Intelligent data labeling task distributed storage method with edge node cooperation

Abstract

The present application relates to the field of data storage, especially to an intelligent data labeling task distributed storage method based on edge node cooperation, and the present application proposes the following scheme: candidate similar sub-trees formed on the edge node side are converted into event clues independent of node identifiers, the event clues are distributed in different audit areas, restricted retrieval and attribution verification are completed in local propagation trees of each edge node by using anchor point routing, and event attribution information and review conclusion labels are determined by combining a cross-area weighted voting mechanism.The present application realizes cross-node consistent decision without requiring consistent propagation tree structures, reduces retrieval and synchronization overhead in a hot spot scenario, and improves the stability and reliability of hot spot data review and attribution determination in a distributed storage environment.

Description

Technical Field

[0001] This invention relates to the field of data storage, and more particularly to a distributed storage method for intelligent data annotation tasks involving edge node collaboration. Background Technology

[0002] With the popularization of short videos, live stream clips, and image-text montages, the dissemination of trending content on online platforms exhibits characteristics of high concurrency, high similarity, and high variation. The same trending event may be repeatedly uploaded by a large number of users in different formats within a short period, with only minor differences in editing methods, playback speed, aspect ratio, subtitle overlay, or audio track processing. Existing technologies typically require similarity identification, event attribution determination, and storage processing of related content after upload to meet the needs of review, compliance, and subsequent retrieval.

[0003] In practical applications, existing technologies often struggle to strike a balance between timeliness and accuracy when handling trending content. On one hand, as the volume of content grows rapidly, the response speed of similarity identification and review processing tends to decrease, causing some content to enter the dissemination stage without proper judgment, thus affecting the overall control effect. On the other hand, in pursuit of processing efficiency, issues such as misjudgment of similar content, unstable event attribution, or inconsistencies in review results can easily arise, leading to the same trending topic being repeatedly identified or processed at different times or at different processing stages.

[0004] Furthermore, existing technologies typically employ relatively simple or static storage strategies at the content storage level, lacking fine-grained differentiation for content at different stages of dissemination, varying degrees of similarity, or different risk levels. This can easily lead to excessive storage resource consumption or insufficient retention of effective information. During the lifecycle of trending topics, some representative content fails to be stably retained, while a large amount of highly repetitive or low-value content continues to occupy storage space, affecting overall resource utilization efficiency.

[0005] To address the above issues, this application proposes a distributed storage method for intelligent data annotation tasks with edge node collaboration. Summary of the Invention

[0006] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a distributed storage method for intelligent data annotation tasks with edge node collaboration. This method transforms candidate similar subtrees formed at edge nodes into event clues independent of node identifiers, distributes these event clues across different review areas, and utilizes anchor point routing to complete restricted retrieval and attribution verification within the local propagation tree of each edge node. A cross-regional weighted voting mechanism is then used to determine event attribution information and review conclusion labels. This application achieves cross-node consistent adjudication without requiring consistent propagation tree structures, reducing retrieval and synchronization overhead in hotspot scenarios, and improving the stability and reliability of hotspot data review and attribution determination in distributed storage environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A distributed storage method for intelligent data annotation tasks with edge node collaboration is applied to the edge nodes of a hotspot review platform. The hotspot review platform also includes multiple review nodes, each corresponding to a review area composed of multiple edge nodes. The method includes: For the uploaded content stream slices, generate slice metadata, and write the slice metadata to the local temporary storage area; Based on the preset propagation tree, the fragment metadata is searched for similarity, and candidate similar subtrees are output. The candidate similar subtrees are then forwarded to the corresponding review nodes so that the review nodes can review the candidate similar subtrees and output the review results. The system receives the review result sent by the corresponding review node, determines the storage method of the shard metadata based on the review result, and performs storage processing on the shard metadata according to the storage method, wherein the storage method includes writing the shard metadata into the local persistent storage area and updating the propagation tree.

[0008] The generation of fragment metadata includes: The content stream is sliced ​​to determine the time range and frame range of each slice; For each slice, content features are extracted, including video fingerprint, audio fingerprint, keyframe hash, and vector features; Obtain the basic attribute information corresponding to the slice, which includes at least the original content identifier, the upload account identifier, the generation timestamp, the edge node identifier, and the region identifier to which the edge node belongs; All content features and basic attribute information of all slices are encapsulated into the slice metadata.

[0009] The step of performing similarity retrieval on the fragmented metadata based on a preset propagation tree and outputting candidate similar subtrees includes: Based on the nodes of the main path of the propagation tree and the content characteristics of the fragmented metadata, it is determined whether it is a new propagation event. If so, the fragmented metadata is converted into a main path node and attached to the end of the main path, and the main path node is output as the candidate similar subtree. If not, the fragmented metadata is converted into branch path nodes and attached to the end of the corresponding trunk path node, and the corresponding trunk path node and its contained branch path nodes are output as the candidate similar subtrees. The main path node includes an event identifier, representative features, and propagation parameters. The propagation parameters are used to determine the propagation number of the fragment metadata, with an initial value of 1. The branch path node includes a parent node identifier, a similarity score, and alignment information with the content corresponding to the parent node. It is configured to perform pruning, merging, and summary storage when preset conditions are met.

[0010] The process of converting the fragmented metadata into branch path nodes and attaching them to the end of the corresponding trunk path node includes: Based on the content features of the fragmented metadata, the representative features of each trunk path node of the propagation tree are matched to determine the trunk path node with the highest similarity to the fragmented metadata as the corresponding trunk path node. Obtain the event identifier of the corresponding trunk path node, and set the event identifier as the parent node identifier of the shard metadata to obtain the branch path node; Calculate the similarity score between the fragment metadata and the corresponding content of the corresponding trunk path node, and write the similarity score into the branch path node; Calculate alignment information and write the alignment information into the branch path node, wherein the alignment information includes time offset, start and end positions of the matching segment, keyframe matching relationship and audio fingerprint matching interval; Insert the branch path node at the end of the branch path rooted at the corresponding trunk path node; Increment the propagation count of the corresponding trunk path node by one, and update the propagation parameters of the corresponding trunk path node based on the propagation count.

[0011] The review result includes a review conclusion label, an event identifier corresponding to the review conclusion label, a review similarity score, and review alignment information; as well as storage strategy parameters corresponding to the review conclusion label, the storage strategy parameters including a storage method identifier and a lifecycle parameter.

[0012] Based on the verification results, the storage method of the fragment metadata is determined, and the fragment metadata is stored and processed according to the storage method, including: Parse the review results to obtain the review conclusion label and the storage strategy parameters corresponding to the review conclusion label; The storage method of the fragment metadata is determined based on the review conclusion label, and the storage method includes at least one of persistent storage, digest storage, frozen evidence storage, temporary retention and deletion. When the storage method is persistent storage, the shard metadata is written to the local persistent storage area and the propagation tree is updated. Updating the propagation tree includes updating the node association relationship of the propagation tree according to the verification alignment information and accumulating the corresponding propagation count.

[0013] A distributed storage method for intelligent data annotation tasks with edge node collaboration is applied to the review nodes of a hotspot review platform. The hotspot review platform also includes multiple edge nodes, each corresponding to a review area composed of multiple edge nodes. The method includes: Receive candidate similar subtrees from edge nodes; Event clues are extracted based on the candidate similar subtrees, wherein the event clues include representative features corresponding to the root node of the candidate similar subtrees and content features and time information of the fragment metadata contained in the candidate similar subtrees; The target audit node for voting is determined, and the event clues are distributed to each target edge node in the audit area corresponding to the target audit node, so that each target edge node performs anchor point routing retrieval based on its own maintained propagation tree and returns the voting results; The system receives the voting summary results returned by each target review node, performs weighted voting based on the voting summary results, determines the review conclusion label and event attribution information corresponding to the candidate similar subtree, generates a review result based on the review conclusion label and event attribution information, and sends the review result to the edge node that initiated the candidate similar subtree, so that the edge node determines the storage method of the shard metadata and performs storage processing according to the review result.

[0014] Extracting event clues based on the candidate similar subtrees includes: Obtain the root node of the candidate similar subtree, and extract the representative feature from the root node; If child nodes exist, obtain the end node associated with the root node from the candidate similar subtree, and extract content features from the fragment metadata of the end node; Obtain the time information corresponding to the shard metadata of the end node, the time information including the generation timestamp and the shard time window; The content features are normalized to generate a cue feature vector, wherein the normalization process includes scale normalization, quantization encoding, and hash mapping; Based on the representative features of the root node and the clue feature vector, a fused clue feature is generated, and the fused clue feature is encapsulated with the time information to obtain an event clue.

[0015] The target edge nodes perform anchor point route retrieval based on their respective maintained propagation trees and return voting results, including: A local propagation tree is obtained at the target edge node, and at least one anchor node is determined from the local propagation tree. The anchor node is a trunk path node that meets preset anchor conditions, wherein the preset anchor conditions include the number of propagations being greater than or equal to a first threshold and the stability of the representative feature being greater than or equal to a second threshold. For each anchor node, anchor features are obtained, wherein the anchor features include representative features of the anchor node and subtree digest signature of the subtree corresponding to the anchor node. Based on the event clues and the anchor features of each anchor node, the target anchor node with the highest matching score and a matching score greater than or equal to a preset score threshold is determined. If it does not exist, the voting result is not returned. Based on the event clues, similarity searches are performed on branch path nodes within the subtree of the target anchor node to obtain candidate attribution information; A voting result is generated based on the candidate attribution information and sent to the review node. The voting result includes the candidate attribution information and the confidence score corresponding to the candidate attribution information.

