Big data advertisement label classification system based on AI analysis

Abstract

The invention discloses a big data advertisement label classification system based on AI analysis, particularly relates to the field of advertisement data processing and label intelligent classification, and is used for solving the problems of interest identification distortion and released label drift caused by coarse granularity of an end-side theme abstract. Judging the combination credibility according to two parameters of a curvature amplitude index and a path stability potential so as to remodel a pseudo-continuous trajectory, and deducing a fine grain interest label under mixed attention modulation; according to the overall architecture, information faults after abstract compression are broken through, audience portraits are close to real-time interests and do not drift along with data reduction, a delivery end bids accurately according to the audience portraits, repeated correction around distortion labels is not needed any more, budget waste is greatly reduced, and 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 always anchors the latest interest coordinate, the advertisement experience meets the user demand, and the brand security risk is inhibited along with the decline of the false bidding.

Description

Technical Field

[0001] The present invention relates to the field of advertising data processing and intelligent label classification, and more specifically, to a big data advertising label classification system based on AI analysis. Background Art

[0002] Mobile advertising is undergoing a privacy sandbox transformation. Mobile operating environments encapsulate user visit traces into on-device topic summaries, exposing only a coarse-grained overview of interests to the outside world. This encapsulation occurs internally on the device, removing the need for external data pipelines to access individual behavioral logs. High-dimensional features such as audience interests, creative semantics, and page context are compressed into limited summaries, disrupting native connections across application touchpoints and significantly reducing the dimensionality of streaming features.

[0003] Ad label classification architectures that rely on AI inference expose subtle shortcomings in this environment. When the training corpus loses its location details, the classification decision area shifts over time, and label outputs tend to float for the same audience. Lacking cross-session alignment cues, online clustering breaks down the same interest into multiple branches, rapidly expanding the label tree and creating mutually exclusive branches. The delivery engine then bids based on distorted interest profiles. When users' true interests have shifted, budgets are repeatedly wasted on ineffective impressions, and resource waste is amplified as the bidding chain progresses. The problem stems from the granularity impact triggered by privacy encapsulation, which manifests itself in feature compression and broken association links. The consequences are concentrated in declining label accuracy and imbalanced delivery.

[0004] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0005] To overcome the above-mentioned shortcomings of the prior art, embodiments of the present invention provide a big data advertising label classification system based on AI analysis. This system concatenates coarse-grained topics on the client side into a coherent behavioral context, then uses the curvature amplitude index and path stability potential dual parameters to determine the combined credibility, thereby reshaping the pseudo-continuous trajectory and deriving fine-grained interest labels under hybrid attention modulation. The overall architecture bridges the information gap after summary compression, ensuring that audience profiles closely align with real-time interests without drifting with data reduction. The delivery end can make precise bids based on this, eliminating the need for repeated corrections around distorted labels, significantly reducing budget waste and balancing exposure frequency distribution. The progressive resonance graph and the label tree mapping library are synchronized, allowing the real-time bidding engine to always anchor the latest interest coordinates. This ensures that the advertising experience is tailored to user needs, and brand safety risks are suppressed as mislabeling decreases, thereby addressing the issues raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: Topic archiving module: The receiving end side topic summary and delivery material context identifier are written into the topic temporary storage table in sequence and appended with the window timestamp; Context 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; Trajectory reconstruction module: First, the semantic drift curvature is calculated for each topic path in the resonance buffer, key features are extracted to evaluate the trajectory merging credibility for each topic path, and nodes are merged to construct a pseudo-continuous trajectory when the criteria are met, while at the same time, the candidate attribution labels are annotated. Label Refinement Module: This module feeds pseudo-continuous trajectories into the label inference engine, uses a hybrid attention parsing algorithm to generate refined interest labels, and then writes them into the label tree mapping library. Mapping delivery module: Synchronizes the latest entries in the label tree mapping library to the bidding decision engine by index, and adds precision tags.

[0007] In a preferred embodiment, the processing content of the subject profile module is as follows: The end-side topic summary and delivery material context identifier are received from the mobile device, and 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 the preset time window length. The end-side topic summary, delivery material context identifier and window timestamp are written as a record to the topic temporary storage table. The topic temporary table stores records in the order of receipt to preserve the timing characteristics of the data. The topic temporary storage table is maintained by clearing records with window timestamps earlier than the current time minus the preset time period to ensure data freshness and storage efficiency.

[0008] In a preferred embodiment, the processing content of the vein weaving module is as follows: Read records containing the end-side topic summary, delivery material context identifier, and window timestamp from the topic temporary table; use the pre-trained semantic embedding model to convert the end-side topic summary into a vector representation; calculate the cosine similarity between the end-side topic summary vectors in adjacent time windows; filter out consecutive topic pairs based on a preset similarity threshold; construct a progressive resonance graph, in which the nodes are end-side topic summaries, the directed edges represent consecutive topic pairs, and the edge attributes include cosine similarity and time difference; store the progressive resonance graph in the resonance buffer in the form of an adjacency list.

[0009] In a preferred embodiment, the processing content of the trajectory reconstruction module is as follows: The topic paths in the progressive resonance graph are obtained from the resonance buffer; the semantic drift curvature is calculated for each topic path, and the semantic drift curvature is determined by measuring the change in the direction of the embedding vector between adjacent nodes; the curvature amplitude index and path stability potential are extracted as key features, the curvature amplitude index measures the fluctuation amplitude of semantic drift by calculating the interquartile range of the opening angle, and the path stability potential comprehensively evaluates the stability of topic similarity and time jumps; a Bayesian sparse regressor is used to evaluate the credibility of trajectory merging based on the curvature amplitude index and path stability potential, and the merging credibility coefficient is output; when the merging credibility coefficient is higher than the preset threshold, the path nodes are merged to construct a pseudo-continuous trajectory, and the attribution label candidates are generated for the pseudo-continuous trajectory.

