Dynamic optimization system and method based on label nebula density analysis and storage medium
By using a tag nebula density analysis system, a tag star map is constructed to identify dense and sparse regions, generate optimization suggestions, solve the problems of redundancy and missing tags, realize the self-diagnosis and optimization of the tag system, and improve the retrieval accuracy and the robustness of the business system.
Patent Information
- Application Number
- CN202511440072.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-10
- Publication Date
- 2026-01-30
AI Technical Summary
Existing tagging systems suffer from problems such as tag overfitting and redundancy, tag missing and content vacuum, and lack of self-optimization in static management. This results in high overlap of search results, poor robustness, waste of computing and storage resources, and a lack of continuous data-driven feedback optimization loop.
A dynamic optimization system based on tag nebula density analysis is adopted. By constructing a tag star map, calculating the correlation probability between tags and identifying dense and sparse regions, optimization suggestions are generated. Combined with human-computer interaction and effect evaluation, a continuously self-evolving optimization loop is formed, realizing the self-diagnosis and optimization of the tag system.
It enables quantitative diagnosis and visualization of the tag system, dynamically adapts to business changes, improves retrieval accuracy and robustness, solves the problems of tag overfitting and missing tags, promotes the improvement of business system content, and the effect is measurable.
Smart Images
Figure CN121434483A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of database structure technology, and specifically to a dynamic optimization system, method, and storage medium based on tag nebula density analysis for large-scale business systems. Background Technology
[0002] With the development of big data and precision services, e-commerce, content recommendation, and user profiling systems commonly adopt tagging for data management. However, existing technologies have the following inherent drawbacks:
[0003] Tag overfitting and redundancy: Different business personnel may create a large number of tags that are semantically highly similar but have subtle differences in expression, such as "loves high-end digital products" and "loyal Apple phone fan". This leads to a bloated tag system, resulting in high overlap of search results and poor robustness when performing tag searches and data analysis, wasting computing and storage resources.
[0004] Missing tags and content vacuum: The construction of tagging systems often lags behind business development, resulting in sparse or even missing tags in certain areas. This makes it impossible to effectively cover and describe all entities, creating a "content vacuum" and reducing the comprehensiveness and accuracy of tag analysis.
[0005] Static management, lack of self-optimization: Existing label management systems rely heavily on manual review and static rules, which cannot automatically and quantitatively perceive the overall health of the label system, and lack a continuous, data-driven feedback optimization loop.
[0006] Therefore, there is an urgent need in this field for a system and method that can dynamically perceive, diagnose, and optimize the labeling system from a global perspective. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a dynamic optimization system and method based on tag nebula density analysis. Its core idea is to treat the entire tag system as a dynamic nebula, calculate the correlation probability between tags as spatial distance, thereby quantitatively analyzing the "dense" and "sparse" regions of tag distribution, and selectively thinning and supplementing these regions to ultimately form a continuously self-evolving healthy tag ecosystem. The specific technical solution is as follows:
[0008] In a first aspect, the present invention provides a dynamic optimization system based on tag nebula density analysis, comprising:
[0009] Data connection and acquisition module; used to connect to the database of the business system and collect all tag data, entity data, and tag-entity relationship data.
[0010] The tag star map construction and calculation module receives collected data and constructs an undirected graph network, i.e., a "tag star map" or "tag nebula," with tags as nodes and the probability of association between tags as edge weights. The distance D(A, B) between tag A and tag B is defined as:
[0011] D(A, B) = 1 - P(A∩B) / sqrt(P(A) * P(B)),
[0012] Where P(A∩B) is the probability that labels A and B are simultaneously associated with the same entity, and P(A) and P(B) are the independent association probabilities of labels A and B, respectively.
[0013] Denseness analysis module; used for graph analysis of the constructed tag nebula;
[0014] Region identification: Calculate the local average density of each region or community in the nebula and compare it with the reference density to identify abnormal regions that are "too dense" and "too sparse".