[0016] Based on the aggregated voting results, a weighted vote is performed to determine the review conclusion label and event attribution information corresponding to the candidate similar subtrees, including: Obtain the voting summary results returned by each target review node, and extract the corresponding candidate attribution information and confidence score from the voting summary results; Based on the confidence score, voting weights are assigned to each voting result, and the weights of voting results with the same candidate attribution information are accumulated. The candidate attribution information with the highest cumulative weight value is determined as the event attribution information, and the review result is determined based on the event attribution information.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a propagation tree at the edge nodes and distinguishes the propagation structure by using trunk paths and branch paths. This allows the similarity identification of hot content to be based on stable representative nodes, effectively reducing the interference of local variants on the judgment results, thereby improving the consistency and convergence of event attribution. By extracting candidate similar subtrees and generating transferable event clues, a cross-edge node voting review mechanism is introduced to achieve collaborative adjudication in a distributed environment. At the same time, the review results are directly associated with multi-level storage strategies, enabling fragmented metadata to match differentiated storage methods at different propagation stages. This ensures the effective retention of representative content while reducing the long-term occupation of storage resources by highly repetitive content. Attached Figure Description

[0018] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 An exemplary application scenario diagram provided for an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the hotspot review platform provided in the embodiments of this application; Figure 3 A schematic diagram of the processor module for an edge node provided in an embodiment of this application; Figure 4 A flowchart illustrating the intelligent distributed data storage method based on edge computing provided in this application embodiment; Figure 5 This is a schematic diagram of the propagation tree structure provided in an embodiment of this application; Figure 6 This is a schematic diagram of a propagation tree update structure provided in an embodiment of this application; Figure 7 This is a flowchart illustrating another intelligent distributed data storage method based on edge computing provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] The solutions described in this application target businesses that are typically operating within a typical cycle of explosive growth, rapid spread, and frequent secondary creation and plagiarism of trending content. In content formats such as short videos, live stream clips, and mixed text and image edits, the same trending topic often exhibits significant temporal and structural differences across different regions and social circles: some content first appears in its original form, followed by numerous derivative versions through template-based editing, partial splicing, speed and pitch changes, adding stickers and filters, and subtitle overlays; simultaneously, for the purpose of copyright infringement or traffic arbitrage, content is repeatedly uploaded to multiple access areas with slight variations but the same intent. The common characteristic of these scenarios is that the amount of content grows exponentially in a short period, and the review and storage chain needs to complete a closed loop of similarity identification, event attribution, risk assessment, and retention strategy within an extremely short window; otherwise, even if the cloud side can retrospectively trace the content, the optimal interception opportunity will be missed, and the spread will already be a fait accompli.

[0022] It's important to note that existing implementations generally employ either centralized retrieval or static distributed indexing. The centralized approach tends to centrally upload sliced ​​features or vectors to the cloud for unified retrieval and verification. Its advantages include simple logic and unified indexing, but during peak periods, bandwidth consumption and computing power queuing often occur, compromising the timeliness of upload-to-check. More importantly, centralized retrieval, to cover multimodal and multi-strategy comparisons, typically requires retaining higher-dimensional features and longer time windows for indexing, leading to significant storage and computational costs. The static distributed approach pre-allocates index shards or replicas across multiple edge nodes. However, hotspot propagation exhibits strong non-uniformity and suddenness, making static strategies prone to structural waste such as insufficient indexing in hot areas and idle resources in less popular areas. When content propagates across regions, the local index evolution of different edge nodes for the same event is not synchronized, easily leading to problems such as duplicate event creation, misattribution, or verification discrepancies.

[0023] For example, existing short video platforms or public content browsing platforms typically set a review window after users publish content, during which basic compliance verification and risk identification are completed. In the case of trending content, this review window is often highly coupled with the speed of content dissemination. For instance, when a trending event is initially published by a small number of accounts, the same content is quickly cut, remixed, and republished in different regions and access nodes due to the combined effects of platform recommendation mechanisms and user forwarding behavior, forming a massive stream of similar content. At this time, the platform needs to quickly determine whether a large number of newly uploaded contents belong to the same dissemination event, whether they constitute duplicate copying or infringement derivatives, and decide accordingly whether to delay publication, restrict recommendations, trigger evidence collection, or implement long-term retention.

[0024] In practice, the aforementioned judgments are often not made at a single node, but rather distributed across multiple edge access areas. For example, the original version of a popular piece of content may first be uploaded and initially indexed at one regional node, while subsequent derivative versions are uploaded by users at multiple regional nodes. Due to differences in access time, caching strategies, index pruning, and resources among regional nodes, the content organization structure formed by the same propagation event within different edge nodes is inconsistent: some nodes have formed relatively complete propagation backbone and branch structures, some nodes only retain summary index information, and some nodes have not yet established a clear event attribution relationship. Under these circumstances, relying solely on single-point centralized retrieval or static distributed indexes makes it difficult to provide stable, consistent, and interpretable attribution judgments within the review window.

[0025] The core concept of this application does not require different edge nodes to share the same propagation judgment mode, nor does it rely on pre-fixed event partitioning or index sharding rules. Instead, it positions the review node as the control plane for event clue orchestration and consistency adjudication: when any edge node completes sharding and feature extraction of the uploaded content and outputs the propagation event, the review node extracts event clues that can be propagated across nodes from the propagation event, so that the event clues can be distributed to other edge nodes and retrieved and verified locally.

[0026] Understandably, in the specific implementation scenario of this application, the review results generated by the review nodes are not only used to provide qualitative conclusions such as whether it is the same event, whether it is suspected of plagiarism, or whether it infringes on rights, but also directly carry the execution strategy of distributed storage, enabling edge nodes to take storage actions that match the risk level during the lifecycle of a hot topic. For example, when the voting results converge consistently across multiple edge nodes and point to the same event, the propagation count of the corresponding backbone path and the index entry update can be triggered to improve the recall efficiency of subsequent nearby searches; when the confidence level is insufficient or the conflict is large, a stricter review strategy or an extended temporary storage window can be triggered; when the conclusion is high-risk plagiarism or infringement, a freeze evidence collection and a higher level of persistence strategy can be triggered; when the hot topic recedes or the branch path nodes reach the pruning conditions, merging and summary storage can be triggered, thereby reducing the consumption of edge resources without sacrificing the event traceability capability.

[0027] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application. Figure 1 This demonstrates the overall operating environment in a scenario where trending content is disseminated, with content accessed from end devices via social platforms and processed by a trending content moderation platform.

[0028] like Figure 1As shown, the edge devices can include, but are not limited to, mobile terminals, camera devices, or other user terminals with content collection and uploading capabilities. After generating content through the edge devices, users upload the content to social media platforms for publication or dissemination. The content can be short videos, live stream clips, mixed text and image content, or other multimedia formats. During trending events, the content is often repeatedly generated and uploaded by a large number of users in different formats within a short period. The social media platform receives content from the edge devices and performs basic content access, distribution, and recommendation processes. A trending topic review platform is connected to the social media platform and participates in cross-node review and collaborative processing during the peak of trending content.

[0029] It is understandable that the hot topic review platform does not directly participate in the content collection on the client side or the basic distribution logic of the social platform. Instead, it acts as an independent collaborative processing unit, coordinating multiple edge nodes to complete consistent judgments and policy distribution in response to the rapid spread of hot topics.

[0030] It should be noted that, Figure 1 The application scenarios shown are merely illustrative examples. This application does not limit the specific implementation form or deployment method of the terminal device, social platform, or hot topic review platform. This application only improves the storage and collaborative processing mechanism of hot topic data in the content dissemination process and does not include functional limitations on content generation, content distribution, recommendation strategies, or user interaction processes.

[0031] In some optional implementations, the edge devices, social platforms, and hot topic review platforms can be deployed by the same business entity or by different business entities. The solution described in this application only focuses on how to collaboratively complete the identification, attribution judgment, and storage strategy control of hot topic data among multiple edge nodes and review nodes in the scenario of rapid dissemination of hot topic content, without imposing restrictions on specific business forms, product forms, or platform architectures.

[0032] refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of the hotspot review platform provided in the embodiments of this application.

[0033] like Figure 2 As shown, the hot topic moderation platform in this application includes multiple moderation areas, each corresponding to a set of logically related node resources. For example, each moderation area includes at least one moderation node and several edge nodes, wherein the edge nodes are used to access content data from social media platforms and perform local processing operations related to hot topic data, and the moderation nodes are used to aggregate, analyze, and collaboratively process the hot topic data from the edge nodes within their respective moderation areas.

[0034] In an exemplary implementation, the division of audit areas can be determined based on geographical location, network topology, service load, or platform policies. For example, a group of edge nodes with similar geographical locations or low network latency can be assigned to the same audit area, with the corresponding audit node responsible for coordination and management within that audit area. It should be understood that this application does not limit the specific rules for dividing audit areas; any method that supports collaborative interaction between edge nodes and audit nodes can be considered an optional implementation of this application.

[0035] Figure 2 The document further illustrates that the hotspot review platform also includes a cloud platform. The cloud platform can be understood as a cloud server or cloud computing resource pool, used to provide cross-regional interaction and support capabilities for multiple review nodes. In an exemplary implementation, the cloud platform can be used to handle operations such as status synchronization, result forwarding, or historical data archiving between review nodes. However, this application does not limit the cloud platform to participating in the specific hotspot data judgment process; the cloud platform's participation method can be flexibly configured according to system scale and business needs.

[0036] exist Figure 2 In a specific embodiment not shown, each edge node may also be configured with a processor and a storage unit. The processor is used to perform content slicing, feature extraction, initial similarity screening, and local computing tasks related to hot data. The storage unit is used to store hot data indexes, temporary data, or summary data related to the edge node.

[0037] It should be noted that, Figure 2 The structure shown is merely an exemplary structure used to illustrate the overall composition of the hotspot review platform in this application. In other embodiments, the connection relationships and deployment numbers between review nodes, edge nodes, and the cloud can be adjusted according to the actual application scenario. This application does not limit the specific number of nodes, hierarchical relationships, or hardware configurations.

[0038] refer to Figure 3 , Figure 3 This is a schematic diagram of the processor module for an edge node provided in an embodiment of this application.

[0039] The content generation module is used to slice the incoming content stream and generate slice metadata; wherein, the slice metadata includes at least a content identifier, a slice time window, a generation timestamp, and a content feature carrying field for subsequent retrieval.

[0040] The propagation tree maintenance module is used to perform similarity retrieval in the local propagation tree based on the content features of the fragment metadata, and update and maintain the local propagation tree according to the retrieval results. For example, the propagation tree maintenance module is used to convert the fragment metadata into trunk path nodes or branch path nodes and complete the attachment, while updating parameters such as propagation number, representative feature stability and subtree summary signature.

[0041] The subtree generation module is used to generate candidate similar subtrees based on the local propagation tree and send the candidate similar subtrees to the corresponding review nodes; it is also used to receive the review results returned by the review nodes to trigger the execution of the local storage policy and the update of the propagation tree.

[0042] The routing voting module is used to receive event clues issued by the review node, perform anchor point routing retrieval in the local propagation tree, generate voting results, and return them to the review node. Specifically, the routing voting module is used to determine anchor nodes that meet preset anchor point conditions from the local propagation tree, match event clues with anchor point features to locate target anchor nodes, and perform similarity retrieval within the subtree of the target anchor node to obtain candidate attribution information, thereby generating voting results containing candidate attribution information and confidence scores.