[0010] In a preferred embodiment, the trajectory reconstruction module further includes the following processing contents: The curvature amplitude index quantifies the fluctuation amplitude of semantic drift by calculating the interquartile range of the opening angle. The specific steps are as follows: First, the opening angles between all adjacent nodes on the topic path are collected. These opening angles represent the degree of change in the semantic direction. Then, these opening angles are sorted by size to find the 25% quantile Q1 and the 75% quantile Q3. Finally, the curvature amplitude index is defined as the difference between Q3 and Q1, that is, the interquartile range IQR = Q3-Q1.

[0011] In a preferred embodiment, the trajectory reconstruction module further includes the following processing contents: Path stability potential measures the overall stability of a path by comprehensively evaluating topic similarity and temporal jump stability. The calculation process includes the following steps: first, local topic similarity is calculated segment by segment along the topic path using a sliding window technique, 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 temporal jump in the path, i.e., the time interval between nodes, is analyzed by calculating the inverse of the temporal jump variance. The smaller the variance, the more balanced and stable the temporal jumps. Finally, path stability potential is defined as the product of the mean topic similarity and the temporal jump stability.

[0012] In a preferred embodiment, the processing content of the tag recommendation module is as follows: Pseudo-continuous trajectories, including the end-side topic summary and its embedding vector, are extracted from the resonance buffer and input into the label inference engine. The pseudo-continuous trajectories are processed using a hybrid attention parsing algorithm. The semantic dependencies between the end-side topic summaries are calculated through the self-attention mechanism to generate a context vector. The context vector is then associated with the embedding vector of the creative context identifier through the cross-attention mechanism to generate a hybrid representation vector. The hybrid representation vector is input into a multi-layer perceptron to generate a probability vector, and the component with the highest probability is selected as the refined interest tag. The refined interest tag is associated with the unique identifier of the pseudo-continuous trajectories and stored in the label tree mapping library as a key-value pair.

[0013] In a preferred embodiment, the tag recommendation module further includes the following processing content: Among them, the hybrid attention mechanism is divided into two parts: self-attention and cross-attention, which are used to process the dependencies within pseudo-continuous trajectories and the association between user interests and advertising content respectively.

[0014] In a preferred embodiment, the processing content of the mapping delivery module is as follows: The latest refined interest tag and its corresponding pseudo-continuous trajectory unique identifier are extracted from the tag tree mapping library, and the refined interest tag is transmitted to the data cache of the bidding decision engine using the pseudo-continuous trajectory unique identifier as the index; based on the tag probability vector generated by the tag inference engine, the maximum component value is selected as the precision mark, and the precision mark and the refined interest tag are attached to the data associated with the pseudo-continuous trajectory unique identifier; with the pseudo-continuous trajectory unique identifier as the key, the refined interest tag and the precision mark are combined as the value to form a key-value pair stored in the data structure of the bidding decision engine.

[0015] The technical effects and advantages of the big data advertising label classification system based on AI analysis of the present invention are as follows: The present invention connects the coarse-grained topics on the terminal side into a coherent behavioral context through time-series resonance retrieval, and then uses the dual parameters of curvature amplitude index and path stability potential to determine the combined credibility, thereby reshaping the pseudo-continuous trajectory and deriving fine-grained interest tags under hybrid attention modulation; the overall architecture breaks through the information gap after summary compression, so that the audience portrait closely follows the real-time interests without drifting with data reduction. The delivery end can bid accurately based on this, and no longer needs to make repeated corrections around distorted tags, which greatly reduces budget waste and balances the exposure frequency distribution; the progressive resonance graph and the label tree mapping library are synchronized, the real-time bidding engine always anchors the latest interest coordinates, the advertising experience is tailored to user needs, and brand safety risks are suppressed as mislabeling decreases; the operation side does not need frequent manual adjustments, the algorithm completes learning and convergence self-consistently, and privacy compliance and delivery efficiency are two-way gains. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a structural diagram of the big data advertising label classification system based on AI analysis in the present invention.

[0017] Figure 2 This is a flow chart of the label recommendation module of the big data advertising label classification system based on AI analysis of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 The present invention provides a big data advertising label classification system based on AI analysis, including: Topic archiving module: The topic summary and delivery material context identifier on the receiving end are written into the topic temporary storage table in sequence and attached with a window timestamp.

[0020] Context weaving module: reads the topic temporary table, uses the cross-window semantic resonance retrieval algorithm to screen out continuous topic pairs, generates a progressive resonance graph and stores it in the resonance buffer.

[0021] Trajectory reconstruction module: First, the semantic drift curvature is calculated for each topic path in the resonance buffer, 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 a pseudo-continuous trajectory, while at the same time, the attribution label candidates are annotated.

[0022] Label refinement module: The pseudo-continuous trajectory is fed into the label inference engine, and refined interest labels are generated using the hybrid attention parsing algorithm, which are then written into the label tree mapping library.

[0023] Mapping delivery module: Synchronizes the latest entries in the label tree mapping library to the bidding decision engine by index, and adds precision tags.

