in the path, i.e., the time interval between nodes, by calculating the reciprocal of the variance of the time jump distance—the smaller the variance, the more balanced and stable the time jumps; Finally, the path stability potential is defined as the product of the mean topic similarity and the stability of time jumps.

[0022] In a preferred embodiment, the tag optimization module processes the following:

[0023] The pseudo-continuous trajectory, including the edge topic summary and its embedding vector, is extracted from the resonant buffer and input into the tag inference engine. The pseudo-continuous trajectory is processed using a hybrid attention parsing algorithm. The semantic dependency between edge topic summaries is calculated through a self-attention mechanism to generate a context vector. The context vector is associated with the embedding vector of the advertising material context identifier through a cross-attention mechanism to generate a hybrid representation vector. The hybrid representation vector is input into a multilayer perceptron to generate a probability vector. The component with the highest probability is selected as the fine interest tag. The fine interest tag is associated with the unique identifier of the pseudo-continuous trajectory and stored in the tag tree mapping library in the form of key-value pairs.

[0024] In a preferred embodiment, the tag optimization module further includes the following processing steps:

[0025] The hybrid attention mechanism is divided into two parts: self-attention and cross-attention, which are used to handle the dependencies within pseudo-continuous trajectories and the association between user interests and advertising content, respectively.

[0026] In a preferred embodiment, the processing content of the mapping and distribution module is as follows:

[0027] The latest fine-grained interest tags and their corresponding pseudo-continuous trajectory unique identifiers are extracted from the tag tree mapping library. The fine-grained interest tags are then transferred to the data cache of the bidding decision engine using the pseudo-continuous trajectory unique identifiers as indexes. Based on the tag probability vector generated by the tag inference engine, the maximum component value is selected as the precision marker, and the precision marker and the fine-grained interest tags are appended together to the data associated with the pseudo-continuous trajectory unique identifiers. Using the pseudo-continuous trajectory unique identifiers as keys, the fine-grained interest tags and precision markers are combined as values ​​to form key-value pairs stored in the data structure of the bidding decision engine.

[0028] The technical effects and advantages of the AI-based big data advertising tag classification system of this invention are as follows:

[0029] This invention uses temporal resonance retrieval to connect coarse-grained topics on the edge into a coherent behavioral framework. It then uses a dual-parameter decision system of curvature amplitude exponent and path stability potential to determine the credibility of the merged data, thereby reshaping pseudo-continuous trajectories and deriving fine-grained interest tags under hybrid attention modulation. The overall architecture breaks down information gaps after summary compression, ensuring that audience profiles closely align with real-time interests without drifting due to data reduction. This allows for precise bidding on the ad delivery side, eliminating repeated adjustments based on distorted tags, significantly reducing budget waste, and achieving a more balanced distribution of exposure frequency. The progressive resonance graph and tag tree mapping library remain synchronized, ensuring the real-time bidding engine is always anchored to the latest interest coordinates. This results in an ad experience that meets user needs, and brand security risks are mitigated as mis-targeting decreases. Operations do not require frequent manual adjustments; the algorithm self-consistently learns and converges, achieving dual benefits in privacy compliance and delivery efficiency. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the structure of the AI-based big data advertising tag classification system of the present invention.

[0031] Figure 2 This is a flowchart illustrating the tag refinement module of the AI-based big data advertising tag classification system of the present invention. Detailed Implementation

[0032] 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.

[0033] Example 1: Figure 1 This invention presents a big data advertising tag classification system based on AI analysis, comprising:

[0034] Topic Filing Module: The topic summary and the context identifier of the delivery material on the receiving end are written into the topic temporary storage table in sequence and a window timestamp is attached.

[0035] The network weaving module reads the topic temporary table, uses the cross-window semantic resonance retrieval algorithm to filter out continuous topic pairs, generates a progressive resonance graph, and stores it in the resonance buffer.

[0036] Trajectory Reconstruction Module: Within the resonance buffer, the semantic drift curvature is first calculated for each topic path, key features are extracted to evaluate the trajectory merging credibility for each topic path, and when the criteria are met, nodes are merged to construct pseudo-continuous trajectories, while candidate attribution labels are marked.

[0037] Tag refinement module: The pseudo-continuous trajectory is fed into the tag inference engine, and a fine interest tag is generated with the help of a hybrid attention parsing algorithm, which is then written into the tag tree mapping library.

[0038] Mapping distribution module: Synchronizes the latest entries in the tag tree mapping library to the bidding decision engine by index, along with a precision tag.

[0039] In mobile advertising scenarios, the privacy sandbox transformation has fundamentally changed how mobile operating environments handle user behavior data. User browsing trajectories are encapsulated into on-device topic summaries, exposing only a coarse-grained overview of interests, such as "sports" or "technology." This encapsulation is done internally, preventing external data pipelines from acquiring individual behavior logs. High-dimensional features such as audience interests, material semantics, and page context are compressed into limited summaries, breaking native links between cross-application touchpoints and significantly reducing feature dimensions. This granularity impact leads to a decline in tag accuracy and delivery imbalance in ad tag classification architectures that rely on AI inference. To address this challenge, this invention proposes a big data ad tag classification system based on AI analysis. By receiving and processing on-device topic summaries and contextual identifiers of delivery materials, it lays the foundation for generating fine-grained interest tags. The topic archiving module, as the starting point of the entire process, is responsible for receiving and initially organizing this coarse-grained data, ensuring that subsequent steps can perform semantic analysis and trajectory reconstruction based on ordered time-series information.