[0043] Next, with reference to the accompanying drawings, the intelligent distributed data storage method based on edge computing provided in the embodiments of this application will be further elaborated. Figure 4 The method shown is applied to the edge nodes of a hotspot review platform, which further includes multiple review nodes, each corresponding to a review area composed of multiple edge nodes. The method includes: S1: Slice the uploaded content stream, generate fragment metadata, and write the fragment metadata into the local temporary storage area; In this embodiment, the uploaded content stream typically originates from multimedia content uploaded by different users at different times during the spread of trending events. This type of content exhibits high uncertainty in its generation method, editing level, and dissemination pace. By slicing the content stream and generating fragment metadata, continuous content can be transformed into the smallest independently processable analytical unit without relying on complete content semantic parsing. Fragment metadata carries the basic information required for subsequent similarity retrieval, propagation relationship determination, and storage decisions. Its specific field format can be flexibly configured according to the operating environment, as long as it can support propagation tree construction and feature comparison; this application does not impose further limitations.

[0044] Writing the fragment metadata to the local temporary storage area first helps to buffer and manage the data before the audit results are returned, thereby avoiding the concentration of storage pressure caused by direct persistent writing during peak periods. It also provides room for operation of subsequent differentiated storage strategies based on the review results, enabling edge nodes to maintain a balance between response speed and resource consumption.

[0045] It is understandable that the local temporary storage area is used for transitional storage and management of the fragment metadata before event attribution and storage strategy decisions are completed. Logically, the local temporary storage area may include different storage areas such as a cache area, a review area, or a priority queue area to distinguish fragment metadata at different processing stages. For example, initially generated fragment metadata can be placed in the review area, and after receiving the corresponding review results, it can be migrated to the persistent storage area or cleaned up depending on the storage method.

[0046] The specific implementation can be understood as a temporary data management space built on memory, disk, or a combination of both. Its capacity, retention period, and scheduling strategy can be dynamically adjusted according to the operating load and resource status of the edge nodes. Those skilled in the art will understand that the organization of the local temporary storage area does not affect the core idea of ​​this application's solution, as long as the traceability and schedulability of the fragment metadata can be guaranteed to a minimum before the review results are returned. This application does not impose further limitations in this regard.

[0047] S2: Perform similarity retrieval on the fragmented metadata according to the preset propagation tree, output candidate similar subtrees, and forward the candidate similar subtrees to the corresponding review nodes so that the review nodes can review the candidate similar subtrees and output the review results; In this embodiment, the propagation tree describes the evolutionary structure of hot content on the edge nodes. It is not a predefined fixed template but rather a dynamically constructed result as content is uploaded and propagated. Similarity retrieval based on the propagation tree can maintain the continuity of event context while avoiding the risk of misjudgment caused by relying solely on flat similarity matching. By locating a set of nodes with high relevance to the current fragment metadata within the propagation tree and organizing this local structure into candidate similarity subtrees, the output of similarity retrieval can be elevated from a few similar entries to structured candidates with propagation semantics, thus providing more sufficient basis for subsequent review. The candidate similarity subtrees are forwarded to the corresponding review nodes to introduce cross-node collaborative analysis capabilities based on local edge node judgments. Review nodes review based on event clues and voting results from different edge nodes, effectively mitigating the misattribution problem caused by limited perspective, index pruning, or resource constraints of a single edge node, thereby improving the consistency and reliability of hot event attribution judgments.

[0048] S3: Receive the review result sent by the corresponding review node, determine the storage method of the shard metadata according to the review result, and perform storage processing on the shard metadata according to the storage method, wherein the storage method includes writing the shard metadata into the local persistent storage area and updating the propagation tree; In this embodiment, the review results are not only used to provide conclusions on event attribution or risk assessment, but also directly serve as the basis for driving the selection of storage strategies. By determining the storage method of fragmented metadata based on the review results, hierarchical management of hot data can be achieved. For example, high-confidence, representative content can be persistently stored and included in the main path of the propagation tree to support subsequent similarity retrieval and propagation statistics; while content with low confidence or short lifecycle can be temporarily stored, summarized, or delayed, thereby reducing unnecessary resource consumption. The propagation tree is updated simultaneously with the storage processing, ensuring that the local propagation structure of edge nodes can promptly reflect the post-review event evolution status, preventing the propagation tree from becoming disconnected from the actual propagation situation.

[0049] Next, we will further elaborate on the technical aspects of the fragmented metadata method in this application.

[0050] In one example, generating shard metadata includes: The content stream is sliced ​​to determine the time range and frame range of each slice; For each slice, content features are extracted, including video fingerprint, audio fingerprint, keyframe hash, and vector features; Obtain the basic attribute information corresponding to the slice, which includes at least the original content identifier, the upload account identifier, the generation timestamp, the edge node identifier, and the region identifier to which the edge node belongs; All content features and basic attribute information of all slices are encapsulated into the slice metadata.

[0051] Specifically, content streams often exhibit continuity and temporal sequence during the upload phase. Treating the entire content directly as the processing object not only increases the burden of single-time computation and storage but also amplifies redundant information when similar content undergoes only local changes. Therefore, this embodiment segments the content stream, dividing continuous content into multiple slices with clearly defined time and frame ranges, ensuring each slice corresponds to a relatively independent analysis unit. The time range describes the temporal position of the slice within the original content, while the frame range defines the visual boundaries of the slice. This dual-definition approach helps to simultaneously consider temporal continuity and spatial consistency in subsequent similarity analysis, thereby avoiding mismatches caused by relying solely on a single dimension of information.

[0052] It should be noted that video and audio fingerprints are used to capture the overall characteristics of content at the audiovisual level, and they have a high tolerance for editing, compression, or slight perturbations. Keyframe hashes are used to characterize representative visual moments in a slice, and can maintain a certain level of distinguishability even when there is local cropping or image overlap. Vector features are used to express higher-dimensional semantic or structural information, facilitating distance calculations during approximate retrieval or clustering. By extracting multiple feature types simultaneously, complementary relationships can be formed under different interference conditions, making subsequent processing independent of a single feature dimension, thereby reducing the risk of misjudgment caused by feature degradation or noise amplification. Those skilled in the art will understand that the specific feature extraction model or algorithm used can be selected according to computing power and business needs. As long as the above feature types can be stably output and the consistent mapping relationship between features can be maintained, the implementation requirements of this embodiment can be met. For example, multimodal feature extraction can be achieved by combining traditional signal processing with deep learning. For instance, in the process of video fingerprint generation, a robust fingerprint description can be constructed based on inter-frame difference, brightness distribution statistics, edge gradient features, or frequency domain features; in the process of audio fingerprint generation, a stable audio fingerprint representation can be formed based on Mel-frequency coefficients, spectral peak features, or time-frequency energy distribution; in the process of keyframe hash extraction, representative frames can be selected through keyframe detection algorithms, and hash features can be generated based on perceptual hashing, differential hashing, or structural hashing; in the process of vector feature extraction, the slice content can be encoded through convolutional neural networks, temporal modeling networks, or multimodal fusion networks to output a vector representation with a unified dimension.

[0053] Next, we will further elaborate on the technical aspects of the propagation tree method in this application.

[0054] refer to Figure 5 , Figure 5 This is a schematic diagram of the propagation tree structure provided in an embodiment of this application.

[0055] It should be noted that the propagation tree in this application can be understood as a structured representation used to depict the evolution of hot data at edge nodes over time. Its core purpose is to organize originally discrete, flat content fragments into structural units with related semantics and propagation order. The propagation tree is not used to describe actual user forwarding paths, but rather to reflect the aggregation and differentiation process of similar content in terms of time, features, and propagation behavior from the perspective of edge nodes. Those skilled in the art will understand that the introduction of the propagation tree does not rely on a complete social relationship or recommendation chain on the platform side, but is entirely based on content features and operational data and is autonomously constructed on the edge side, thus possessing strong deployment independence.

[0056] In this embodiment, each edge node is configured with a corresponding propagation tree to manage the hotspot-related shard metadata accessed and processed by that edge node. Logically, the propagation tree can be represented as a directed structure, where nodes represent shard metadata or content entities abstracted from shard metadata, and edges represent the relationships between nodes in terms of similarity and temporal evolution. For example, the data structure of the propagation tree can be a tree structure based on node and parent-child relationships, or it can be implemented as a graph structure with hierarchical identifiers and association pointers, as long as it supports the distinction between the main path and branch paths, the maintenance of node attachment relationships, and the extraction of local subtrees. This application does not impose specific limitations on this.

[0057] The propagation tree construction process is driven by the continuous arrival of fragmented metadata. Upon receiving new fragmented metadata, the edge node first performs a similarity match within the existing propagation tree based on its content characteristics to determine if the fragmented metadata belongs to an existing propagation event. If no set of nodes satisfying the similarity criteria is found in the propagation tree, the fragmented metadata is considered a potential starting point for a new propagation event and added to the propagation tree as a trunk path node to carry any subsequent derivative content. If a matching node exists, the fragmented metadata is attached to the corresponding trunk path node or branch path node as a branch path node, based on factors such as similarity score, temporal proximity, and propagation parameters, thus reflecting the derivative relationships of the content within that event.

[0058] In one example, the step of performing similarity retrieval on the fragment metadata based on a preset propagation tree and outputting candidate similar subtrees includes: S2.1: Based on the nodes of the main path of the propagation tree and the content characteristics of the fragment metadata, determine whether it is a new propagation event; Specifically, the main path nodes in the propagation tree carry representative content that has been recognized by edge nodes as the entry point or key point of evolution for propagation events. If the entire tree nodes are used as the matching object directly, a large number of branch variants are easily introduced into the comparison scope, resulting in uncontrolled retrieval computation. Furthermore, during the hotspot outbreak phase, the rapid growth of branches may cause the matching results to be pulled by locally similar variants, leading to frequent fluctuations in event attribution. Limiting the judgment object to the main path nodes ensures that the determination of whether it is a new propagation event is based on a stable set of representative content. Even when there are a large number of disturbances such as editing, splicing, occlusion, and speed and pitch changes within the same event, the consistency of the event entry point is maintained, avoiding misjudging short-term noisy variants as new events.

[0059] In this embodiment, the edge node reads the current set of backbone path nodes from the propagation tree and obtains its representative features for each backbone path node. The representative features can be directly formed from the content features of the slice corresponding to the backbone path node, or they can be formed by aggregating and updating the content features of several branch path nodes under the backbone path node. The aggregation method can be to normalize the vector features, calculate the mean and normalize again, take the intersection or stable subset of the fingerprint / hash set, etc., as long as the representative features can reflect the core features of the propagation event corresponding to the backbone path node. Subsequently, the content features of the fragment metadata to be judged are matched with the representative features of each backbone path node. The matching process preferably adopts a hierarchical comparison: first, a coarse screening is performed using video fingerprints or keyframe hashes to quickly eliminate obviously dissimilar backbone path nodes, and then the similarity scores of vector features are calculated for the small number of candidate backbone path nodes retained after coarse screening. When audio fingerprints are available, the audio matching results are used to sort or correct the confidence of candidate backbone path nodes. If the similarity scores of all candidate trunk path nodes do not reach the preset event attribution threshold, or if they reach the threshold but are incompatible with the active window of existing events in terms of time information, they are determined to be new propagation events. The active window can be determined by the timestamp of the most recent update of the trunk path node, the propagation number growth rate, or the branch growth rate. Those skilled in the art can configure the window width according to the running data to adapt to the differences in the life cycle of different hotspots.