[0024] In the mobile advertising delivery scenario, the privacy sandbox transformation has brought about fundamental changes in the way the mobile operating environment handles user behavior data. User access trajectories are encapsulated as end-side topic summaries, exposing only coarse-grained interest overviews to the outside world, such as "sports" or "technology." This encapsulation method is completed within the device, resulting in the inability of external data pipelines to obtain individual behavior logs. High-dimensional features such as audience interests, material semantics, and page context are compressed into limited summaries, and the native links between cross-application touchpoints are broken, and the feature dimensions are significantly reduced. This granularity impact causes the advertising label classification architecture that relies on artificial intelligence inference to face the problems of declining label accuracy and imbalanced delivery. To address this challenge, the present invention proposes a big data advertising label classification system based on AI analysis, which lays the foundation for the subsequent generation of refined interest tags by receiving and processing end-side topic summaries and delivery material context identifiers. As the starting point of the entire process, the topic archiving module is responsible for receiving and preliminarily organizing these coarse-grained data to ensure that subsequent steps can perform semantic analysis and trajectory reconstruction based on ordered time series information.

[0025] The goal of the topic archiving module is to receive the topic summary and context identifiers of the delivered material on the receiving end, write them sequentially into the topic temporary storage table, and attach a window timestamp to each record. The following is the detailed technical logic: S1.1, the subject summary and delivery material context identification on the receiving end side; In mobile advertising scenarios, due to increasing 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 the ad creative to the system through predefined application programming interfaces (APIs). This design effectively balances privacy protection with ad delivery requirements. The system receives a client-side topic summary and ad creative context identifier from the mobile device via the API. 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 tag associated with the ad creative. For example, "Ad A" indicates 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 and accuracy of data, providing the necessary information source for subsequent classification processing.

[0026] S1.2, generate window timestamp; To preserve the temporal nature of data during processing, a timestamp must be attached to each received data record to track the temporal patterns of data generation. A window timestamp is generated for each data record. This generation process is based on the current time when the data is received and adjusted according to the preset time window length. This time alignment ensures that data records within the same window have the same window timestamp, making it easier for the system to identify temporal correlations during subsequent analysis.

[0027] S1.3, write to the topic temporary storage table; The received end-side topic summary, delivery material 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 end-side topic summary, delivery material context identifier, and window timestamp are written as a complete record to the topic staging table. The topic staging table is a storage structure used to temporarily store recent data. Each record contains three fields. For example, a record might be "Sports|Advertisement A|2025-06-27 10:58:00." When writing, records are added in the order they were received to ensure that the data's temporal characteristics are preserved.

[0028] S1.4, maintain the topic temporary table; As data is continuously received, the number of records in the topic temporary table will gradually increase. In order to avoid the storage space being occupied indefinitely, the system needs to regularly update the content in the table to keep it running efficiently. The topic temporary table is maintained by cleaning up expired records, for example, setting it to only retain records within the last 24 hours. The specific operation is to check the window timestamp of each record. If the time corresponding to the window timestamp is earlier than the current time minus 24 hours, the record will be deleted from the topic temporary table. Taking the current time as 10:58:00 on June 28, 2025 as an example, the time 24 hours ago was 10:58:00 on June 27, 2025. All records with a window timestamp earlier than this time will be removed. This maintenance strategy ensures the timeliness of data in the topic temporary table while optimizing the efficiency of storage resources.

[0029] The topic archiving module receives the end-side topic summary and delivery material context identifier from the mobile device, and combines this information with the generated window timestamp to write it into the topic temporary storage table in a structured form, providing orderly data input for the advertising label classification system based on artificial intelligence analysis. The coarse-grained nature of the end-side topic summary meets the requirements of privacy protection, while the introduction of the window timestamp preserves the timing information of the data, enabling the system to conduct in-depth analysis through the time dimension in subsequent steps. The topic temporary storage table stores records in the form of end-side topic summary, delivery material context identifier, and window timestamp, ensuring the integrity of the data in terms of time and content, making it easier for subsequent steps to directly use this information for label inference and decision support.

[0030] The topic archiving module has completed receiving the end-side topic summary and delivery material context identifier from the mobile device, and has generated a window timestamp for each record, writing this data into the topic temporary storage table in chronological order. The topic temporary storage table provides a coarse-grained user interest data foundation for subsequent processing. However, due to the limitations of the privacy protection mechanism, the end-side topic summary only provides a general interest overview and lacks fine-grained behavioral details, which makes it difficult to directly infer precise interest tags. To solve this problem, 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 the resonance buffer. The aim is to string together discrete end-side topic summaries into a semantically coherent topic path, providing richer semantic information for subsequent trajectory reconstruction and label inference.

[0031] The purpose of the context weaving module is to read data from the topic temporary table, use the cross-window semantic resonance retrieval algorithm to select consecutive topic pairs, generate a progressive resonance graph, and store it in the resonance buffer. The following is a detailed technical logic breakdown: S2.1, compute the semantic embedding of the client-side topic summary; When processing cross-window semantic resonance retrieval, the on-device topic summary must first be converted from text to a vector representation suitable for mathematical calculations. Text data itself cannot be directly compared for similarity. However, semantic embedding technology can map the semantic information of text into 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 the Transformer model based on a bidirectional encoder representation, to convert each on-device topic summary into a fixed-dimensional vector. The specific operation is as follows: Taking the on-device topic summary "sports" as an example, it is input into the pre-trained model. The model processes it through a multi-layer neural network and outputs a 768-dimensional vector, which is the semantic embedding representation of "sports." This process leverages the model's ability to understand the deep semantics of language and capture the complex semantic features contained in the text. In this way, the system converts the on-device topic summary from discrete text data to continuous vector data, providing an accurate and consistent input basis for subsequent semantic similarity calculations. The advantage of this method is that it can effectively preserve the semantic information of the text, so that semantically similar topics have measurable proximity in the vector space, thereby improving the accuracy of subsequent processing.