[0040] The goal of the topic archiving module is to send the topic summary and the context identifier of the delivery materials to the receiving end in sequence to the topic temporary storage table, and to attach a window timestamp to each record. The detailed technical logic is as follows:

[0041] S1.1, Receiver-side topic summary and context identifier for delivered materials;

[0042] In mobile advertising scenarios, due to increased privacy concerns, mobile devices no longer directly provide detailed user interest data. Instead, they transmit a coarse-grained overview of user interests and contextual information about ad creatives to the system via predefined application programming interfaces (APIs). This design effectively balances privacy protection and ad delivery requirements. The API receives a client-side topic summary and ad creative context identifier from the mobile device. The client-side topic summary is a brief description of the user's current interests; for example, when a user browses sports content, the device generates "sports" as the client-side topic summary. The ad creative context identifier is a unique identifier associated with the ad creative; for example, "Ad A" represents a specific ad content. The data record received by the system may be a combination of "sports" and "Ad A." This reception method ensures the real-time nature and accuracy of the data, providing the necessary information source for subsequent classification processing.

[0043] S1.2, Generate window timestamp;

[0044] To preserve the temporal sequence of data during processing, a timestamp needs to be attached to each received data record to track the temporal patterns of data generation. A window timestamp is generated for each data record, based on the current time at which the data is received and adjusted according to a preset time window length. This time alignment ensures that data records within the same window have the same window timestamp, facilitating the system's identification of temporal correlations in subsequent analysis.

[0045] S1.3, Write to the topic temporary table;

[0046] The received client-side topic summary, delivery creative context identifier, and generated window timestamp need to be integrated and stored so that the system can manage this data in an orderly manner. The client-side topic summary, delivery creative context identifier, and window timestamp are written as a complete record to the topic temporary storage table. The topic temporary storage table is a storage structure used to temporarily store the most recent data. Each record contains three fields; for example, a record might be "Sports|Advertising A|2025-06-27 10:58:00". During writing, records are added sequentially according to the order of receipt to ensure that the temporal characteristics of the data are preserved.

[0047] S1.4, Maintain the topic temporary table;

[0048] As data is continuously received, the number of records in the topic staging table gradually increases. To prevent storage space from being consumed indefinitely, the system needs to periodically update the table's contents to maintain its efficient operation. This is achieved by cleaning up expired records, for example, by setting it to retain only records from the most recent 24 hours. Specifically, the window timestamp of each record is checked. If the time corresponding to the window timestamp is earlier than the current time minus 24 hours, the record is deleted from the topic staging table. For example, if the current time is 10:58:00 on June 28, 2025, then 24 hours ago was 10:58:00 on June 27, 2025. Any record with a window timestamp earlier than this time will be removed. This maintenance strategy ensures the timeliness of data in the topic staging table while optimizing the efficiency of storage resource utilization.

[0049] The topic documentation module receives topic summaries and creative context identifiers from mobile devices and combines them with generated window timestamps. This information is then written into a structured topic temporary storage table, providing ordered data input for the AI-based ad tagging system. The coarse-grained nature of the topic summaries accommodates privacy requirements, while the introduction of window timestamps preserves the temporal information of the data, enabling the system to conduct in-depth analysis over time in subsequent steps. The topic temporary storage table stores records in the form of topic summaries, creative context identifiers, and window timestamps, ensuring data integrity in both time and content. This facilitates the direct use of this information for tag inference and decision support in subsequent steps.

[0050] The topic archiving module has completed the process of receiving topic summaries and context identifiers for delivered materials from the mobile device, and generating a window timestamp for each record. This data is then written to a topic temporary storage table in chronological order. The topic temporary storage table provides a coarse-grained foundation of user interest data for subsequent processing. However, due to privacy protection mechanisms, the topic summaries on the mobile device only provide a general overview of interests, lacking fine-grained behavioral details, making it difficult to directly infer accurate interest tags. To address this issue, the context weaving module uses a cross-window semantic resonance retrieval algorithm to extract continuous topic associations from the topic temporary storage table, generate a progressive resonance graph, and store it in a resonance buffer. This aims to connect discrete topic summaries on the mobile device into semantically coherent topic paths, providing richer semantic information for subsequent trajectory reconstruction and tag inference.

[0051] The purpose of the network weaving module is to read data from the topic temporary table, use a cross-window semantic resonance retrieval algorithm to filter consecutive topic pairs, generate a progressive resonance graph, and store it in the resonance buffer. The following is a detailed breakdown of the technical logic:

[0052] S2.1, Semantic embedding of topic summaries on the computing edge;

[0053] In processing cross-window semantic resonance retrieval, the first step is to convert the client-side topic summaries from text to a vector representation suitable for mathematical calculations. Text data itself cannot be directly compared for similarity, but semantic embedding techniques can map the semantic information of the text to a high-dimensional vector space, making the representations of semantically similar texts more similar in this space. The system uses a pre-trained semantic embedding model, such as a transformer model based on a bidirectional encoder representation, to transform each client-side topic summary into a fixed-dimensional vector. Specifically, taking the client-side topic summary "sports" as an example, it is input into the pre-trained model. After processing through a multi-layer neural network, the model outputs a 768-dimensional vector, which is the semantic embedding representation of "sports." This entire process leverages the model's ability to understand deep semantic meaning, capturing the complex semantic features inherent in the text. In this way, the system transforms client-side topic summaries from discrete text data into continuous vector data, providing an accurate and consistent input foundation for subsequent semantic similarity calculations. The advantage of this method is that it can effectively preserve the semantic information of the text, making semantically similar topics have measurable proximity in the vector space, thereby improving the accuracy of subsequent processing.