[0060] It should be noted that video fingerprint similarity can be obtained by segmenting and matching the video fingerprint sequence of the segment metadata with the representative fingerprint sequence of the backbone path nodes. Specifically, a sliding window method can be used to calculate the Hamming distance or edit distance between fingerprint segments, and the proportion of successfully matched segments is used as the fingerprint similarity score. Keyframe hash similarity can be obtained by statistically analyzing the overlap ratio between the keyframe hash set of the segment metadata and the stable set of keyframe hashes of the backbone path nodes. This overlap ratio can be expressed as the ratio of the number of intersections to the size of the smaller set, where the smaller set size represents the smaller set size in the keyframe hash set of the segment metadata and the stable set of keyframe hashes of the backbone path nodes. Alternatively, frame position weights can be introduced, so that keyframes with similar temporal positions contribute more to the matching results. The above two types of similarity are mainly used in the coarse screening stage. They have low computational complexity, a certain tolerance for editing perturbations, and can quickly compress the number of candidate backbone path nodes while ensuring recall.

[0061] In the fine matching stage, the similarity score of vector features can be obtained by calculating the distance between the segment metadata vector features and the representative vectors of the backbone path nodes. For example, both can be normalized to the same scale first, and then the inverse mapping value of cosine similarity or Euclidean distance can be calculated to characterize the degree of proximity in the high-dimensional feature space. The inverse mapping value refers to converting the cosine similarity or Euclidean distance into a similarity score that is opposite to the distance in magnitude, i.e., the smaller the distance, the larger the value. When audio fingerprints exist, the matching coverage rate or continuous matching length ratio of audio fingerprints on the time axis can be further calculated. The matching coverage rate refers to the proportion of the time period of successful matching to the total time length of the comparison after aligning the audio fingerprints of the segment metadata and the representative audio fingerprints of the backbone path nodes on the time axis, and the result is mapped to the audio representative fingerprint matching consistency score. The final similarity score can be obtained by combining vector feature similarity and audio representative fingerprint matching consistency according to preset weights. The weight values ​​can be configured based on content type or operational experience. For example, in music content, the weight of audio representative fingerprint matching consistency can be increased, while in visually dominant content, the weight of vector feature similarity can be increased. Those skilled in the art will understand that as long as the calculation results of different feature dimensions remain comparable in numerical range, fusion can be achieved using methods such as linear weighting, piecewise mapping, or confidence correction.

[0062] Furthermore, the involvement of time information is not simply a matter of time difference filtering, but rather it, together with the similarity score, constitutes the criteria for determining event attribution. In this embodiment, the time difference between the generation timestamp of the fragment metadata and the most recent active time of the trunk path node is used to determine whether it falls within the active window range of the event corresponding to that trunk path node.

[0063] S2.2: If so, convert the fragmented metadata into a trunk path node and attach it to the end of the trunk path, and output the trunk path node as the candidate similar subtree; Specifically, when fragmented metadata is determined to be a new propagation event, the propagation tree needs to generate a stable entry point that can carry subsequent derivative content aggregation. Otherwise, when subsequent content of the same event arrives, it can only be determined to be a new event, causing fragmentation of the propagation tree entry point and affecting the unified execution of subsequent storage strategies.

[0064] In this embodiment, when fragment metadata is transformed into a trunk path node, at least three types of fields are generated: event identifier, representative feature, and propagation parameters. The event identifier is used to uniquely identify the propagation event locally on the edge node. The event identifier can be constructed from the edge node identifier, generation timestamp, and fragment content feature summary to reduce the probability of cross-event collisions. The representative feature preferably uses a combination of vector features and fingerprint / hash of the fragment metadata, and stores the feature version number or update timestamp in the trunk path node for subsequent updates, so that it can be traced when the representative feature is modified by subsequent branch content. The propagation parameters include at least the propagation count, with an initial value set to 1, and can synchronously record the first appearance time of the trunk path node, the most recent active time, and the local popularity count. The process of attaching to the end of the trunk path includes: assigning a node identifier to the trunk path node in the propagation tree structure, pointing its parent pointer to the previous trunk path node (if it exists), and updating the trunk path index table or trunk linked list, so that the trunk path node can be quickly traversed and reached through the trunk path entry point.

[0065] S2.3: If not, convert the fragmented metadata into branch path nodes and attach them to the end of the corresponding trunk path node, and output the corresponding trunk path node and its contained branch path nodes as the candidate similar subtrees; Specifically, when shard metadata shares a consistent attribution with existing propagation events, the propagation tree should incorporate it into the derived structure of that event to organize and manage different variants within the same event. Directly incorporating it into the event set without recording its correspondence with the main path nodes will result in a lack of traceable evidence chains for subsequent verification. Furthermore, if a large number of variants are retained during hotspot propagation without distinguishing between hierarchy and scalability, the propagation tree will rapidly expand and affect the stability of edge node resources.

[0066] In this embodiment, the determination of the corresponding main path node is based on the matching result of S2.1: the main path node with the highest similarity score is selected as the attachment target, and the node identifier of the main path node is written into the parent node identifier field of the branch path node. In addition to the parent node identifier, the branch path node also has a similarity score and alignment information written into it. The similarity score is preferably the final confidence score after multi-feature fusion, and the alignment information is used to describe the consistency evidence between the current segment metadata and the corresponding content of the parent node in terms of time and structure. The alignment information may include the relative offset of the slice time window, the index position of the keyframe matching pair, the segment range of the audio fingerprint matching, and the matching coverage ratio when there is image cropping or occlusion. The above alignment information can be generated by aligning the keyframe sequence, sliding matching the audio fingerprint sequence, or dynamic window comparison of the vector feature sequence. It is only necessary to provide a verifiable evidence summary for subsequent verification. The process of attaching to the end of the corresponding trunk path node includes: inserting the branch path node into the child node list or branch list of the trunk path node, and updating the propagation count of the trunk path node so that the propagation count reflects the cumulative occurrence intensity of the event at this edge node; the accumulation of the propagation count is not only used for statistics, but also for subsequent anchor point selection and representative feature stability evaluation, thereby forming a closed loop of self-maintenance of the propagation tree.

[0067] In one example, converting the shard metadata into branch path nodes and attaching them to the end of the corresponding trunk path node includes: Based on the content features of the fragmented metadata, the representative features of each trunk path node of the propagation tree are matched to determine the trunk path node with the highest similarity to the fragmented metadata as the corresponding trunk path node. Obtain the event identifier of the corresponding trunk path node, and set the event identifier as the parent node identifier of the shard metadata to obtain the branch path node; Calculate the similarity score between the fragment metadata and the corresponding content of the corresponding trunk path node, and write the similarity score into the branch path node; Calculate alignment information and write the alignment information into the branch path node, wherein the alignment information includes time offset, start and end positions of the matching segment, keyframe matching relationship and audio fingerprint matching interval; Insert the branch path node at the end of the branch path rooted at the corresponding trunk path node; Increment the propagation count of the corresponding trunk path node by one, and update the propagation parameters of the corresponding trunk path node based on the propagation count.

[0068] refer to Figure 6 , Figure 6This is a schematic diagram of a propagation tree update structure provided in an embodiment of this application.

[0069] Figure 6 by Figure 5 Based on the propagation tree structure shown, two output methods for candidate similar subtrees are illustrated under different propagation decision results.

[0070] In one example scenario, when newly arrived fragment metadata is identified as a new propagation event, the propagation tree generates a new trunk path node based on the existing trunk path. At this point, the candidate similar subtree is output with this newly generated trunk path node as its sole node. This output method corresponds to the scenario of establishing a new event entry point in the propagation tree, characterized by: a minimal candidate similar subtree structure containing only node information that can represent the representativeness of the new event. By outputting the candidate similar subtree as a single node, the event entry point can be quickly submitted to the review node for cross-node confirmation before the new event has branched and evolved, thus avoiding the introduction of too many uncertain derivative structures in the early stages of the event and reducing the complexity of subsequent review.

[0071] In another example scenario, when newly arrived fragment metadata is determined to be derived from an existing propagation event, the propagation tree generates branch path nodes under the corresponding trunk path node. In this case, the output of the candidate similar subtree uses the corresponding trunk path node as its root node and includes branch path nodes under that trunk path node that are related to the current fragment metadata. This output method preserves the parent-child relationship between trunk path nodes and branch path nodes, ensuring that the candidate similar subtree not only reflects the event attribution of the current fragment metadata but also provides the evolutionary context information of the event within the local propagation tree. This supports the review node's judgment on event consistency and the reasonableness of derivation during the review phase.

[0072] For example, the following numerical example is provided to illustrate how, in a "non-new propagation event" scenario, the computational link of matching main path nodes, branching, and outputting candidate similar subtrees from fragmented metadata is generated. This example is only used to explain the processing flow and dimensional relationships; the parameters and values ​​listed are for illustrative purposes only and do not constitute a limitation on the implementation scope.

[0073] In this example, there are three main path nodes on the local propagation tree of a certain edge node, denoted as T1, T2, and T3, with corresponding event identifiers E1, E2, and E3, respectively. Each main path node maintains representative features and active window information. The most recent active time for T1 is 10:02:10, for T2 it is 10:05:30, and for T3 it is 09:40:00. The corresponding active window width is set to 15 minutes for example. The newly arrived fragment metadata is denoted as Mx, with a fragment time window of 10:06:12-10:06:18, a generation timestamp of 10:06:19, and video fingerprint, keyframe hash, audio fingerprint, and vector features have been extracted. For ease of explanation, the event attribution thresholds are set as follows: vector feature similarity threshold 0.82, audio fingerprint consistency threshold 0.70; the specific value of any of these thresholds can be determined by the running data calibration, and this application does not impose further limitations.

[0074] When matching backbone path nodes, a coarse screening is first performed to control the scale of comparisons. The coarse screening stage uses video fingerprints and keyframe hashes as rapid filtering criteria: the video fingerprint of Mx is compared with representative fingerprints of T1, T2, and T3, resulting in three sets of coarse screening similarity scores of 0.41, 0.86, and 0.48, respectively; simultaneously, the keyframe hash set of Mx is matched with stable subsets of keyframe hashes of each backbone path node, yielding matching coverage ratios of 18%, 62%, and 21%, respectively. Based on coarse screening rules (e.g., coarse screening similarity not lower than 0.75 and keyframe coverage ratio not lower than 50%), only candidate backbone path node T2 is retained, while T1 and T3 are excluded, thus avoiding high-cost vector comparisons of a large number of irrelevant backbone path nodes.