[0032] The training of semantic embedding models relies on large-scale text corpora, generating high-quality semantic representations through self-supervised learning. Taking the pre-trained BERT model as an example, its training logic primarily involves the following steps: First, leveraging large-scale text data (such as Wikipedia and news articles), the model learns word-level contextual relationships through the Masked Language Model (MLM) task. This involves randomly masking some words in the input text, and the model predicts these masked words based on their 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 employs a bidirectional transformer architecture that simultaneously captures contextual information on both the left and right sides of a word, thereby generating a more comprehensive semantic representation. During training, the model parameters are adjusted using a backpropagation algorithm to minimize prediction error. After multiple rounds of optimization, the high-dimensional embedding vectors generated by the model accurately reflect the semantic similarity of text in a high-dimensional space, providing a reliable foundation for subsequent tasks.

[0033] S2.2, calculate the semantic similarity of the topic summaries on the adjacent window sides; To identify semantically similar pairs of end-side topic summaries within adjacent time windows, it is necessary to calculate the similarity of the end-side topic summaries' embedding vectors within different time windows. Since the similarity between vectors cannot be directly determined through text comparison, the cosine of the angle between the vectors is used as a metric. The degree of semantic association between the two vectors 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, for example, 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-side topic summary A within window timestamp m and the embedding vector of an end-side topic summary C within window timestamp n. Next, the dot product of these two vectors is calculated, which is the sum of the elements in the corresponding dimensions. The modulus of the two vectors is then calculated, which is 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 moduli to obtain the cosine similarity value. Assume that the calculation results show that the cosine similarity between end-side topic summaries A and C is 0.85, indicating high semantic similarity between the two.

[0034] S2.3, screening of consecutive subject pairs; After obtaining the semantic similarity between the end-side topic summaries within adjacent windows, it is necessary to filter out topic pairs with sufficiently strong semantic associations from these similarity values ​​to form continuous topic pairs. Because not all calculated similarities reflect meaningful semantic continuity, a similarity threshold, such as 0.8, is set as a screening criterion. The specific processing process is as follows: for each pair of end-side topic summaries in adjacent windows, for example, end-side topic summary A in window timestamp m and end-side topic summary C in window timestamp n, if their cosine similarity is 0.85, which is greater than the preset threshold of 0.8, the two are considered semantically similar enough to form a continuous topic pair.

[0035] The system then records the pair of client-side topic summaries and their corresponding window timestamp information. For example, it marks client-side topic summaries A and C as a continuous topic pair and associates window timestamps m and n. This filtering allows the system to exclude topic pairs with weak semantic connections and retain only those with strong connections, ensuring that subsequent analysis is based on a coherent semantic context.

[0036] S2.4, generate progressive resonance graph; 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. Separate consecutive topic pairs cannot intuitively represent the overall evolutionary path, but 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 each correspond to two nodes. For each selected consecutive topic pair, 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 direction of the edge being from window timestamp m (earlier) to window timestamp n (later). Furthermore, each edge records associated attributes, including a semantic similarity value (e.g., 0.85) and a time difference (e.g., the time interval between window timestamps n and m is 300 seconds). In this way, the progressive resonance graph not only displays the semantic connections between end-side topic summaries but also preserves their temporal order. This structuring can clearly reflect the path of user interest evolution over time, facilitating subsequent analysis of semantic drift trends and progressive relationships between topics.

[0037] S2.5, stored in resonance buffer; The constructed progressive resonance graph needs to be stored in a temporary area for easy access and use to support further processing by the system. Because the progressive resonance graph contains multiple nodes and edges, storing it directly as a graph structure may not be convenient for fast querying. Instead, 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 table entry is generated, recording the node it points to and the attributes of the edge. For example, the adjacency table 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 via an edge, along with the corresponding attribute information. The resonance buffer stores the entire progressive resonance graph in this format, allowing the system to quickly retrieve subsequent nodes and associated attributes for any end-side topic summary. Furthermore, when the topic temporary storage table receives new data, the progressive resonance graph is updated with the latest data and stored again in the resonance buffer to maintain the real-time nature of the structure. The advantage of this storage method is that it not only efficiently represents the structure of the graph, but also facilitates dynamic updating and rapid access to data, providing convenient support for subsequent semantic analysis and trajectory processing.

[0038] The thread weaving module reads data from the topic temporary storage table and, using a cross-window semantic resonance retrieval algorithm, gradually completes the entire process of computing the semantic embedding of the end-user topic summary, calculating adjacent window similarities, filtering consecutive topic pairs, generating a progressive resonance graph, and storing it in the resonance buffer. This connects discrete end-user topic summaries into semantically and temporally related topic paths, overcoming the limitations of coarse data granularity. Ultimately, a structured progressive resonance graph is formed, providing precise and coherent context for the system's subsequent semantic drift analysis and trajectory reconstruction.

[0039] The context weaving module has used a cross-window semantic resonance retrieval algorithm to concatenate the client-side topic summaries into a progressive resonance graph and store it in the resonance buffer. Each topic path in the progressive resonance graph reflects the potential trajectory of user interests evolving over time, providing a basis for subsequent tag classification. However, due to feature compression and broken association links caused by privacy encapsulation, topic paths may experience semantic drift or instability, and direct label inference based on these paths will reduce accuracy. To address this issue, the trajectory reconstruction module calculates semantic drift curvature and extracts key features to evaluate the credibility of trajectory merging. It then merges nodes to construct pseudo-continuous trajectories and annotates candidate attribution labels, providing high-quality input for the label refinement module to generate refined interest labels.