[0054] The training of semantic embedding models relies on large-scale text corpora, generating high-quality semantic representations through self-supervised learning methods. Taking the pre-trained model BERT as an example, its training logic mainly includes the following steps: First, using large-scale text data (such as Wikipedia, news articles, etc.), the model learns word-level contextual relationships through the Masked Language Model (MLM) task. Specifically, some words in the input text are randomly masked, and the model predicts these masked words based on the context. Simultaneously, the model learns sentence-level semantic associations through the Next Sentence Prediction (NSP) task, determining whether two sentences are adjacent in the original text. Second, the model adopts a bidirectional transformer architecture, which can simultaneously capture the contextual information on both sides of a word, thereby generating a more comprehensive semantic representation. During training, the model parameters are adjusted through the backpropagation algorithm to minimize prediction errors. After multiple rounds of optimization, the high-dimensional embedding vectors generated by the model can accurately reflect the semantic similarity of the text in high-dimensional space, providing a reliable basic representation for subsequent tasks.

[0055] S2.2, Calculate the semantic similarity of topic summaries at the ends of adjacent windows;

[0056] To identify semantically similar end-topic summary pairs within adjacent time windows, similarity calculations are needed for the embedding vectors of end-topic summaries in different time windows. Since the similarity between vectors cannot be directly derived from text comparison, the cosine similarity of the angle between the vectors is used as a metric. The semantic association is determined by calculating the directional consistency of the two vectors in high-dimensional space. The specific calculation process is as follows: For two adjacent window timestamps, such as window timestamp m and window timestamp n, where window timestamp n is later than window timestamp m, the system first extracts the embedding vector of an end-topic summary A from window timestamp m and the embedding vector of an end-topic summary C from window timestamp n. Next, the dot product of these two vectors is calculated, i.e., the summation of the product of corresponding elements; then, the magnitudes of the two vectors are calculated, i.e., the square root of the sum of the squares of the elements in each dimension of each vector; finally, the dot product is divided by the product of the two magnitudes to obtain the cosine similarity value. Assuming the calculation results show that the cosine similarity between end-topic summary A and end-topic summary C is 0.85, this indicates that they have high semantic similarity.

[0057] S2.3, Filter consecutive topic pairs;

[0058] After obtaining the semantic similarity between end-topic summaries within adjacent windows, it is necessary to filter out topic pairs with sufficiently strong semantic association from these similarity values ​​to form continuous topic pairs. Since not all calculated similarities reflect meaningful semantic continuity, a similarity threshold, such as 0.8, is set as a filtering criterion. The specific processing procedure is as follows: for each pair of end-topic summaries from adjacent windows, such as end-topic summary A in window timestamp m and end-topic summary C in window timestamp n, if their cosine similarity is 0.85, which is greater than the preset threshold of 0.8, then these two are considered to be semantically similar enough to form a continuous topic pair.

[0059] The system then records the topic summaries from both ends and their corresponding window timestamps. For example, it labels topic summaries A and C as consecutive topic pairs and associates them with window timestamps m and n. Through this filtering, the system can exclude topic pairs with weak semantic connections and retain only strongly related topic pairs, ensuring that subsequent analysis is based on a coherent semantic context.

[0060] S2.4, generating a progressive resonance diagram;

[0061] To organize the selected consecutive topic pairs into a structured representation that reflects temporal and semantic relationships, a directed graph called a progressive resonance graph was constructed. Individual consecutive topic pairs cannot intuitively demonstrate the overall evolutionary path, while a directed graph can clearly express the connections and temporal order between topics through nodes and edges. The specific generation process is as follows: each end-side topic summary is treated as a node in the graph; for example, end-side topic summary A and end-side topic summary C correspond to two nodes respectively. For the selected consecutive topic pairs, such as end-side topic summary A and end-side topic summary C, a directed edge is added from end-side topic summary A to end-side topic summary C, with the edge direction pointing from window timestamp m (earlier) to window timestamp n (later). Furthermore, each edge also records associated attributes, including semantic similarity values ​​(e.g., 0.85) and time differences (e.g., the time interval between window timestamp n and window timestamp m is 300 seconds). In this way, the progressive resonance graph not only shows the semantic connections between end-side topic summaries but also preserves their temporal order. This structuring clearly reflects the path of user interests evolving over time, facilitating subsequent analysis of semantic drift trends and the progressive relationships between topics.

[0062] S2.5, stored in the resonance buffer;

[0063] The completed progressive resonance graph needs to be stored in a temporary area for easy access and use later, supporting further system processing. Since the progressive resonance graph contains multiple nodes and edges, directly storing it as a graph structure may not facilitate fast querying; therefore, it is stored in the resonance buffer in the form of an adjacency list. The specific storage process is as follows: For each node in the progressive resonance graph, such as end-side topic summary A, an adjacency list entry is generated, recording the node it points to and the edge attributes. For example, the adjacency list entry for end-side topic summary A is "points to end-side topic summary C, similarity is 0.85, time difference is 300 seconds," indicating that end-side topic summary A is connected to end-side topic summary C through an edge, along with the corresponding attribute information. The resonance buffer saves the entire progressive resonance graph in this form, enabling the system to quickly retrieve the subsequent nodes and associated attributes of any end-side topic summary. Furthermore, when the topic temporary storage table receives new data, the progressive resonance graph is updated according to the latest data and re-stored in the resonance buffer to maintain the real-time nature of the structure. The advantage of this storage method is that it efficiently represents the structure of the graph and facilitates dynamic updates and fast access to data, providing convenient support for subsequent semantic analysis and trajectory processing.