[0075] The process then proceeds to the fine-matching stage. Vector feature similarity scores are calculated for candidate backbone path nodes T2, and confidence is adjusted using the consistency of audio representative fingerprint matching. For example, the similarity score between the vector features of Mx and the representative vector of T2 is 0.88, higher than the vector feature similarity threshold of 0.82. Simultaneously, the consistency between the audio fingerprint of Mx and the audio representative fingerprint of T2 is 0.74, higher than the audio fingerprint consistency threshold of 0.70. Regarding temporal consistency, the generation timestamp of Mx, 10:06:19, falls within the active window range of T2 (within 15 minutes from 10:05:30), satisfying the active window compatibility condition. Based on the above three types of evidence (vector feature similarity, audio representative fingerprint consistency, and temporal compatibility), Mx is determined to be a "non-new propagation event," and its associated event identifier is identified as E2, corresponding to backbone path node T2.

[0076] During the branch attachment phase, the fragment metadata Mx is converted into a branch path node Bx, and necessary fields are written to ensure subsequent verification and traceability. The parent node identifier is set to the node identifier of T2; the similarity score is written as the final confidence score after multi-feature fusion, exemplarily set to 0.90 (which can be understood as the comprehensive evaluation result after combining the vector feature similarity of 0.88 and the audio representative fingerprint matching consistency of 0.74; the specific fusion method can be implemented according to the project configuration); the alignment information is used to describe the matching position relationship between Mx and the corresponding content of the parent node, exemplarily including: time offset +3.2 seconds (indicating that the main matching area of ​​Mx and the T2 representative fragment has shifted relatively backward on the time axis), the start and end positions of the matching fragment 10:06:13–10:06:17, the number of keyframe matching pairs 7, and the audio matching interval 4.0 seconds. After writing the fields, insert the branch path node Bx into the end of the branch list of the trunk path node T2, and update the propagation count of T2: if the original propagation count of T2 is 27, then it will be 28 after the update.

[0077] During the candidate similar subtree output phase, a candidate similar subtree rooted at T2 is constructed and output. This candidate similar subtree contains at least the main path node T2 and the set of branch path nodes associated with it.

[0078] The processing flow illustrated in the above example can be summarized as follows: In the coarse screening stage, lightweight features are used to quickly narrow down the range of candidate backbone path nodes; in the fine matching stage, vector features and audio representative fingerprints are used to confirm event attribution; in the linking stage, traceable relationships are established through parent node identifiers, confidence scores, and alignment information; and in the output stage, structured evidence is provided in the form of local subtrees. The values ​​of each threshold, candidate number, and time window can be determined by calibration using runtime data, and those skilled in the art can implement the same processing flow accordingly.

[0079] Next, we will further elaborate on the technical aspects of the method in this application regarding the verification results.

[0080] In one example, the review result is not merely a simple judgment of similarity or whether the review has passed, but rather a comprehensive decision-making vehicle driving subsequent storage behavior and propagation tree update behavior of edge nodes. Therefore, the review result is designed to include multiple interrelated fields to simultaneously express the event attribution conclusion, evidence strength, and resource management instructions. Specifically, the review conclusion label characterizes the judgment result of the candidate similar subtree after cross-node review. This can include, but is not limited to, states such as confirmation of the same event, high-confidence derived event, low-confidence related event, and unrelated event, used to distinguish different degrees of event consistency. The event identifier corresponding to the review conclusion label clarifies the propagation event entry point to which the conclusion points, enabling edge nodes to accurately map shard metadata back to the corresponding trunk path node in the local propagation tree, avoiding event drift after review. The similarity score is used to quantify the overall consistency of the review stage. This score can be obtained by weighted fusion of voting results from multiple edge nodes and is used to reflect the statistical reliability of the review conclusion. The alignment information is used to describe the key correspondence between the candidate similar subtree and the target event at the time, structure or content level, and to provide an interpretable basis for subsequent auditing, backtracking or anomaly analysis.

[0081] In this embodiment, storage strategy parameters, as components of the review results, correspond one-to-one with the review conclusion labels, and are used to directly translate logical judgments into executable storage behavior instructions. The storage method identifier clarifies the storage format that fragmented metadata should adopt on the edge node side, while the lifecycle parameters limit the duration, retention conditions, or expiration rules of this storage behavior. By binding storage strategy parameters to the review conclusions, inconsistent storage decision logic adopted by edge nodes at different stages can be avoided, thus ensuring a unified strategy source for the storage evolution of hot data throughout its lifecycle.

[0082] Those skilled in the art will understand that the lifecycle parameter can be represented by time length, access count threshold, popularity decay threshold, or a combination thereof, as long as it can support subsequent automatic cleanup or policy switching. This application does not impose any further limitations.

[0083] It is understood that the storage methods described in this application are not a simple binary choice, but rather a multi-level storage approach for hot content at different stages of dissemination and risk levels. Persistent storage is used to preserve fragment metadata that has been confirmed to belong to a stable dissemination event and has representative or subsequent retrieval value. This fragment metadata is typically included in the trunk or key branch path nodes of the dissemination tree to support subsequent similarity searches and event backtracking. Summary storage is used to compress and retain branch fragments with high dissemination frequency but highly similar content. This can be achieved by retaining only feature summaries, alignment information, and statistical parameters, while deleting high-cost original features or complete slice information, thereby reducing storage usage without disrupting the semantics of the dissemination structure. Freezing for evidence storage is used to protectively preserve fragment metadata that is disputed, potentially infringing, or requires subsequent manual review. In this method, fragment metadata and its alignment evidence are locked in a designated storage area, and the lifecycle parameters are preferably controlled by compliance or business policies to prevent automatic cleanup before the investigation is completed.

[0084] In another example, the storage method of the fragment metadata is determined based on the review result, and the fragment metadata is stored according to the storage method, including: Parse the review results to obtain the review conclusion label and the storage strategy parameters corresponding to the review conclusion label; The storage method of the fragment metadata is determined based on the review conclusion label, and the storage method includes at least one of persistent storage, digest storage, frozen evidence storage, temporary retention and deletion. When the storage method is persistent storage, the shard metadata is written to the local persistent storage area and the propagation tree is updated. Updating the propagation tree includes updating the node association relationship of the propagation tree according to the verification alignment information and accumulating the corresponding propagation count.

[0085] In another example, when storage processing is completed based on the review results, the propagation tree update is configured to be executed in conjunction with the storage method to avoid the propagation tree structure becoming disconnected from the actual storage form, thus affecting subsequent retrieval and review. The propagation tree update can be understood as a combination of four types of operations: node addition, node persistence, node merging, and node rollback. Its triggering criteria include the review conclusion label, event identifier, review alignment information, and storage method identifier.

[0086] In the persistent storage mode, the propagation tree update adopts the processing logic of solidifying newly added nodes: when the review conclusion label indicates a new propagation event, the node corresponding to the currently pending shard metadata is solidified as a main path node. Specifically, this includes assigning an event identifier to the node, writing representative features and propagation parameters, and initializing the propagation count to 1. At the same time, its position on the main path is determined according to the review alignment information. For example, it can be attached to the end of the main path, and the main path index is updated so that subsequent similar searches can use the main path node as the event entry point for recall. When the verification conclusion label indicates a non-new propagation event, the current node is fixed as a branch path node at the end of the corresponding main path node. This includes writing the parent node identifier, similarity score, and verification alignment information, and inserting the branch path node at the end of the branch list or child node list of the corresponding main path node. At the same time, the propagation count of the corresponding main path node is cumulatively updated according to each arrival event, and the time order of the branch list can be maintained based on the verification alignment information, so that branch path nodes within the same time window are kept adjacent to each other, which facilitates the subsequent extraction of candidate similar subtrees according to the time window.

[0087] In the summary-based storage method, the propagation tree update employs a feature merging logic, and the verification results under this method typically indicate that the current shard metadata does not need to form an independent long-term searchable entity. To ensure the traceability of the propagation structure and the updability of the index, the summary-based storage method still performs propagation tree updates, but the update targets are existing nodes rather than newly created nodes. For example, when the review conclusion label originally indicated a new propagation event, it is not directly fixed as a new main path node. Instead, the target main path node for merging is re-determined from the set of main path nodes. Based on the review similarity score and review alignment information, the main path node with the highest matching degree with the current fragment metadata and meeting the merging threshold can be selected as the merging object from the representative features of the main path nodes. During merging, the content features of the current fragment metadata are written into the summary storage area in a summary form, and the representative features and propagation parameters of the main path node of the merging object are updated. The update method may include weighted fusion of vector features, expansion or replacement of fingerprint / hash stable subsets, and accumulation of propagation times. The weighting coefficient can be determined by the review similarity score or node stability to make the contribution of high-confidence samples to the representative features higher. When the review conclusion label originally indicates a non-new propagation event, the target branch path node for merging is determined in the branch path of the corresponding main path node: First, the candidate branch range can be limited based on the review alignment information (e.g., limited to a set of branches with similar time windows or similar alignment positions). Then, the branch path node with the highest similarity score to the current fragment metadata and that meets the merging threshold is selected as the merging object from the candidate set. During merging, no new branch path nodes are added. Instead, the summary field of the branch path node of the merging object is updated in the summary storage area. The update content includes at least similarity score statistics, alignment information summary merging, and branch occurrence count accumulation. If necessary, the summary signature corresponding to the branch path node is recalculated, so that the "representativeness" of the branch path node in the propagation tree is enhanced as more samples converge.

[0088] For the temporary retention, deletion, and frozen evidence storage methods, the propagation tree update uses the logic of deleting or rolling back newly added nodes, but the differences exist depending on the preservation requirements. Under the temporary retention method, the current fragment metadata can remain in the temporary storage area, and the corresponding newly added nodes in the propagation tree are not fixed: if a trunk or branch path node has been temporarily created for candidate similar subtree output during the similarity retrieval stage, the node is marked as pending confirmation, and minimal reference information is retained in the propagation tree, with an expiration time set; if no further convergence verification results or subsequent sample support are received within the expiration time, node rollback deletion is performed, removing the node from the parent node's child node list or branch chain. Under the deletion method, node rollback deletion and temporary storage cleanup are performed directly. The deletion action includes at least: removing parent-child associations, deleting node records, synchronously updating the parent node's branch count or index pointer, and clearing the temporary feature cache related to the node to avoid dangling references in the propagation tree. In the frozen evidence storage method, to balance evidence preservation and controllability of the propagation tree, whether a newly added node is deleted is determined by the evidence collection strategy: When the evidence collection requires maintaining a traceable link, the newly added node can remain in the propagation tree but is marked as a frozen node. Frozen nodes do not participate in subsequent digest merging and pruning; at the same time, fragment metadata and verification digests are written to the frozen evidence storage area, and the index relationship corresponding to the event identifier and parent node identifier is recorded. If the evidence collection strategy requires minimizing propagation tree disturbance, the newly added node is still rolled back and deleted, but an independent record related to the event identifier and alignment information is retained in the frozen evidence storage area, so that evidence preservation does not depend on the long-term existence of propagation tree nodes.