[0015] Optimize suggestion generation module;
[0016] For dense areas: This module will use the center label of the area as a reference to find a set of redundant labels that are too close to it; and generate "relocation suggestions";
[0017] For sparse regions: This module analyzes the semantics of the boundary labels of sparse regions and generates a set of candidate "mediator labels" to connect isolated label clusters through natural language processing (NLP) technology or knowledge graph-based association. At the same time, it also provides a reverse hint to the business system that there may be a certain type of "content vacuum", that is, a lack of entities that can be described by such mediator labels.
[0018] Preferred options also include:
[0019] Human-computer interaction and confirmation module: This module presents suggestions for scrambling and replacing tags, suggestions for adding tags, and suggestions for optimizing content vacuum prompts to the tag system administrator in a visual interface and list format.
[0020] Preferably, it also includes: an effect evaluation and feedback closed-loop module, which is used to continuously monitor the application effect of the business system after optimization; this module generates a quantitative analysis report of the effect comparison data over a period of time, and submits it to the administrator again, so that the administrator can judge whether the optimization is successful, needs further improvement, or is a failure; thus forming a continuous improvement closed loop of "analysis-optimization-evaluation-re-optimization".
[0021] Furthermore, the performance evaluation and feedback closed-loop module collects the following metrics:
[0022] Search precision / recall; use tags before and after optimization to search and compare the relevance of the results;
[0023] Search bounce rate; the proportion of users who revise their search after performing it. A decrease in bounce rate indicates improved robustness.
[0024] Tag usage concentration / coverage; analyze whether the distribution of tag usage is healthier.
[0025] Secondly, the present invention provides a dynamic optimization method based on tag nebula density analysis, comprising the following process:
[0026] Step 1: Data Collection; Extract all tags and tag-entity association data from the business system;
[0027] Step 2: Construct a label nebula; calculate the pairwise association probabilities between all labels and convert them into distance weights to construct a graph data structure;
[0028] Step 3: Set up a reference frame and analyze the density; calculate and identify dense and sparse regions in the nebula;
[0029] Step 4: Generate optimization suggestions; for dense areas, generate a relief scheme to replace redundant surrounding tags with core tags; for sparse areas, generate an intermediate tag supplement scheme to connect isolated tag clusters or a content vacuum prompt for the business system.
[0030] Step 5: Manual review and confirmation; Submit the suggestions to the administrator for manual review, correction, and final confirmation;
[0031] Step 6: Implement optimization; update the confirmed solution to the tag system of the business system.
[0032] Preferably, the process following step 6 includes:
[0033] Step 7: Performance monitoring and evaluation; After the predetermined evaluation period, collect the performance indicators of the business system and conduct a comparative analysis before and after optimization;
[0034] Step 8: Feedback and Decision-Making; Generate an evaluation report, and the administrator decides on the next action based on the data. The types of actions in the next round include: maintain, deepen optimization, or rollback. The process ends or jumps to step S101 to start a new optimization cycle.
[0035] Furthermore, the distance calculation method in step 2 is as follows:
[0036] D(A, B) = 1 - P(A∩B) / sqrt(P(A) * P(B));
[0037] Where P(A∩B) is the probability that labels A and B are simultaneously associated with the same entity, and P(A) and P(B) are the independent association probabilities of labels A and B, respectively.
[0038] Furthermore, in step 4, the core label is the node with the highest degree centrality; the mediator label is obtained by finding the semantic midpoint through NLP word vector calculation or by searching for related concepts from the knowledge graph.
[0039] Furthermore, the reference system is set according to excellent sample reference or uniform distribution reference;
[0040] The excellent sample is a set of high-quality labels that have been predefined by experts and are recognized as excellent. The distribution characteristics of this set in the nebula are used as a reference for the ideal density.
[0041] The uniform distribution reference is to automatically find the "tag cluster" in the nebula with the most uniform tag distribution and the best connectivity, and use its density as a reference benchmark.
[0042] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described optimization methods.