[0064] The aforementioned context weaving module reads data from the topic temporary storage table and, through a cross-window semantic resonance retrieval algorithm, progressively completes the entire process of semantic embedding calculation for edge-side topic summaries, adjacent window similarity calculation, continuous topic pair filtering, progressive resonance graph generation, and storage in the resonance buffer. This connects discrete edge-side topic summaries into topic paths with semantic and temporal relevance, overcoming the limitations of coarse data granularity and ultimately forming a structured progressive resonance graph. This provides accurate and coherent contextual information for subsequent semantic drift analysis and trajectory reconstruction.

[0065] The network weaving module uses a cross-window semantic resonance retrieval algorithm to concatenate edge-side topic summaries into a progressive resonance graph, which is then stored in a resonance buffer. Each topic path in the progressive resonance graph reflects the potential trajectory of user interests evolving over time, providing a foundation for subsequent label classification. However, due to feature compression and broken association links caused by privacy encapsulation, topic paths may exhibit semantic drift or instability, reducing accuracy when directly inferring labels based on these paths. To address this issue, the trajectory reconstruction module calculates semantic drift curvature and extracts key features, evaluates the credibility of trajectory merging, and then merges nodes to construct pseudo-continuous trajectories and labels candidate attributes, providing high-quality input for the label refinement module to generate refined interest labels.

[0066] The trajectory reconstruction module processes each topic path in the progressive resonance graph within the resonance buffer, aiming to overcome semantic drift and path instability issues. The following is a detailed breakdown of the technical logic:

[0067] S3.1, Calculate the semantic drift curvature of each topic path;

[0068] In the progressive resonance graph, each topic path reflects the trajectory of user interests evolving over time. Due to privacy protection mechanisms, edge topic summaries only provide a coarse-grained overview of interests, and semantic incoherence or drift may occur in the paths. To quantify the degree of this drift and assess the reliability of the paths, it is necessary to calculate the semantic drift curvature by measuring the change in the direction of the embedding vectors between adjacent nodes. The specific calculation process is as follows: For every three consecutive nodes on the path, such as the previous edge topic summary, the current edge topic summary, and the next edge topic summary, firstly, the vector difference from the previous edge topic summary to the current edge topic summary is calculated, that is, the embedding vector of the current edge topic summary is subtracted component by component from the embedding vector of the previous edge topic summary to obtain the first vector difference. Next, the vector difference from the current edge topic summary to the next edge topic summary is calculated, that is, the embedding vector of the next edge topic summary is subtracted component by component from the embedding vector of the current edge topic summary to obtain the second vector difference. Next, the angle between these two vector differences is calculated. This is done by first calculating the length of each vector difference (the square root of the sum of the squares of each vector component), then calculating the dot product of the two vector differences (multiplying the corresponding components and summing them), dividing the dot product by the product of the two lengths, and taking the inverse cosine of the result. This angle is called the open angle. After calculating all open angles along the path, the average of these angles is taken (the sum of all open angles divided by the total number of open angles), defined as the semantic drift curvature. A small open angle indicates smooth semantic changes, while a large open angle indicates drastic semantic drift. This calculation method can intuitively reflect the dynamic characteristics of semantic changes along the path, providing a quantitative basis for subsequent trajectory merging, thereby improving the accuracy of label classification.

[0069] S3.2, Extract the key feature – curvature amplitude index;

[0070] When analyzing the stability of a thematic path, simply using the average value of the semantic drift curvature may not fully capture the fluctuations in the opening angles within the path. To more comprehensively describe the path's fluctuation characteristics, it is necessary to extract the curvature amplitude index by calculating the dispersion of the opening angles. The specific calculation process is as follows: First, collect all the opening angle values ​​calculated in the sub-trajectory reconstruction module 1 along the path, forming an opening angle sequence. Then, arrange these opening angle values ​​in ascending order to determine the value at the 25th position (25th quantile) and the value at the 75th position (75th quantile). The curvature amplitude index is defined as the difference between the 75th quantile and the 25th quantile. This difference reflects the degree of concentration or dispersion in the opening angle distribution. If the difference is small, the opening angle changes are more concentrated, and the semantic drift is relatively stable; if the difference is large, the opening angle fluctuations are significant, and the semantic drift is unstable. Features extracted in this way can effectively resist outlier interference, accurately characterize the fluctuation range of the path, provide a more detailed evaluation standard for subsequent trajectory merging, and thus improve the reliability of merging decisions.

[0071] S3.3, Extract the key feature – path stability potential;

[0072] The stability of a path is related not only to the curvature and fluctuation amplitude of semantic drift, but also to topic similarity and time jumps. To comprehensively consider these factors, an index is needed to simultaneously evaluate semantic similarity and temporal stability, called the path stability potential. The specific calculation process is as follows: First, a sliding window technique is used on the path, with a window size of 3, meaning three consecutive end-side topic summaries are taken each time. For each sliding window, the cosine similarity between adjacent end-side topic summaries is calculated, i.e., the dot product of the two embedding vectors is divided by the product of their lengths to obtain the similarity value within each sliding window. Then, the average of these similarity values ​​is calculated, called the mean similarity within the sliding window. Next, the mean similarity within all sliding windows on the path is summed and divided by the total number of sliding windows to obtain the overall mean similarity of the path. Then, the time difference between each edge in the path is calculated, i.e., the time interval between adjacent nodes, forming a jump distance sequence. The average of these jump distances is calculated, and the sum of the absolute values ​​of the deviations of each jump distance from the average is divided by the total number of jump distances to obtain the average jump distance deviation. The reciprocal of this deviation is taken as the jump number balance. Finally, the path stability potential is defined as the product of the overall mean path similarity and the hop count balance. This metric combines local semantic consistency and the stationarity of temporal jumps, enabling a comprehensive assessment of path stability and providing more accurate reference information for trajectory merging.