[0089] Those skilled in the art will understand that the settings for the form of frozen node marking, the timing of rollback deletion triggering, and the expiration time can all be configured according to compliance requirements and resource status, as long as the traceability of evidence is ensured while preventing the spread of the tree from getting out of control. This application does not impose any further limitations.

[0090] Next, with reference to the accompanying drawings, another intelligent distributed data storage method based on edge computing provided in this application will be further described. Figure 7 The method shown is applied to the review nodes of a hot topic review platform, which also includes multiple edge nodes. Each review node corresponds to a review area composed of multiple edge nodes. The method includes: A1: Receive candidate similar subtrees from edge nodes; A2: Extract event clues based on the candidate similar subtrees, wherein the event clues include representative features corresponding to the root node of the candidate similar subtrees and content features and time information of the fragment metadata contained in the candidate similar subtrees; A3: Determine the target audit node for voting, and distribute the event clues to each target edge node in the audit area corresponding to the target audit node, so that each target edge node performs anchor point routing retrieval based on its own maintained propagation tree and returns the voting results; A4: Receive the voting summary results returned by each target review node, perform weighted voting based on the voting summary results, determine the review conclusion label and event attribution information corresponding to the candidate similar subtree, generate a review result based on the review conclusion label and event attribution information, and send the review result to the edge node that initiated the candidate similar subtree, so that the edge node can determine the storage method of the shard metadata and perform storage processing according to the review result.

[0091] It should be noted that the review node in this application is not concerned with improving the similarity calculation accuracy of a single node, but rather with the inherent structural heterogeneity problem faced by hot data organized in a propagation tree form at the edge. Because different edge nodes are located at different access positions, the generation and maintenance of the propagation tree will vary depending on the local data arrival order, pruning strategy, summarization strategy, and heat decay mechanism. The tree structures formed by the same propagation event on different edge nodes do not have directly alignable node identifiers or fixed hierarchical relationships.

[0092] Those skilled in the art will understand that attempting to align node identifiers or force tree structure consistency will typically introduce high-frequency synchronization and consistency maintenance overhead, making it difficult for the edge side to maintain stable throughput during hotspot outbreaks. On the other hand, completely abandoning structural information and using only flat indexes for full-domain retrieval will result in a lack of propagation semantics, leading to a lack of contextual support for verification and making it difficult to reliably distinguish between new events, derived branches, and transported variants.

[0093] Based on this, this embodiment introduces a transferable clue representation in the organization of the verification link. Its core lies in transforming the candidate similar subtrees generated on the edge side into event clues that do not rely on local node identifiers. This allows the clues to be distributed to other edge nodes and located and verified within their respective propagation trees. At the representation level, event clues emphasize a combination of representativeness and temporality: representativeness is used to carry the feature centers of the propagating event, while temporality is used to limit the local window of propagation evolution, thereby preventing clues from being mistakenly absorbed by irrelevant branches during cross-node propagation.

[0094] Furthermore, to avoid the problem of uncontrollable resource consumption caused by cross-node retrieval degenerating into full tree traversal, this embodiment uses anchor point routing to achieve rapid positioning within the propagation tree. Anchor nodes are not arbitrarily selected, but rather nodes with higher representativeness and stronger stability are selected from the main path as routing entry points, enabling the retrieval to converge from a finite candidate set in most cases. In this embodiment, anchor point conditions can be composed of propagation count, representative feature stability, popularity parameters, or combinations thereof. In addition to containing the representative features of the anchor node, anchor point features can also carry a summary signature of the anchor-covered subtree, used to filter out significantly irrelevant tree domains during the coarse matching stage.

[0095] Those skilled in the art will understand that the scope of the subtree covered by the anchor point can be limited by tree depth, number of branches or time window, and its configuration value can be set in combination with the resource status of the edge node and the hotspot evolution speed, as long as it can ensure that enough candidate attribution information is obtained within the limited search space.

[0096] Furthermore, after forming candidate attribution information, this embodiment does not rely on the judgment result of a single edge node as the final conclusion, but instead uses a voting method to achieve cross-node consensus convergence. When the voting result is returned to the review node, it preferably carries the candidate attribution information and confidence score, and may also carry an evidence summary to assist in weight allocation if necessary. The introduction of weighted voting makes the review link more robust to noise: when individual edge nodes make judgment deviations due to local index pruning, different levels of summarization, or incomplete short-term caching, the decision result can still converge to a more stable event attribution under the common evidence of most nodes.

[0097] In one example, extracting event clues based on the candidate similar subtree includes: Obtain the root node of the candidate similar subtree, and extract the representative feature from the root node; If child nodes exist, obtain the end node associated with the root node from the candidate similar subtree, and extract content features from the fragment metadata of the end node; Obtain the time information corresponding to the shard metadata of the end node, the time information including the generation timestamp and the shard time window; The content features are normalized to generate a cue feature vector, wherein the normalization process includes scale normalization, quantization encoding, and hash mapping; Based on the representative features of the root node and the clue feature vector, a fused clue feature is generated, and the fused clue feature is encapsulated with the time information to obtain an event clue.

[0098] It is understandable that the extraction of event clues is not a simple summary of existing information in candidate similar subtrees, but rather a reorganization and compression of representative information in candidate similar subtrees, centered around the goals of cross-node portability, locatability, and adjudicability. Since the propagation trees maintained locally by different edge nodes naturally differ in structural form, node numbering, and branch depth, directly transmitting the complete subtree structure or node-level index information would make it difficult for the receiving side to effectively locate the information within its own propagation tree. Therefore, this embodiment uses the root node of the candidate similar subtree as the core, extracting its representative features as a stable anchor point for the event, and then supplementing it with the latest propagation form reflected by the terminal nodes. This allows the event clues to simultaneously possess the dual semantics of event center and latest evolution, thereby maintaining information sufficiency while avoiding dependence on specific tree structures.

[0099] In this embodiment, when there are child nodes in the candidate similar subtree, the terminal node associated with the root node is preferentially selected as the object for extracting content features and time information. This is because the fragment metadata corresponding to the terminal node usually represents the latest arriving sample of the event on the current edge node side, which can reflect the latest changes in the event in terms of editing method, content perturbation, or propagation rhythm. By extracting content features from the fragment metadata of the terminal node and simultaneously obtaining its generation timestamp and slice time window, a clear temporal positioning basis can be introduced into the event clue, enabling the receiving side to prioritize matching within the propagation structure range with similar time when performing anchor routing or similarity retrieval, reducing the probability of false adsorption across events. If there are no child nodes in the candidate similar subtree, the fragment metadata corresponding to the root node itself is used as the feature source. At this time, the event clue naturally degenerates into a single-node event entry form, which can still meet the needs of new event confirmation or cross-node consensus voting.

[0100] Furthermore, to ensure the stability and comparability of event clues during cross-node transmission and matching, this embodiment performs normalization processing on the extracted content features and generates clue feature vectors. The purpose of normalization is to eliminate differences in feature extraction models, scale ranges, or precision configurations among different edge nodes, enabling event clues to be matched with consistent standards in the propagation trees of different nodes. For example, scale normalization is used to map continuous features of different dimensions to a unified numerical range, quantization encoding is used to reduce feature precision to improve the efficiency of cross-node transmission and comparison, and hash mapping is used to generate stable low-dimensional representations to support fast coarse matching. Based on this, the representative features of the root node are fused with the clue feature vectors. This can be achieved through weighted concatenation, feature merging, or mapping to the same vector space, so that the fused clue features retain both the long-term stable features of the event and the local change information of the latest samples. Finally, the fusion clue features and time information are encapsulated together as an event clue, which can be used for anchor point routing retrieval, voting decision and event attribution determination when it is subsequently distributed to other edge nodes, without relying on any identification information directly related to the original propagation tree structure, thereby realizing the transferable event expression in a distributed environment.

[0101] In one example, determining the target review node includes: determining the active time interval of the event based on the time information in the event clues corresponding to the candidate similar subtrees; Based on the active time interval, review nodes that cover the time interval and are currently in a valid working state are selected in the hotspot review platform and are used as a candidate review node set. Obtain information on the load status, index completeness, and historical voting participation of edge nodes within the review area corresponding to each candidate review node; Based on the load status, index completeness, and historical voting participation, each candidate review node is comprehensively evaluated to determine at least one target review node.

[0102] The index completeness is used to characterize the maintenance integrity of the propagation tree structure and shard metadata by the edge nodes within the corresponding review area, and the historical voting participation is used to reflect the effective contribution of the review nodes to the event adjudication in the past review process.

[0103] It is understood that the calculation methods for the edge node load status, index completeness, and historical voting participation can be implemented using the following exemplary methods: The edge node load status can be calculated comprehensively based on computing resource load, storage resource load, and network resource load. The computing resource load may include at least one of CPU utilization, memory utilization, inference acceleration unit utilization, and task queue length; the storage resource load may include at least one of storage space utilization, write latency, and I / O throughput; and the network resource load may include at least one of bandwidth utilization, link latency, and packet loss rate. In one optional implementation, each load indicator is first normalized to obtain the node load value corresponding to each edge node. Then, the node load values ​​of multiple edge nodes within the audit area are aggregated to obtain the regional load status of the corresponding audit node. The aggregation method can be median, weighted average, or truncated average to reduce the impact of abnormal nodes on the evaluation results.

[0104] The index completeness can be determined jointly by structural coverage metrics, metadata completeness metrics, and index freshness metrics. The structural coverage metrics reflect the maintenance status of the propagation tree structure and may include at least one of trunk path coverage, effective branch coverage, and anchor node availability. The metadata completeness metrics reflect the retrievability of shard metadata and may include at least one of shard metadata existence rate, digest signature completeness rate, and time information completeness rate. The index freshness metrics reflect the timeliness of index updates and can be calculated based on the time difference between the most recent index update time and the current time. In an optional implementation, the node index completeness of each edge node can be calculated separately and adjusted by combining index freshness and pruning strength correction factors. Then, the node index completeness of multiple edge nodes within the review area is weighted and aggregated to obtain the regional index completeness of the corresponding review node.