[0040] The trajectory reconstruction module processes each topic path of the progressive resonance graph in the resonance buffer, aiming to overcome the problems of semantic drift and path instability. The following is a detailed technical logic breakdown: S3.1, calculate the semantic drift curvature of each topic path; In the progressive resonance graph, each topic path reflects the trajectory of user interests evolving over time. Because on-device topic summaries are limited by privacy protection mechanisms and only provide a coarse-grained overview of interests, paths may exhibit semantic incoherence or drift. To quantify the extent of this drift and assess the reliability of the path, semantic drift curvature is calculated 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, current, and next on-device topic summaries, the vector difference between the previous and current on-device topic summaries is first calculated. This is done by subtracting the embedding vector of the previous on-device topic summaries from the embedding vector of the current on-device topic summaries, component by component, to obtain the first vector difference. Next, the vector difference between the current and next on-device topic summaries is calculated. This is done by subtracting the embedding vector of the current on-device topic summaries from the embedding vector of the next on-device topic summaries, component by component, to obtain the second vector difference. Then, the angle between the two vector differences is calculated by first calculating the lengths of the two vector differences, that is, taking the square root of the sum of the squares of each vector component, and then calculating the dot product of the two vector differences, that is, multiplying the corresponding components and summing them. The arc cosine of the dot product is divided by the product of the two lengths to obtain the included angle, which is called the opening angle. After all the opening angles on the path are calculated, the average of these opening angles is taken, that is, the sum of all the opening angles is divided by the total number of opening angles, and this is defined as the semantic drift curvature. A small opening angle means smooth semantic changes, while a large opening angle indicates drastic semantic drift. This calculation method can intuitively reflect the dynamic characteristics of semantic changes in the path, provide a quantitative basis for subsequent trajectory merging, and thus improve the accuracy of label classification.

[0041] S3.2, extract key features - curvature amplitude index; When analyzing the stability of a topic path, using only the average value of semantic drift curvature may not fully capture the fluctuations in the path's opening angles. To more comprehensively characterize the path's fluctuations, it is necessary to calculate the degree of dispersion of the opening angles to extract the curvature amplitude index. The specific calculation process is as follows: First, all the opening angle values ​​calculated in the sub-trajectory reconstruction module 1 on the path are collected to form an opening angle sequence. Then, these opening angle values ​​are sorted from smallest to largest, and the value at the 25th percentile (the 25th percentile) and the value at the 75th percentile (the 75th percentile) in the sequence are determined. The curvature amplitude index is defined as the difference between the 75th percentile and the 25th percentile. This difference reflects the degree of concentration or dispersion of the opening angle distribution. If the difference is small, the opening angle variation is relatively concentrated and the semantic drift is relatively stable. If the difference is large, the opening angle fluctuates significantly and the semantic drift is unstable. The features extracted in this way can effectively resist outlier interference and accurately characterize the fluctuation range of the path, providing more detailed evaluation criteria for subsequent trajectory merging, thereby improving the reliability of merging decisions.

[0042] S3.3, extract key features - path stability potential; Path stability is not only related to the curvature and fluctuation of semantic drift, but also influenced by topic similarity and temporal jumps. To comprehensively account for these factors, a metric is designed to simultaneously assess semantic similarity and temporal stability, called 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. This means that three consecutive end-side topic summaries are taken at a time. For each window, the cosine similarity between adjacent end-side topic summaries is calculated. This is done by taking the dot product of the two embedding vectors and dividing it by the product of their lengths. This yields the similarity value within each window. The average of these similarities, called the mean within-window similarity, is then calculated. Next, the mean within-window similarity values ​​for all windows on the path are summed and divided by the total number of windows to obtain the mean similarity for the entire path. The time difference between each edge in the path, that is, the time interval between adjacent nodes, is calculated to form a sequence of hops. The average of these hops is then calculated. The sum of the absolute deviations of each hop from the average is then divided by the total number of hops to obtain the average hop deviation. The inverse of this sum is taken as the hop balance. Finally, path stability is defined as the product of the mean of the overall path similarity and the hop balance. This metric combines local semantic consistency and temporal jump stability to comprehensively assess path stability and provide more accurate reference information for trajectory merging.

[0043] S3.4, Evaluate the credibility of trajectory merging; After obtaining the curvature amplitude index and path stability potential, these two features need to be converted into a confidence assessment for trajectory merging. Due to the potential sparsity and uncertainty in historical data, a Bayesian sparse regressor is used 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 index and path stability potential of each path. The output is the manually annotated merge confidence level, which takes a value of 0 or 1, indicating unreliability or reliability. The training process optimizes the regression parameters using a Bayesian method to obtain a model capable of predicting merge confidence. For the current path, its curvature amplitude index and path stability potential are calculated and input into the trained regressor. The model outputs a value between 0 and 1, called the merge confidence coefficient. A value close to 1 indicates high merge confidence, while a value close to 0 indicates low confidence. This method uses machine learning techniques to automatically map features to confidence levels, improving the intelligence and accuracy of merging decisions.