[0073] S3.4, Evaluate the reliability of trajectory merging;

[0074] After obtaining the curvature amplitude exponent and path stability potential, these two features need to be transformed into a reliability assessment for trajectory merging. Due to the potential sparsity and uncertainty in historical data, a Bayesian sparse regressor is chosen as the evaluation model. The specific calculation process is as follows: First, the regressor is trained using historical data. The input is the curvature amplitude exponent and path stability potential for each path, and the output is the manually labeled merging reliability, with a value of 0 or 1, representing unreliable or reliable. The training process optimizes the regression parameters using Bayesian methods to obtain a model capable of predicting merging reliability. For the current path, its curvature amplitude exponent and path stability potential are calculated, and these two values ​​are input into the trained regressor. The model outputs a value between 0 and 1, called the merging reliability coefficient. A value close to 1 indicates high merging reliability, while a value close to 0 indicates low reliability. This method uses machine learning technology to achieve automatic mapping from features to reliability, improving the intelligence and accuracy of merging decisions.

[0075] S3.5, merge nodes to construct pseudo-continuous trajectories and label candidate attribution labels;

[0076] After calculating the merging confidence coefficient, it is necessary to determine whether to merge paths based on this coefficient and generate candidate labels for the merged trajectory. The specific process is as follows: A threshold is set, for example, 0.7. If the merging confidence coefficient of the current path is greater than this threshold, the path is considered sufficiently stable, and the merging operation is performed. During merging, all end-side topic summary nodes in the path are integrated into a pseudo-node, representing a continuous interest trajectory. The content of the pseudo-node is jointly represented by the included end-side topic summaries. Then, candidate labels are generated based on the content of the pseudo-node, for example, by extracting keywords or semantic clustering to obtain label descriptions like "sports-football-match". If the merging confidence coefficient is lower than the threshold, merging is not performed, and the original state of the path is maintained. This processing method transforms discrete topic paths into coherent interest expressions, providing high-quality input for subsequent label inference, thereby improving the accuracy and practicality of label classification.

[0077] The above steps calculate semantic drift curvature, extract curvature amplitude exponent and path stability potential, and use a Bayesian sparse regressor to evaluate merging credibility, ultimately achieving topic path merging and candidate labeling. Each sub-step is closely linked, progressively resolving the semantic drift and path instability issues caused by privacy encapsulation, generating semantically coherent pseudo-continuous trajectories, and ensuring the accuracy and reliability of subsequent label inference.

[0078] The topic documentation module and trajectory reconstruction module have successfully constructed pseudo-continuous trajectories and labeled candidate tags by summarizing topics on the receiving end, building a progressive resonance graph, and merging topic paths, thus initially connecting the user's coarse-grained interest context. However, the pseudo-continuous trajectories and their tag candidates only reflect the surface expression of user interests and lack in-depth analysis of the correlation between user behavior and the context of advertising materials. To further improve tag classification accuracy and support real-time bidding, the tag refinement module introduces a tag inference engine, which uses a hybrid attention parsing algorithm to refine the pseudo-continuous trajectories, generating high-precision interest tags and storing them in a tag tree mapping library, providing reliable data support for subsequent steps.

[0079] like Figure 2 As shown, the tag refinement module takes the pseudo-continuous trajectory generated by the trajectory reconstruction module as input, uses the tag inference engine combined with a hybrid attention parsing algorithm to generate refined interest tags, and writes them into the tag tree mapping library. The following is a detailed breakdown of the technical logic:

[0080] S4.1, Input the pseudo-continuous trajectory to the label inference engine;

[0081] In advertising tag classification systems, to accurately capture the dynamic evolution of user interests, pseudo-continuous trajectories obtained through preprocessing need to be input into a specially designed tag inference engine for analysis. The pseudo-continuous trajectory is the core data structure for user interest evolution, consisting of a series of end-side topic summaries, each corresponding to an embedding vector containing rich semantic information. The processing begins by extracting the pseudo-continuous trajectory from the resonance buffer; specifically, this involves obtaining all end-side topic summary nodes and their corresponding embedding vectors contained within the pseudo-continuous trajectory. For example, suppose a pseudo-continuous trajectory contains three end-side topic summaries: A, C, and E, with corresponding embedding vectors A, C, and E, respectively. After extraction, these embedding vectors are sequentially fed into the tag inference engine according to their order within the pseudo-continuous trajectory. Upon receiving these vectors, the tag inference engine performs further calculations and analysis to identify the semantic relationships and trends in interest evolution between the embedding vectors. In this way, the system can comprehensively understand the dynamic changes in user interests, providing reliable data support for generating accurate interest tags. By fully utilizing the semantic information of the embedded vectors, the accuracy and comprehensiveness of the interest tag inference process are ensured.