[0105] The historical voting participation can be calculated comprehensively using participation rate, response quality, and consistency metrics. The participation rate metric characterizes the actual participation of edge nodes when requested to vote; the response quality metric characterizes the timeliness and completeness of the voting results returned by edge nodes; and the consistency metric characterizes the degree of consistency between the historical voting results of edge nodes and the final ruling. In one optional implementation, the historical voting participation of each edge node within the audit area can be calculated first, and then the historical voting participation of multiple edge nodes can be aggregated to obtain the regional historical voting participation of the corresponding audit node.

[0106] In one example, the target edge nodes perform anchor point route retrieval based on their respective maintained propagation trees and return voting results, including: A3.1: Obtain a local propagation tree at the target edge node and determine at least one anchor node from the local propagation tree. The anchor node is a trunk path node that satisfies preset anchor conditions, wherein the preset anchor conditions include the number of propagations being greater than or equal to a first threshold and the stability of the representative feature being greater than or equal to a second threshold. Specifically, the propagation tree at edge nodes is influenced by the arrival order of local data, pruning strategies, summarization granularity, and heat decay mechanisms during hotspot evolution. The tree structures formed by the same event on different edge nodes often cannot be aligned using node identifiers or fixed levels. If a full tree traversal is used to find matching branches during cross-node retrieval, uncontrollable retrieval overhead will be introduced during hotspot outbreaks. Furthermore, the candidate set can easily become too large due to noisy branches, transport variants, and temporal drift, affecting the verification convergence. Therefore, at the target edge node, a more stable route entry point is prioritized from the propagation backbone path, allowing the retrieval to converge within a limited candidate space. This incorporates cross-node transferable clues into the local propagation semantic structure, forming the basis for subsequent similarity retrieval and voting decisions.

[0107] In this embodiment, after receiving event clues, the target edge node reads the local propagation tree from local storage or memory index; it selects nodes that meet the preset anchor point conditions from the main path node set as anchor points; the propagation number in the preset anchor point conditions can be accumulated according to the number of times the node is referenced, forwarded, or associated as a parent node, and the stability of the representative feature can be measured by the degree of drift or update consistency of the node's representative feature in multiple time windows. For example, the change range of the node's representative feature in the most recent maintenance cycles can be mapped to a stability index, and this index can be maintained synchronously when the feature is updated.

[0108] Furthermore, the setting of the first threshold and the second threshold can be configured in combination with the hotspot type, edge node resources and the expected number of routing entry points, so as to avoid the subsequent matching cost increasing due to too many anchor points, or the recall decreasing due to insufficient routing entry points caused by too few anchor points.

[0109] In one example, the first threshold can be adaptively set according to the quantile of the propagation frequency distribution of the backbone path nodes in the review area over a recent period. For example, the 70th quantile of the propagation frequency of the backbone path nodes can be used as the first threshold, with a minimum value of no less than 30 times. When resources allow and the hotspot intensity is high, the first threshold can also be fixed at 50 times to improve the representativeness of the anchor points. The second threshold can be set according to the empirical distribution representing the stability of the features. For example, after normalizing the stability to the range of 0 to 1, 0.85 can be used as the second threshold to exclude nodes that are in a rapid drift or frequently rewritten in the summary. When the model version difference on the edge node side is large, the second threshold can be appropriately lowered to 0.80 to ensure that the number of anchor points meets the routing entry requirements.

[0110] A3.2: Obtain anchor features for each anchor node, wherein the anchor features include representative features of the anchor node and subtree digest signature of the subtree corresponding to the anchor node; Specifically, relying solely on the representative features of anchor nodes for cross-node matching can easily lead to mismatches between events with similar semantics but different propagation contexts. This is especially true when trending content is templated, rewritten in title, or subject to local perturbations, where the representative features of different events may appear similar in certain dimensions, causing routing entry point shifts. Furthermore, using the complete structure of the anchor-covered subtree or all fragment metadata for comparison introduces high overhead from cross-node transmission and local comparison, which is detrimental to throughput during high-concurrency phases of trending content. Therefore, in addition to carrying the representative features of the anchor node, anchor features also need to carry a compressed representation of the anchor-covered subtree. This allows the coarse matching stage to determine whether the subtree domain is in the same propagation context as the event clue with a lower cost, thus limiting the fine-tuning scope to a more reasonable candidate domain.

[0111] In this embodiment, for each anchor node, the target edge node reads its representative features as anchor representative features. These representative features can be multimodal features or unified embedding representations, and their sources may include text semantic vectors, visual semantic vectors, audio fingerprint vectors, or their fusion representations. The anchor node has already updated and solidified its representative features during the propagation tree maintenance phase. Simultaneously, the target edge node also generates a subtree summary signature for the covering subtree corresponding to the anchor node. This subtree summary signature is used to compress and represent the content form and propagation evolution of the covering subtree, preferably containing at least two of the following information types: a quantitative summary of the fragmented metadata content features within the covering subtree, a summary of the time window distribution within the covering subtree, a statistical summary of the branch structure within the covering subtree, and a fingerprint set summary of the key nodes within the covering subtree. The generation of the subtree summary signature can be maintained offline or incrementally: when inserting, pruning, or rewriting the summary of propagation tree nodes, the summary signature of the covering subtree is incrementally updated, ensuring that the anchor features remain available as the tree is maintained.

[0112] A3.3: Based on the event clues and the anchor features of each anchor node, the target anchor node with the highest matching score and a matching score greater than or equal to the preset score threshold is determined. If it does not exist, the voting result is not returned. Specifically, event clues emphasize the combination of representativeness and temporality, aiming to avoid being absorbed by irrelevant branches when migrating across nodes. However, on the target edge node side, the propagation tree may have multiple anchor nodes that are similar to the clues in terms of representative feature dimension. If a hierarchical matching strategy and threshold constraints are not introduced, the anchor selection is prone to deviation, which may lead to subsequent similarity searches in the wrong subtree domain, resulting in incorrect candidate attribution information or no search results, and causing unnecessary computation and storage access overhead.

[0113] In this embodiment, after receiving an event clue, the target edge node parses the fused clue features and time information from the event clue and uses this information to perform matching evaluation on anchor nodes. The matching process can adopt a two-stage strategy of coarse matching and fine matching: In the coarse matching stage, the fused clue features and subtree digest signatures are quickly judged for consistency. For example, the overlap between the time window in the event clue and the time distribution in the subtree digest signature is checked, and the consistency filtering between the quantized summary of the content features in the event clue and the quantized summary in the subtree digest signature is performed, thereby eliminating anchor candidates with non-overlapping time windows or excessively different summaries; In the fine matching stage, the similarity between the fused clue features and the anchor representative features of the candidate anchors that passed the coarse matching is compared, and a comprehensive matching score is formed by combining the consistency results of the subtree digest signature. The anchor node with the highest matching score and not lower than a preset score threshold is determined as the target anchor node; If the matching scores of all candidate anchors are lower than the preset score threshold, it is considered that the event clue cannot find a reliable route entry in the local propagation tree, and the target edge node does not return a voting result to avoid injecting uncertain attribution information into cross-node voting and interfering with the decision convergence.

[0114] Furthermore, the preset score threshold can be set based on the statistical analysis of historical review results within the audit area to achieve a balance between recall and accuracy.

[0115] In one example, a preset score threshold of 0.78 can be used to filter out anchors that are only similar in a few feature dimensions but have inconsistent propagation contexts. When hot content has a large amount of templated rewriting that leads to a decrease in the distinguishability of representative features, the threshold can be increased to 0.82 to reduce the probability of misrouting. When the edge node side pruning is strong and the number of anchors is small, the threshold can be lowered to 0.75 to ensure the necessary supply of routing entry points. Through the above matching and threshold control, the selection of target anchors can achieve a locatable and verifiable landing point under the premise of heterogeneous cross-node structure, providing a stable starting point for subsequent subtree retrieval.

[0116] A3.4: Based on the event clues, perform similarity retrieval on the branch path nodes within the subtree of the target anchor node to obtain candidate attribution information; Specifically, anchor nodes only provide routing entry points and are not equivalent to the final event attribution node. The same anchor node may cover multiple branch evolution paths within its subtree, corresponding to different derived branches, transport variants, or propagation sub-events within different time periods. If anchor nodes are used solely as the attribution criterion, it's easy to mistakenly merge derived branches into the main event or confuse transport propagation within different time periods, thus affecting the accuracy of the verification conclusion labels and the selection of subsequent sharding metadata storage strategies.

[0117] In this embodiment, after determining the target anchor node, the target edge node expands its covering subtree downwards from the anchor node and performs a similarity search on the branch path nodes within a preset range. The preset range can be jointly limited by tree depth, number of branches, and time window to avoid traversing all nodes in the covering subtree; for example, it can limit the search to only the set of nodes within three levels below the anchor node, and prioritize nodes with popularity parameters in the top few percentiles to enter the candidate search list, while limiting the search to the node time window index range near the event clue slice time window. During the search, the target edge node can jointly compare the fragmented metadata content features, node representative features, and node time information of the candidate nodes: at the content feature level, it compares the consistency of the fused clue features of the event clue with the content feature summary of the candidate nodes; at the representative feature level, it evaluates the similarity between the representative components in the event clue and the representative features of the candidate nodes to maintain event center consistency; at the time level, it aligns the generation timestamp and slice time window corresponding to the candidate node with the time information of the event clue, preferably selecting candidate nodes falling into the same or adjacent time windows to reduce the probability of false adsorption across events. The search results can form a candidate attribution list. Each candidate attribution entry includes at least: the attribution node identifier (which can be a local identifier), the attribution path summary, the matching consistency description with the event clue, and optional evidence summary information. The evidence summary information may include the metadata summary signature of the matched key fragment or the set of key node fingerprints.

[0118] A3.5: Generate voting results based on the candidate attribution information and send them to the review node. The voting results include the candidate attribution information and the confidence score corresponding to the candidate attribution information. Specifically, cross-node voting decisions rely on the fact that the attribution judgments returned by each target edge node can be compared and weighted by the auditing node. If the voting results only return a single attribution identifier without confidence and evidence description, the auditing node will find it difficult to distinguish between high-confidence consistency and low-confidence accidental hits. Furthermore, in cases where edge node pruning, digestion levels differ, or short-term caching is incomplete, a small number of low-quality votes may interfere with the decision.