[0043] Compared with the prior art, the present invention has the following significant advantages:
[0044] I. Global Perspective, Quantitative Diagnosis: The abstract issue of "quality" of labels is transformed into the concrete issue of "density" of graph networks, realizing the quantitative and visual diagnosis of the health of the label system.
[0045] 2. Dynamic Adaptation: The system is not a static tool, but a dynamic optimization closed loop embedded in the business operation process. It can continuously evolve itself as business data changes, ensuring the long-term vitality of the tag system.
[0046] Third, human-machine collaboration ensures reliable decision-making: It leverages the powerful computing and discovery capabilities of computers while fully respecting the domain knowledge and final decision-making authority of human experts, thus balancing automation and accuracy.
[0047] Fourth, two-way optimization: It not only solves the problems of "overfitting" and "missing" data, but also promotes the improvement of business system content, making it even more valuable.
[0048] 5. Measurable Results: The optimization effect is evaluated through rigorous A / B testing and quantitative indicators such as accuracy and return rate, so that each iteration of improvement is based on evidence and avoids blind optimization. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the overall system architecture in an embodiment of the present invention.
[0050] Figure 2 This is a schematic diagram of the method flow in an embodiment of the present invention.
[0051] Figure 3 This is a schematic diagram illustrating the optimized effect of the tag nebula in an embodiment of the present invention. Detailed Implementation
[0052] A dynamic optimization system based on tag nebula density analysis includes:
[0053] The data connection and acquisition module is used to connect to the database of the business system and collect all tag data, such as entity data of users, products, and content, as well as tag-entity relationship data.
[0054] The tag star map construction and calculation module is the core computing engine of the system. It receives the collected data and constructs an undirected graph network, namely the "tag star map" or "tag nebula," with tags as nodes and the probability of association between tags as edge weights. The distance D(A, B) between tag A and tag B is defined as follows:
[0055] D(A, B) = 1 - P(A∩B) / sqrt(P(A) * P(B)),
[0056] Where P(A∩B) is the probability that labels A and B are simultaneously associated with the same entity, and P(A) and P(B) are the independent association probabilities of labels A and B, respectively.
[0057] The smaller the distance value, the greater the probability that the two labels will appear simultaneously in the business, and the more similar they are semantically.
[0058] Denseness analysis module; used for graph analysis of the constructed tag nebula;
[0059] Region identification: Calculate the local average density of each region or community in the nebula and compare it with the reference density to identify abnormal regions that are "too dense" and "too sparse".
[0060] Optimize suggestion generation module;
[0061] For dense regions: This module uses the central label of the region as a benchmark, which is the core label with the highest relevance, and finds a set of redundant labels that are too close to it; it generates "relocation suggestions"; for example: it suggests replacing redundant labels L_redundant1, L_redundant2... with the core label L_core; or it suggests merging multiple redundant labels into a new label with a broader meaning;
[0062] For sparse regions: This module analyzes the semantics of the boundary labels of sparse regions and generates a set of candidate "mediator labels" to connect isolated label clusters through natural language processing (NLP) technology or knowledge graph-based association. At the same time, it may also provide a reverse hint that there may be a certain type of "content vacuum" in the business system, that is, a lack of entities that can be described by such mediator labels.
[0063] The human-computer interaction and confirmation module presents suggestions for shunting and replacing tags, suggestions for adding tags, and suggestions for optimizing content vacuums to the tag system administrator in a visual interface, such as highlighted annotations on a nebula map. All add, delete, and modify operations must be manually reviewed and confirmed by the administrator before execution, ensuring that the decision-making power of the automated process remains with humans.
[0064] The effect evaluation and feedback closed-loop module is used to continuously monitor the application effect of the business system after optimization. This module generates a quantitative analysis report of the effect comparison data over a period of time, and submits it to the administrator again. The administrator judges whether the optimization is successful and should be continued, whether further improvement is needed, or whether it is a failure and should be rolled back. This forms a continuous improvement closed loop of "analysis-optimization-evaluation-re-optimization".