[0044] S3.5, merge nodes to construct pseudo-continuous trajectories and mark the candidate attribution labels; After calculating the merge confidence coefficient, it is necessary to use this coefficient to decide whether to merge the paths and generate label candidates for the merged trajectory. The specific processing process is as follows: a threshold is set, such as 0.7. If the merge confidence coefficient of the current path is greater than this threshold, the path is considered sufficiently stable and the merge operation is performed. During the merge, 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 end-side topic summaries it contains. Next, attribution label candidates are generated based on the content of the pseudo-node. For example, by extracting keywords or semantic clustering, a label description such as "sports-football-game" is obtained. If the merge confidence coefficient is lower than the threshold, the merge is not performed and the path is retained in its original state. This processing method transforms discrete topic paths into coherent expressions of interest, providing high-quality input for subsequent label inference, thereby improving the accuracy and practicality of label classification.

[0045] The above steps calculate semantic drift curvature, extract curvature amplitude index and path stability potential, and use a Bayesian sparse regressor to assess merging credibility, ultimately achieving the merging of topic paths and labeling candidate tags. These closely linked sub-steps gradually address 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.

[0046] The topic archiving module and the trajectory reconstruction module have successfully constructed pseudo-continuous trajectories and annotated candidate labels by receiving end-side topic summaries, building a progressive resonance graph, and merging topic paths, thus preliminarily connecting the user's coarse-grained interest context. However, pseudo-continuous trajectories and their label candidates only reflect the surface expression of user interests and lack a deep correlation analysis between user behavior and the context of the advertising material. To further improve the accuracy of label classification and support real-time bidding, the label recommendation module introduces a label inference engine. It uses a hybrid attention parsing algorithm to refine the pseudo-continuous trajectories, generate high-precision interest labels, and store them in the label tree mapping library, providing reliable data support for subsequent steps.

[0047] like Figure 2 As shown in the figure, the label refinement module takes the pseudo-continuous trajectory generated by the trajectory reconstruction module as input, uses the label inference engine combined with the hybrid attention parsing algorithm to generate refined interest tags, and writes them into the label tree mapping library. The following is a detailed technical logic breakdown: S4.1, input pseudo-continuous trajectory to label inference engine; In order to accurately capture the dynamic evolution of user interests in an ad tag classification system, the pseudo-continuous trajectory generated through pre-processing needs 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 and consists of a series of on-device topic summaries. Each on-device topic summary corresponds to an embedding vector, which contains rich semantic information. The processing begins by extracting the pseudo-continuous trajectory from the resonance buffer. Specifically, this involves obtaining all on-device topic summary nodes contained in the pseudo-continuous trajectory and their corresponding embedding vectors. For example, suppose a pseudo-continuous trajectory contains three on-device topic summaries: topic summary A, topic summary C, and topic summary E. The corresponding embedding vectors are vector A, vector C, and vector E, respectively. After extraction, these embedding vectors are fed into the tag inference engine in the order they appear in the pseudo-continuous trajectory. Upon receiving these vectors, the tag inference engine performs further calculations and analysis to identify the semantic relationships between the embedding vectors and the trends in interest evolution. In this way, the system can fully understand the dynamic evolution of user interests, providing reliable data support for the subsequent generation of accurate interest tags. The semantic information of the embedded vector is fully utilized to ensure the accuracy and comprehensiveness of the interest tag inference process.

[0048] S4.2, hybrid attention parsing algorithm; To extract labels that closely match user interests from pseudo-continuous trajectories, a hybrid attention parsing algorithm is used to deeply mine the semantic associations between on-device topic summaries and their matching relationships with the creative context identifiers. The hybrid attention mechanism consists of two components: self-attention and cross-attention, which are used to handle dependencies within the pseudo-continuous trajectories and the association between user interests and ad content, respectively. The specific calculation process is as follows. First, for each on-device topic summary in the pseudo-continuous trajectories, an attention score is calculated between it and other on-device topic summaries. The attention score is calculated by taking the dot product of the two embedding vectors, which reflects the similarity between the vectors; then normalizing the dot product by dividing it by the product of the two vector moduli; and finally dividing it by a preset temperature parameter, which smoothes the distribution of the attention scores to avoid excessive concentration or dispersion. Based on these attention scores, a context vector is generated for each on-device topic summary by weighted summing the embedding vectors of all other on-device topic summaries according to their corresponding attention scores to obtain a comprehensive representation vector. Next, the cross-attention mechanism is used to associate the generated context vector with the embedding vector of the creative context identifier. The calculation method is similar. First, the attention score between the context vector and the embedding vector of each creative context identifier is calculated, that is, the score is obtained through dot product, modulus normalization and temperature parameter adjustment; then, based on these attention scores, the embedding vectors of the creative context identifier are weighted and summed to generate a mixed representation vector. Finally, the mixed representation vector is input into a multi-layer perceptron, which is processed through a multi-layer neural network and outputs a probability vector. Each component of the probability vector corresponds to the possibility of a label category. The system selects the label category corresponding to the component with the highest value in the probability vector as the refined interest label for the end-side topic summary. This calculation method can effectively capture the semantic dependencies within the pseudo-continuous trajectory and the deep association between user interests and advertising content, thereby significantly improving the accuracy of label classification and the relevance of advertising delivery.

[0049] Among them, the multi-layer perceptron (MLP) is a feedforward neural network with a clear and structured construction and processing process, which mainly includes steps such as network architecture design, forward propagation, loss calculation, backpropagation and parameter optimization. The following is its complete description: 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 this data through fully connected layers and nonlinear activation functions (such as ReLU) to enhance the model's expressiveness, and the output layer generates results based on task requirements (such as the probability distribution of label categories, using the Softmax function). For example, a typical MLP might consist of a 768-dimensional input layer, two hidden layers with 256 and 128 neurons, and an output layer with the same number of label categories.