[0082] S4.2, Hybrid Attention Parsing Algorithm;

[0083] To extract tags highly matching user interests from pseudo-continuous trajectories, a hybrid attention parsing algorithm is employed to deeply mine the semantic associations between edge-side topic summaries and their matching relationships with ad creative context identifiers. The hybrid attention mechanism consists of two parts: self-attention and cross-attention, used to handle dependencies within the pseudo-continuous trajectory and the association between user interests and ad content, respectively. The specific calculation process is as follows: First, for each edge-side topic summary in the pseudo-continuous trajectory, its attention score is calculated relative to other edge-side topic summaries. The attention score is calculated by taking the dot product of two embedding vectors, which reflects the similarity between the vectors; then, the dot product is divided by the product of the magnitudes of the two vectors for normalization; finally, it is divided by a preset temperature parameter, which adjusts the smoothness of the attention score distribution to prevent it from being overly concentrated or dispersed. Based on these attention scores, a context vector for each edge-side topic summary is generated. Specifically, the embedding vectors of all other edge-side topic summaries are weighted and summed according to their corresponding attention scores to obtain a comprehensive representation vector. Next, the generated context vector is associated with the embedding vector of the ad creative context identifier using the cross-attention mechanism. The calculation method is similar. First, the attention score between the context vector and the embedding vector of each ad creative's context identifier is calculated, i.e., the score is obtained by adjusting the dot product, modulus normalization, and temperature parameter. Then, based on these attention scores, the embedding vectors of the ad creative's context identifiers are weighted and summed to generate a hybrid representation vector. Finally, the hybrid representation vector is input into a multilayer perceptron, which processes it through a multilayer neural network and outputs a probability vector, where each component of the probability vector corresponds to the probability of a label category. The system selects the label category corresponding to the component with the highest value in the probability vector as the fine-grained interest label for the topic summary on that side. This calculation method can effectively capture the semantic dependencies within pseudo-continuous trajectories and the deep correlation between user interests and ad content, thereby significantly improving the accuracy of label classification and the relevance of ad delivery.

[0084] Among them, the Multilayer Perceptron (MLP) is a type of feedforward neural network. Its construction and processing are clear and structured, mainly including network architecture design, forward propagation, loss calculation, backpropagation, and parameter optimization. The following is its complete description:

[0085] First, the network architecture consists of an input layer, multiple hidden layers, and an output layer. The input layer receives data (such as a mixed representation vector), the hidden layers process the data through fully connected layers and non-linear activation functions (such as ReLU) to enhance the model's expressive power, and the output layer generates results according to task requirements (such as the probability distribution of label categories, using the Softmax function). For example, a typical MLP might contain a 768-dimensional input layer, two hidden layers with 256 and 128 neurons respectively, and an output layer with the same number of label categories.

[0086] Forward propagation is the core step in data processing: the input data undergoes linear transformation through the weight matrix and bias vector of each layer, and is then calculated through the activation function to finally output the result.

[0087] During training, a loss function (such as cross-entropy loss) measures the difference between the prediction and the true value. Backpropagation adjusts the weights through gradient calculation, and optimization algorithms (such as Adam) update the parameters. This process is repeated in multiple iterations until the model performance improves. During inference, predictions are generated from the input simply through forward propagation.

[0088] Through the nonlinear transformation of multilayer neural networks, multilayer perceptrons can effectively process complex data and complete tasks such as classification.

[0089] S4.3, write to the tag tree mapping library;

[0090] After generating fine-grained interest tags, to ensure efficient management and rapid access by the system, these tags need to be associated with their corresponding pseudo-continuous trajectories and stored in a dedicated database. The tag tree mapping library is a structured storage system designed to preserve fine-grained interest tags and their mapping relationships with pseudo-continuous trajectories. The specific operation process is as follows: For each pseudo-continuous trajectory, its unique identifier is used as the key, and the generated fine-grained interest tag is used as the corresponding value, recorded in the tag tree mapping library as a key-value pair. For example, assuming the unique identifier of a pseudo-continuous trajectory is "PCT1", and the fine-grained interest tag generated through the aforementioned steps is "football", then a record is stored in the tag tree mapping library indicating that "PCT1" corresponds to "football". Through this key-value pair storage method, the system can quickly retrieve the corresponding fine-grained interest tag based on the unique identifier of the pseudo-continuous trajectory in subsequent operations. The advantage of this storage method is that it achieves efficient management and rapid access to tag data, providing timely data support for real-time ad delivery and bidding decisions, thereby improving the overall response speed and delivery efficiency of the system.

[0091] From the topic creation module to the tag refinement module, the user's coarse-grained interest context has been successfully transformed into high-precision tag data and stored in the tag tree mapping library by summarizing topics on the receiving end, constructing a progressive resonance graph, merging topic paths, and generating refined interest tags. These refined interest tags overcome the information gap problem caused by privacy encapsulation through the construction of pseudo-continuous trajectories and the derivation of a hybrid attention parsing algorithm. However, to achieve accuracy and efficiency in ad delivery, the latest data in the tag tree mapping library needs to be integrated with the bidding decision engine to ensure that the engine can make real-time decisions based on the latest and most accurate user interest data. Therefore, the mapping distribution module focuses on the complete process of data synchronization and accuracy evaluation to support the optimization of subsequent bidding strategies.