[0119] In this embodiment, the confidence score can comprehensively reflect the following factors: the reliability of the target anchor matching, the similarity and consistency of candidate attribution entries in the subtree retrieval, the time window alignment, and the completeness of the evidence summary. To reduce transmission overhead, the candidate attribution information can preferably only carry the top-ranked attribution candidates and their path summaries, for example, carrying the top 3 candidate attribution entries; at the same time, to facilitate consistency aggregation by the review nodes, the voting results can also carry the event clue summary identifier (e.g., clue hash or clue sequence number), local propagation tree version identifier, or index freshness identifier, which are used by the review nodes to identify the index status of the voting source during the aggregation phase.

[0120] In another example, a weighted vote is performed based on the aggregated voting results to determine the review conclusion label and event attribution information corresponding to the candidate similar subtrees, including: Obtain the voting summary results returned by each target review node, and extract the corresponding candidate attribution information and confidence score from the voting summary results; Based on the confidence score, voting weights are assigned to each voting result, and the weights of voting results with the same candidate attribution information are accumulated. The candidate attribution information with the highest cumulative weight value is determined as the event attribution information, and the review result is determined based on the event attribution information.

[0121] Understandably, after receiving the aggregated voting results from all target review nodes, the review nodes perform unified parsing and normalization of the voting results. The aggregated voting results are not the judgment of a single edge node, but rather a regional-level aggregation of the voting results of all target edge nodes within their respective review areas. Therefore, each aggregated voting result is associated with at least one candidate attribution information and its corresponding confidence score. To ensure the comparability of voting results returned from different review areas during the weighting stage, the review nodes perform consistency processing on the confidence scores during the parsing stage. For example, they perform interval mapping and normalization on the local confidence scales used in different review areas, ensuring that all confidence scores fall within a unified numerical range. They also perform pruning on obviously abnormal confidence values ​​to avoid excessively high or low scores in certain areas due to missing local indexes or short-term anomalies. After completing the above normalization, the review nodes form a standardized voting set of candidate attribution information and confidence scores, which serves as the input for subsequent weighted voting.

[0122] In this embodiment, the allocation of voting weights is not a simple equal-weighted sum, but is determined by combining the confidence score and the reliability of the voting source.

[0123] Specifically, the review nodes use the confidence scores from the aggregated voting results as the base weights, and introduce a regional correction factor to adjust the base weights. This regional correction factor can be set based on the stability of the corresponding review node in the historical review process, the completeness of the edge node index within its review region, and the recent voting effectiveness rate, thus reflecting the long-term contribution of different review regions to the decision. In one example, the base weight can be directly taken as the confidence score itself, and the regional correction factor is limited to a preset range, such as between 0.8 and 1.2, to prevent a single region from excessively dominating the final result. When a review region has high recent index freshness and good historical voting consistency, its correction factor can approach the upper limit; when its index coverage is insufficient or there are multiple recent deviations from the decision, its correction factor can be reduced accordingly. The corrected voting weights are used to accumulate the weights of voting results with the same candidate attribution information, thereby forming a comprehensive weight value corresponding to each candidate attribution information.

[0124] Furthermore, after weight accumulation, the review node performs a consistency assessment of the comprehensive weight distribution of each candidate attribution information to avoid directly issuing unstable conclusions when weights are close or significantly divergent. In one example, when the weight difference between the candidate attribution information with the highest accumulated weight value and the second-highest weight value reaches a preset difference threshold, the candidate attribution information can be directly identified as the event attribution information, and a corresponding review conclusion label can be generated accordingly. When the weight difference does not reach the difference threshold, the review node can mark the review conclusion label as pending confirmation or trigger a further cross-regional supplementary voting mechanism to complete the decision with more supporting evidence.

[0125] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A distributed storage method for intelligent data annotation tasks with edge node collaboration, applied to the edge nodes of a hotspot review platform, characterized in that... The hotspot review platform also includes multiple review nodes, each review node corresponding to a review area composed of multiple edge nodes, and the method includes: For the uploaded content stream slices, generate slice metadata, and write the slice metadata to the local temporary storage area; Based on the preset propagation tree, the fragment metadata is searched for similarity, and candidate similar subtrees are output. The candidate similar subtrees are then forwarded to the corresponding review nodes so that the review nodes can review the candidate similar subtrees and output the review results. The system receives the review result sent by the corresponding review node, determines the storage method of the shard metadata based on the review result, and performs storage processing on the shard metadata according to the storage method, wherein the storage method includes writing the shard metadata into the local persistent storage area and updating the propagation tree.

2. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 1, characterized in that, The generation of fragment metadata includes: The content stream is sliced ​​to determine the time range and frame range of each slice; For each slice, content features are extracted, including video fingerprint, audio fingerprint, keyframe hash, and vector features; Obtain the basic attribute information corresponding to the slice, which includes at least the original content identifier, the upload account identifier, the generation timestamp, the edge node identifier, and the region identifier to which the edge node belongs; All content features and basic attribute information of all slices are encapsulated into the slice metadata.

3. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 1, characterized in that, The step of performing similarity retrieval on the fragmented metadata based on a preset propagation tree and outputting candidate similar subtrees includes: Based on the nodes of the main path of the propagation tree and the content characteristics of the fragmented metadata, it is determined whether it is a new propagation event. If so, the fragmented metadata is converted into a main path node and attached to the end of the main path, and the main path node is output as the candidate similar subtree. If not, the fragmented metadata is converted into branch path nodes and attached to the end of the corresponding trunk path node, and the corresponding trunk path node and its contained branch path nodes are output as the candidate similar subtrees. The main path node includes an event identifier, representative features, and propagation parameters. The propagation parameters are used to determine the propagation number of the fragment metadata, with an initial value of 1. The branch path node includes a parent node identifier, a similarity score, and alignment information with the content corresponding to the parent node. It is configured to perform pruning, merging, and summary storage when preset conditions are met.

4. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 3, characterized in that, The process of converting the fragmented metadata into branch path nodes and attaching them to the end of the corresponding trunk path node includes: Based on the content features of the fragmented metadata, the representative features of each trunk path node of the propagation tree are matched to determine the trunk path node with the highest similarity to the fragmented metadata as the corresponding trunk path node. Obtain the event identifier of the corresponding trunk path node, and set the event identifier as the parent node identifier of the shard metadata to obtain the branch path node; Calculate the similarity score between the fragment metadata and the corresponding content of the corresponding trunk path node, and write the similarity score into the branch path node; Calculate alignment information and write the alignment information into the branch path node, wherein the alignment information includes time offset, start and end positions of the matching segment, keyframe matching relationship and audio fingerprint matching interval; Insert the branch path node at the end of the branch path rooted at the corresponding trunk path node; Increment the propagation count of the corresponding trunk path node by one, and update the propagation parameters of the corresponding trunk path node based on the propagation count.

5. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 1, characterized in that, The review result includes a review conclusion label, an event identifier corresponding to the review conclusion label, a review similarity score, and review alignment information; as well as storage strategy parameters corresponding to the review conclusion label, the storage strategy parameters including a storage method identifier and a lifecycle parameter.

6. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 5, characterized in that, Based on the verification results, the storage method of the fragment metadata is determined, and the fragment metadata is stored and processed according to the storage method, including: Parse the review results to obtain the review conclusion label and the storage strategy parameters corresponding to the review conclusion label; The storage method of the fragment metadata is determined based on the review conclusion label, and the storage method includes at least one of the following: persistent storage, digest storage, frozen evidence storage, temporary retention, and deletion. When the storage method is persistent storage, the shard metadata is written to the local persistent storage area and the propagation tree is updated. Updating the propagation tree includes updating the node association relationship of the propagation tree according to the verification alignment information and accumulating the corresponding propagation count.

7. A distributed storage method for intelligent data annotation tasks with edge node collaboration, applied to the review nodes of a hotspot review platform, characterized in that... The hot topic review platform also includes multiple edge nodes, each review node corresponding to a review area composed of multiple edge nodes, and the method includes: Receive candidate similar subtrees from edge nodes; Event clues are extracted based on the candidate similar subtrees, wherein the event clues include representative features corresponding to the root node of the candidate similar subtrees and content features and time information of the fragment metadata contained in the candidate similar subtrees; The target audit node for voting is determined, and the event clues are distributed to each target edge node in the audit area corresponding to the target audit node, so that each target edge node performs anchor point routing retrieval based on its own maintained propagation tree and returns the voting results; The system receives the voting summary results returned by each target review node, performs weighted voting based on the voting summary results, determines the review conclusion label and event attribution information corresponding to the candidate similar subtree, generates a review result based on the review conclusion label and event attribution information, and sends the review result to the edge node that initiated the candidate similar subtree, so that the edge node determines the storage method of the shard metadata and performs storage processing according to the review result.

8. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 7, characterized in that, Extracting event clues based on the candidate similar subtrees includes: Obtain the root node of the candidate similar subtree, and extract the representative feature from the root node; If child nodes exist, obtain the end node associated with the root node from the candidate similar subtree, and extract content features from the fragment metadata of the end node; Obtain the time information corresponding to the shard metadata of the end node, the time information including the generation timestamp and the shard time window; The content features are normalized to generate a cue feature vector, wherein the normalization process includes scale normalization, quantization encoding, and hash mapping; Based on the representative features of the root node and the clue feature vector, a fused clue feature is generated, and the fused clue feature is encapsulated with the time information to obtain an event clue.

9. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 7, characterized in that, The target edge nodes perform anchor point route retrieval based on their respective maintained propagation trees and return voting results, including: A local propagation tree is obtained at the target edge node, and at least one anchor node is determined from the local propagation tree. The anchor node is a trunk path node that meets preset anchor conditions, wherein the preset anchor conditions include the number of propagations being greater than or equal to a first threshold and the stability of the representative feature being greater than or equal to a second threshold. For each anchor node, anchor features are obtained, wherein the anchor features include representative features of the anchor node and subtree digest signature of the subtree corresponding to the anchor node. Based on the event clues and the anchor features of each anchor node, the target anchor node with the highest matching score and a matching score greater than or equal to a preset score threshold is determined. If it does not exist, the voting result is not returned. Based on the event clues, similarity searches are performed on branch path nodes within the subtree of the target anchor node to obtain candidate attribution information; A voting result is generated based on the candidate attribution information and sent to the review node. The voting result includes the candidate attribution information and the confidence score corresponding to the candidate attribution information.

10. The distributed storage method for intelligent data annotation tasks with edge node collaboration according to claim 9, characterized in that, Based on the aggregated voting results, a weighted vote is performed to determine the review conclusion label and event attribution information corresponding to the candidate similar subtrees, including: Obtain the voting summary results returned by each target review node, and extract the corresponding candidate attribution information and confidence score from the voting summary results; Based on the confidence score, voting weights are assigned to each voting result, and the weights of voting results with the same candidate attribution information are accumulated. The candidate attribution information with the highest cumulative weight value is determined as the event attribution information, and the review result is determined based on the event attribution information.

Images