[0050] Forward propagation is the core step in data processing: the input data is linearly transformed by the weight matrix and bias vector of each layer, then calculated by the activation function, and finally the result is output.

[0051] During training, a loss function (such as cross-entropy loss) measures the difference between the prediction and the true value. Backpropagation adjusts the weights by calculating gradients, and an optimization algorithm (such as Adam) updates the parameters. This process is repeated over multiple rounds until model performance improves. During inference, only forward propagation is required to generate predictions from the input.

[0052] Through the nonlinear transformation of multi-layer neural networks, multi-layer perceptrons can effectively process complex data and complete tasks such as classification.

[0053] S4.3, write to the tag tree mapping library; After generating fine-grained interest tags, to ensure efficient management and rapid access by the system, they 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 store fine-grained interest tags and their mappings to pseudo-continuous trajectories. The specific operation is as follows: For each pseudo-continuous trajectories, its unique identifier is used as the key, and the generated fine-grained interest tag is used as the corresponding value. These tags are recorded in the tag tree mapping library as key-value pairs. For example, if the unique identifier of a pseudo-continuous trajectories is "PCT1," and the fine-grained interest tag generated through the above steps is "football," then a record is stored in the tag tree mapping library, indicating that "PCT1" corresponds to "football." This key-value storage method allows the system to quickly retrieve the corresponding fine-grained interest tag based on the pseudo-continuous trajectories' unique identifiers in subsequent operations. The benefit of this storage method is that it enables efficient management and rapid access to tag data, providing timely data support for real-time advertising and bidding decisions, thereby improving the overall responsiveness and delivery efficiency of the system.

[0054] The topic archiving module to the tag precision recommendation module has successfully converted the user's coarse-grained interest context into high-precision label data and stored it in the label tree mapping library by receiving the end-side topic summary, constructing a progressive resonance graph, merging the topic path and generating fine-grained interest labels. These labels are constructed by pseudo-continuous trajectories and derived by the hybrid attention parsing algorithm, overcoming the information gap problem caused by privacy encapsulation. However, in order to achieve the accuracy and efficiency of advertising delivery, the latest data in the label tree mapping library needs to be connected to the bidding decision engine to ensure that the bidding decision engine can make real-time decisions based on the latest and accurate user interest data. Therefore, the mapping and distribution module focuses on the complete process of data synchronization and accuracy evaluation to support the optimization of subsequent bidding strategies.

[0055] The mapping delivery module is designed to synchronize the latest entries in the label tree mapping library to the bidding decision engine by index and attach precision tags. The following is a detailed technical logic breakdown: S5.1, synchronize the latest entry by index; In the ad 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 ensure the bidding decision engine can utilize the latest user interest information, the most recently generated entries in the tag tree mapping library must be extracted and synchronized. The specific processing logic is as follows: First, the system accesses the tag tree mapping library and filters 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 pseudo-continuous trajectory's unique identifier as an index to transfer and store the corresponding refined interest tag in the bidding decision engine's data cache. For example, suppose there is a record in the tag tree mapping library where the pseudo-continuous trajectory's unique identifier is a specific string and the corresponding refined interest tag is the word "football," representing the user's interest. The system then associates "football" with the pseudo-continuous trajectory's unique identifier and stores it in the bidding decision engine's data cache. This synchronization process is performed periodically or triggered when user behavior is updated to ensure data remains up-to-date. This allows the bidding decision engine to quickly retrieve the user's current interest information when processing ad placement requests, enabling ad placement that matches the user's real-time interests, improving delivery effectiveness and user experience.

[0056] S5.2, additional precision markings; Based on the synchronization of refined interest tags, a precision flag is attached to each tag to quantify the tag's accuracy, thereby assisting the bidding decision engine in formulating more appropriate bidding strategies. The specific processing logic is as follows: The precision flag is derived from the confidence calculation when the tag inference engine generates refined interest tags. In the tag inference engine, a multi-layer perceptron processes user behavior data and outputs a probability vector, where each component represents the likelihood of a particular interest category. The system selects the interest category corresponding to the component with the largest value in this probability vector as the refined interest tag and uses the value of that component as the precision flag. For example, suppose the multi-layer 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 refined interest tag and uses 0.85 as the precision flag. The system then appends "football" and 0.85 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 enables the bidding decision engine to adjust its decisions based on the accuracy of the label. For example, when processing high-accuracy labels, it will give priority to ads with higher matching degrees, thereby optimizing the accuracy and effectiveness of advertising delivery.

[0057] S5.3, synchronized to the bidding decision engine; After preparing the refined 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, and the corresponding refined interest tag and precision marker combination as the value, a key-value pair is stored in the bidding decision engine's data structure. For example, for a pseudo-continuous trajectory's unique identifier, assuming its corresponding refined 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 pseudo-continuous trajectory's unique identifier and the value is the combination of "football" and 0.85. Once stored, when the bidding decision engine receives an ad placement request, it uses the pseudo-continuous trajectory's unique identifier to query the corresponding key-value pair, obtain the refined interest tag and precision marker, and adjust the bidding strategy based on the precision marker value, for example, giving higher matching priority to tags with higher precision. This key-value storage method, through structured data organization, ensures that the bidding decision engine can efficiently retrieve and utilize user information, thereby achieving highly targeted and efficient decision-making in ad placement.