[0065] The performance evaluation and feedback closed-loop module collects the following indicators:
[0066] Search precision / recall; use tags before and after optimization to search and compare the relevance of the results;
[0067] Search bounce rate; the proportion of users who revise their search after performing it. A decrease in bounce rate indicates improved robustness.
[0068] Tag usage concentration / coverage; analyze whether the distribution of tag usage is healthier.
[0069] A dynamic optimization method based on tag nebula density analysis includes the following process:
[0070] Step 1: Data Collection; Extract all tags and tag-entity association data from the business system;
[0071] Step 2: Construct a label nebula; calculate the pairwise association probabilities between all labels and convert them into distance weights to construct a graph data structure; the distance calculation method is as follows:
[0072] D(A, B) = 1 - P(A∩B) / sqrt(P(A) * P(B));
[0073] Where P(A∩B) is the probability that labels A and B are simultaneously associated with the same entity, and P(A) and P(B) are the independent association probabilities of labels A and B, respectively.
[0074] Step 3: Set a reference system and analyze the density; the reference system is set according to excellent sample reference or uniform distribution reference;
[0075] The excellent sample is a set of high-quality labels that have been predefined by experts and are recognized as excellent. The distribution characteristics of this set in the nebula are used as a reference for the ideal density.
[0076] The uniform distribution reference is to automatically find the "tag cluster" in the nebula with the most uniform tag distribution and the best connectivity, and identify it through community discovery algorithms such as Louvain, using its density as a reference benchmark.
[0077] Calculate and identify dense and sparse regions in a nebula;
[0078] Step 4: Generate optimization suggestions; for dense areas, generate a relief scheme that replaces redundant surrounding tags with core tags; for sparse areas, generate a supplementary scheme for intermediate tags to connect isolated tag clusters or a vacuum prompt for business system content; the core tag is the node with the highest degree centrality; the intermediate tags are used to find semantic midpoints through NLP word vector calculation or to find related concepts from the knowledge graph;
[0079] Step 5: Manual review and confirmation; Submit the suggestions to the administrator for manual review, correction, and final confirmation;
[0080] Step 6: Implement optimization; update the confirmed solution to the tag system of the business system;
[0081] Step 7: Performance monitoring and evaluation; After the predetermined evaluation period, collect the performance indicators of the business system and conduct a comparative analysis before and after optimization;
[0082] Step 8: Feedback and Decision-Making; Generate an evaluation report, and the administrator decides on the next action based on the data. The types of actions in the next round include: maintain, deepen optimization, or rollback. The process ends or jumps to Step 1 to start a new optimization cycle.
[0083] Comparison Figure 3 This figure visually demonstrates the changes in the Tag Nebula after being processed by the method of this invention by comparing the "before optimization" and "after optimization".
[0084] The left image shows the situation before optimization. There is a very dense label cluster with many nodes crowded together and short connections, indicating label redundancy and overfitting. There is also a sparse label cluster with few nodes that are far from the core area, indicating missing labels or content vacuum. Finally, there is a relatively evenly distributed label cluster, which can serve as an internal reference system.
[0085] The right image shows the optimized result. Dense regions are thinned out: redundant nodes in dense clusters are removed or merged, retaining only one or a few core labels, resulting in a clearer and simpler structure. Sparse regions are supplemented: new intermediary labels are added between sparse and core regions, effectively connecting previously isolated areas, making the entire nebula more connected, structurally more complete, and more evenly distributed.