[0092] The mapping distribution module aims to synchronize the latest entries in the tag tree mapping library to the bidding decision engine by index, and attach precision tags. The following is a detailed breakdown of the technical logic:

[0093] S5.1, Synchronize the latest entries by index;

[0094] In the advertising tag classification system, the tag tree mapping library stores refined interest tags generated based on user behavior data. These tags reflect users' real-time interests. To enable the bidding decision engine to use the latest user interest information in a timely manner, it needs to extract and synchronize the most recently generated entries from the tag tree mapping library. The specific processing logic is as follows: First, the system accesses the tag tree mapping library and filters out the records with the latest timestamps. Each record contains a unique identifier for a pseudo-continuous trajectory and a corresponding refined interest tag. Next, the system uses the unique identifier of the pseudo-continuous trajectory as an index to transmit and store the corresponding refined interest tag in the bidding decision engine's data cache. For example, assuming there is a record in the tag tree mapping library where the unique identifier for the pseudo-continuous trajectory is a specific string, and the corresponding refined interest tag is the word "football" representing the user's interest, the system will associate "football" with the unique identifier of the pseudo-continuous trajectory and store it in the bidding decision engine's data cache. This synchronization process is executed periodically or triggered when user behavior is updated to ensure that the data remains up-to-date. In this way, when processing ad delivery requests, the bidding decision engine can quickly retrieve the user's current interest information, thereby achieving ad delivery that matches the user's real-time interests, improving ad delivery effectiveness and user experience.

[0095] S5.2, Additional precision markings;

[0096] Based on synchronized fine-grained interest tags, a precision marker is attached to each tag to quantify its accuracy, thereby assisting the bidding decision engine in formulating more reasonable bidding strategies. The specific processing logic is as follows: The precision marker originates from the confidence calculation when the tag inference engine generates fine-grained interest tags. In the tag inference engine, a multilayer perceptron processes user behavior data and outputs a probability vector, where each component represents the likelihood of a certain interest category. The system selects the interest category corresponding to the component with the largest value from this probability vector as the fine-grained interest tag and uses the value of that component as the precision marker. For example, suppose the multilayer perceptron processes user data and outputs a probability vector containing three components: 0.1, 0.85, and 0.05. If 0.85 corresponds to the interest category "football," the system determines "football" as the fine-grained interest tag and uses 0.85 as the precision marker. Subsequently, the system appends "football" and 0.85 together to the data associated with the unique identifier of the pseudo-continuous trajectory and transmits it to the bidding decision engine. This additional process of precision tagging allows the bidding decision engine to adjust its decisions based on the accuracy of the tags. For example, when dealing with highly accurate tags, it can prioritize ads with higher matching scores, thereby optimizing the accuracy and effectiveness of ad delivery.

[0097] S5.3, synchronized to the bidding decision engine;

[0098] After preparing the fine-grained interest tags and their precision markers, this data is organized and stored in the bidding decision engine to support fast access and efficient decision-making. The specific processing logic is as follows: Using the unique identifier of the pseudo-continuous trajectory as the key, the corresponding fine-grained interest tag and precision marker are combined as the value, forming key-value pairs stored in the bidding decision engine's data structure. For example, for a unique identifier of a pseudo-continuous trajectory, assuming its corresponding fine-grained interest tag is "football" and its precision marker is 0.85, the system creates a record in the bidding decision engine's data storage, where the key is the unique identifier of the pseudo-continuous trajectory, and the value is a combination of "football" and 0.85. After storage, when the bidding decision engine receives an ad delivery request, it queries the corresponding key-value pair using the unique identifier of the pseudo-continuous trajectory to obtain the fine-grained interest tag and precision marker, and adjusts the bidding strategy based on the precision marker value, for example, assigning a higher matching priority to tags with higher precision. This key-value pair storage method, through structured data organization, ensures that the bidding decision engine can efficiently retrieve and utilize user information, thereby achieving a highly targeted and efficient decision-making process in ad delivery.

[0099] The bidding decision engine operates as follows: Upon receiving an ad delivery request, it first retrieves the corresponding pseudo-continuous trajectory identifier from the data cache based on the user information (such as user ID or device ID) in the request. Then, it uses this identifier to query the stored key-value pairs to obtain the user's refined interest tags and precision markers. Next, the bidding decision engine matches the ad inventory based on the interest tags, filters out relevant ad candidate sets, and prioritizes matching ads with high-precision tags based on the precision markers. Subsequently, the bidding decision engine combines ad characteristics (such as type, budget, and historical click-through rate) and real-time user behavior data to calculate the expected revenue (combining metrics such as click-through rate and conversion rate) for each ad using a machine learning model. Finally, based on the expected revenue and the advertiser's bidding strategy, the bidding decision engine determines the ad display order and bid amount to maximize platform revenue and user experience while meeting the advertiser's goals.

[0100] 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.

[0101] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0102] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0103] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0104] 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 big data advertising tag classification system based on AI analysis, characterized in that, include: Topic documentation module: The topic summary and the context identifier of the delivery material on the receiving end are written into the topic temporary storage table in sequence and a window timestamp is attached; The network weaving module reads the topic temporary table, uses the cross-window semantic resonance retrieval algorithm to filter out continuous topic pairs, generates a progressive resonance graph and stores it in the resonance buffer; The processing content of the vein weaving module is as follows: Retrieves records containing edge-side topic summaries, delivery material context identifiers, and window timestamps from the topic temporary storage table; converts edge-side topic summaries into vector representations using a pre-trained semantic embedding model; calculates the cosine similarity between edge-side topic summary vectors in adjacent time windows; filters out consecutive topic pairs based on a preset similarity threshold; constructs a progressive resonance graph, where nodes are edge-side topic summaries, directed edges represent consecutive topic pairs, and edge attributes include cosine similarity and time difference; stores the progressive resonance graph in the resonance buffer in the form of an adjacency list. Trajectory reconstruction module: First, calculate the semantic drift curvature for each topic path within the resonance buffer, extract key features to evaluate the trajectory merging credibility for each topic path, merge nodes to construct pseudo-continuous trajectories when the criteria are met, and label candidate attribution labels at the same time; The tag inference module feeds pseudo-continuous trajectories into the tag inference engine, uses a hybrid attention parsing algorithm to generate fine interest tags, and then writes them into the tag tree mapping library. Mapping distribution module: Synchronizes the latest entries in the tag tree mapping library to the bidding decision engine by index, along with a precision tag.