[0058] Among them, the operation processing logic of the bidding decision engine is as follows: when an advertising delivery request is received, the corresponding pseudo-continuous trajectory identifier is first retrieved from the data cache based on the user information in the request (such as user ID or device ID), and then the stored key-value pairs are queried using the identifier to obtain the user's refined interest tags and precision tags. Next, the bidding decision engine matches the advertising inventory based on the interest tags, filters out the relevant advertising candidate sets, and prioritizes matching advertisements with high-precision tags based on the precision tags. Subsequently, the bidding decision engine combines advertising characteristics (such as type, budget, historical click-through rate) and real-time user behavior data to calculate the expected revenue of each advertisement (comprehensive indicators such as click-through rate and conversion rate) through a machine learning model. Finally, the bidding decision engine determines the order of ad display and bid amount based on the expected revenue and advertiser bidding strategy to maximize platform revenue and user experience while meeting advertiser goals.

[0059] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0060] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0061] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0062] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0063] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A big data advertising label classification system based on AI analysis, characterized by: include: Topic archiving module: The receiving end side topic summary and delivery material context identifier are written into the topic temporary storage table in sequence and appended with the window timestamp; Context 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; Trajectory reconstruction module: First, the semantic drift curvature is calculated for each topic path in the resonance buffer, key features are extracted to evaluate the trajectory merging credibility for each topic path, and nodes are merged to construct a pseudo-continuous trajectory when the criteria are met, while at the same time, the candidate attribution labels are annotated. Label Refinement Module: This module feeds pseudo-continuous trajectories into the label inference engine, uses a hybrid attention parsing algorithm to generate refined interest labels, and then writes them into the label tree mapping library. Mapping delivery module: Synchronizes the latest entries in the label tree mapping library to the bidding decision engine by index, and adds precision tags.

2. The big data advertising label classification system based on AI analysis according to claim 1 is characterized in that: The processing content of the theme archiving module is as follows: The end-side topic summary and delivery material context identifier are received from the mobile device, and 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 the preset time window length. The end-side topic summary, delivery material context identifier and window timestamp are written as a record to the topic temporary storage table. The topic temporary table stores records in the order of receipt to preserve the timing characteristics of the data. The topic temporary storage table is maintained by clearing records with window timestamps earlier than the current time minus the preset time period to ensure data freshness and storage efficiency.

3. The big data advertising label classification system based on AI analysis according to claim 2 is characterized in that: The processing content of the vein weaving module is as follows: Read records containing the end-side topic summary, delivery material context identifier, and window timestamp from the topic temporary table; use the pre-trained semantic embedding model to convert the end-side topic summary into a vector representation; calculate the cosine similarity between the end-side topic summary vectors in adjacent time windows; filter out consecutive topic pairs based on a preset similarity threshold; construct a progressive resonance graph, in which the nodes are end-side topic summaries, the directed edges represent consecutive topic pairs, and the edge attributes include cosine similarity and time difference; store the progressive resonance graph in the resonance buffer in the form of an adjacency list.

4. The big data advertising label classification system based on AI analysis according to claim 3 is characterized in that: The processing content of the trajectory reconstruction module is as follows: The topic paths in the progressive resonance graph are obtained from the resonance buffer. The semantic drift curvature is calculated for each topic path. The semantic drift curvature is determined by measuring the change in the direction of the embedding vector between adjacent nodes. The curvature amplitude index and path stability potential are extracted as key features. The curvature amplitude index measures the fluctuation amplitude of semantic drift by calculating the interquartile range of the opening angle, and 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 index and path stability potential, and output the merging credibility coefficient. When the merging credibility coefficient is higher than the preset threshold, the path nodes are merged to construct a pseudo-continuous trajectory, and attribution label candidates are generated for the pseudo-continuous trajectory.

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

6. The big data advertising label classification system based on AI analysis according to claim 5 is characterized in that: The trajectory reconstruction module also includes the following processing content: The path stability potential measures the overall stability of the path by comprehensively evaluating the stability of topic similarity and time jump; The calculation process includes: first, using the sliding window technology to calculate the local topic similarity section by section on the topic path, and then taking the mean of the similarities in all sliding windows to reflect the semantic consistency of the path; second, analyzing the stability of each time jump in the path, that is, the time interval between nodes, by calculating the inverse of the time jump variance. 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 time jump stability.

7. The big data advertising label classification system based on AI analysis according to claim 6 is characterized in that: The processing content of the tag recommendation module is as follows: Pseudo-continuous trajectories, including the end-side topic summary and its embedding vector, are extracted from the resonance buffer and input into the label inference engine. The pseudo-continuous trajectories are processed using a hybrid attention parsing algorithm. The semantic dependencies between the end-side topic summaries are calculated through the self-attention mechanism to generate a context vector. The context vector is then associated with the embedding vector of the creative context identifier through the cross-attention mechanism to generate a hybrid representation vector. The hybrid representation vector is input into a multi-layer perceptron to generate a probability vector, and the component with the highest probability is selected as the refined interest tag. The refined interest tag is associated with the unique identifier of the pseudo-continuous trajectories and stored in the label tree mapping library as a key-value pair.

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

9. The big data advertising label classification system based on AI analysis according to claim 8 is characterized in that: The processing content of the mapping delivery module is as follows: Extract the latest refined interest tag and its corresponding pseudo-continuous trajectory unique identifier from the tag tree mapping library, and transmit the refined interest tag to the data cache of the bidding decision engine using the pseudo-continuous trajectory unique identifier as an index; Based on the label probability vector generated by the label inference engine, the maximum component value is selected as the precision mark, and the precision mark is attached together with the fine interest tag to the data associated with the pseudo-continuous trajectory unique identifier; with the pseudo-continuous trajectory unique identifier as the key, the fine interest tag and the precision mark are combined as the value to form a key-value pair stored in the data structure of the bidding decision engine.

Images