Claims
1. A dynamic optimization system based on label constellation density analysis, characterized in that, Comprise: Data connection and collection module; for connecting the database of the business system, collecting all label data, entity data and label-entity association data; Label star map construction and calculation module; for receiving the collected data, constructing a label as a node, and a label association probability as an edge weight undirected graph network, namely "label star map" or "label star cloud"; wherein the distance D(A, B) between label A and label B is defined as: D(A, B) = 1 - P(A∩B) / sqrt(P(A) * P(B)), Wherein, P(A∩B) is the probability that label A and B are associated with the same entity at the same time, P(A) and P(B) are the independent association probabilities of label A and B respectively; Dense analysis module; for graph analysis on the constructed label star cloud; Region identification: calculate the local average density of each region or community in the star cloud, and compare it with the reference density to identify the "too dense" and "too sparse" abnormal regions; Optimization suggestion generation module; For dense regions: this module will take the center label of the region as the benchmark, find out the redundant label set around it with too close distance; generate "thinning suggestion"; For sparse regions: this module will analyze the semantics of the boundary label of the sparse region, generate a group of candidate "intermediate labels" to connect the isolated label group through natural language processing NLP technology or based on knowledge graph association; at the same time, it also reversely prompts the business system that there may be a certain "content vacuum", that is, lack of entities that can be described by this kind of intermediate label.
2. The system of claim 1, wherein the system is configured to: Also include: Human-computer interaction and confirmation module; This module presents the thinning replacement suggestion, supplementary label suggestion and content vacuum prompt optimization suggestion to the label system administrator in the form of visual interface and list.
3. The system of claim 1, wherein the system is configured to: Also include: effect evaluation and feedback closed loop module, for continuously monitoring the application effect of the business system after the implementation of optimization; this module generates quantitative analysis report for the effect comparison data in a period of time, and submits it to the administrator again, and the administrator judges whether this optimization is successful, needs further improvement or fails; form a "analysis-optimization-evaluation-re-optimization" continuous improvement closed loop.
4. The system of claim 3, wherein the system is configured to: The effect evaluation and feedback closed loop module collects indexes including: Search accuracy / recall rate; use the labels before and after optimization to search, compare the relevance of the results; Search return rate; the proportion of users who modify the search again after performing the search, which decreases indicates that the robustness is improved; Label use concentration / coverage rate; analyze whether the distribution of label use is healthier.
5. A dynamic optimization method based on label constellation density analysis, characterized in that, Include the following processes: Step 1: data collection; extract all labels and label-entity association data from the business system; Step 2: build label star cloud; calculate the association probability between all labels and convert it into distance weight to build graph data structure; Step 3: set reference system and analyze density; Calculate and identify the dense and sparse areas in the star cloud; Step 4: generate optimization suggestions; generate thinning scheme for replacing redundant labels around the core label for dense areas; generate intermediate label supplement scheme for connecting isolated label groups or business system content vacuum prompt for sparse areas; Step 5: Artificially review and confirm; submit the suggestion to the administrator for artificial review, correction and final confirmation; Step 6: Implement optimization; update the confirmed scheme to the label system of the business system.
6. The dynamic optimization method based on label constellation density analysis according to claim 5, wherein, After step 6, the following processes are also included: Step 7: Effect monitoring and evaluation; after the scheduled evaluation period, collect the performance indicators of the business system, and compare and analyze before and after optimization; Step 8: Feedback and decision; Generate an evaluation report, and the administrator decides the next action based on the data, including: keep, deepen optimization or rollback, end the process or jump to step S101 to start a new round of optimization cycle.
7. The dynamic optimization method based on label cloud density analysis according to claim 5, wherein, The distance calculation method in step 2 is as follows: D(A, B) = 1 - P(A∩B) / sqrt(P(A) * P(B)); Where P(A∩B) is the probability of label A and B being associated with the same entity at the same time, and P(A) and P(B) are the independent association probabilities of label A and B respectively.
8. The dynamic optimization method based on label cloud density analysis according to claim 5, wherein, The core label in step 4 is the node with the highest degree centrality; the intermediate label is calculated by NLP word vector to find the semantic midpoint, or find the associated concept from the knowledge graph.
9. The dynamic optimization method based on label cloud density analysis according to claim 5, wherein, The reference system is set according to the excellent sample reference or uniform distribution reference; The excellent sample is a set of excellent labels with high quality recognized by experts, and the distribution characteristics of this set in the star cloud are used as the ideal density reference; The uniform distribution reference is to automatically find a "label group" in the star cloud with the most uniform label distribution and the best connectivity, and use its density as a reference benchmark.
10. A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the optimization method of any one of claims 5-9.