2. The big data advertising tag classification system based on AI analysis according to claim 1, characterized in that, The processing content of the subject file creation module is as follows: The device receives a topic summary and a delivery material context identifier from the receiving end. A window timestamp is generated for each received data record. The window timestamp is based on the current time when the data is received and aligned to a preset time window length. The device-side topic summary, delivery material context identifier, and window timestamp are written as a record into a topic temporary storage table. The topic temporary storage table stores records in the order of receipt to preserve the temporal characteristics of the data. The topic temporary storage table is maintained by cleaning up records whose window timestamps are earlier than the current time minus a preset time period to ensure data freshness and storage efficiency.

3. The big data advertising tag classification system based on AI analysis according to claim 1, characterized in that, The trajectory reconstruction module processes the following: The topic paths in the progressive resonance graph are obtained from the resonance buffer; the semantic drift curvature is calculated for each topic path, which is determined by measuring the change in the direction of the embedding vector between adjacent nodes; the curvature amplitude exponent and the path stability potential are extracted as key features, the curvature amplitude exponent is used to measure the fluctuation of semantic drift by calculating the interquartile range of the open angle, and the path stability potential comprehensively evaluates the topic similarity and the stability of time jump. The reliability of trajectory merging is evaluated using a Bayesian sparse regressor based on the curvature amplitude exponent and the path stability potential, and the merging reliability coefficient is output. When the merging reliability coefficient is higher than a preset threshold, pseudo-continuous trajectories are constructed by merging path nodes, and candidate attribution labels are generated for pseudo-continuous trajectories.

4. The big data advertising tag classification system based on AI analysis according to claim 3, characterized in that, The trajectory reconstruction module also includes the following processing: The curvature amplitude index quantifies the fluctuation of semantic drift by calculating the interquartile range of the open angles. The specific steps are as follows: First, collect the open angles between all adjacent nodes on the topic path. These open angles represent the degree of change in semantic direction. Then, sort these open angles by size and find the 25th quantile Q1 and the 75th quantile Q3. Finally, the curvature amplitude index is defined as the difference between Q3 and Q1, that is, the interquartile range IQR = Q3 - Q1.

5. The AI-based big data advertising tag classification system according to claim 4, characterized in that, The trajectory reconstruction module also includes the following processing: The path stability potential measures the overall stability of a path by comprehensively evaluating the topic similarity and the stability of time jumps. The calculation process includes: First, using the sliding window technique, the local topic similarity is calculated segment by segment along the topic path, and then the mean of the similarities within all sliding windows is taken to reflect the semantic consistency of the path; Second, the stability of each time jump in the path is analyzed, that is, the time interval between nodes, which is measured by calculating the reciprocal of the variance of the time jump. The smaller the variance, the more balanced and stable the time jump; Finally, the path stability potential is defined as the product of the mean topic similarity and the stability of the time jump.

6. The big data advertising tag classification system based on AI analysis according to claim 5, characterized in that, The tag recommendation module processes the following: The pseudo-continuous trajectory, including the edge topic summary and its embedding vector, is extracted from the resonant buffer and input into the tag inference engine. The pseudo-continuous trajectory is processed using a hybrid attention parsing algorithm. The semantic dependency between edge topic summaries is calculated through a self-attention mechanism to generate a context vector. The context vector is associated with the embedding vector of the advertising material context identifier through a cross-attention mechanism to generate a hybrid representation vector. The hybrid representation vector is input into a multilayer perceptron to generate a probability vector. The component with the highest probability is selected as the fine interest tag. The fine interest tag is associated with the unique identifier of the pseudo-continuous trajectory and stored in the tag tree mapping library in the form of key-value pairs.

7. The big data advertising tag classification system based on AI analysis according to claim 6, characterized in that, The tag recommendation module also includes the following processing functions: The hybrid attention mechanism is divided into two parts: self-attention and cross-attention, which are used to handle the dependencies within pseudo-continuous trajectories and the association between user interests and advertising content, respectively.

8. The big data advertising tag classification system based on AI analysis according to claim 7, characterized in that, The processing content of the mapping and distribution module is as follows: Extract the latest fine interest tags and their corresponding pseudo-continuous trajectory unique identifiers from the tag tree mapping library, and use the pseudo-continuous trajectory unique identifiers as indexes to transfer the fine interest tags to the data cache of the bidding decision engine. Based on the label probability vector generated by the label inference engine, the maximum component value is selected as the precision marker, and the precision marker and the fine interest label are attached together to the data associated with the pseudo-continuous trajectory unique identifier. Using the pseudo-continuous trajectory unique identifier as the key, the fine interest label and the precision marker are combined as the value to form a key-value pair stored in the bidding decision engine's data structure.

Images