Cosmetic global regulation intelligent compliance examination system based on knowledge graph

Abstract

According to the cosmetic global regulation intelligent compliance examination system based on the knowledge graph, the data acquisition module acquires multi-dimensional data related to cosmetics, the graph construction module constructs the knowledge graph according to the preprocessed data, the multi-dimensional examination module carries out semantic matching on product information and regulation terms based on the SimBERT model, and the knowledge graph is obtained. An examination rule base is constructed based on the knowledge graph, and automatic compliance judgment is carried out through rule reasoning; processing fuzzy expressions in regulatory terms through a probabilistic reasoning model, and outputting to obtain compliance probability; and integrating the deterministic result of the rule reasoning and the fuzzy expression probability of the probability reasoning, and generating a unified compliance probability score and risk level. According to the invention, compliance automatic examination of the full life cycle of cosmetics can be realized, and compliance examination efficiency and compliance judgment accuracy are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a cosmetic global regulation intelligent compliance review system based on a knowledge graph. BACKGROUND

[0002] With the intensification of global layout of the cosmetic industry, different countries or regions have formulated differentiated regulations on the use of raw materials, labeling, safety testing, and claims of cosmetics, such as the EU REACH regulation, the US FDA regulations, and the Chinese Cosmetics Supervision and Management Regulations. In the process of product research and development, production, import and export, enterprises need to spend a lot of manpower and resources to sort out compliance requirements. Traditional manual review methods have low efficiency, are prone to omissions, and are difficult to keep up with regulatory updates.

[0003] Therefore, it is urgent to develop a system that can integrate multi-regional regulation data, product data, and raw material data, and realize automatic review of the compliance of cosmetics throughout their life cycle. SUMMARY

[0004] To solve the above problems, the present application provides a cosmetic global regulation intelligent compliance review system based on a knowledge graph, which constructs a cosmetic global regulation intelligent compliance review system based on knowledge graph technology. By integrating multi-dimensional data of cosmetics, based on knowledge modeling, intelligent reasoning, and semantic matching technologies, automatic review of the compliance of cosmetics throughout their life cycle is realized, effectively improving the compliance efficiency of enterprises and reducing compliance risks.

[0005] To achieve the above purpose, the technical solution adopted by the present application is: The present application provides a cosmetic global regulation intelligent compliance review system based on a knowledge graph, comprising: A data acquisition module for acquiring and preprocessing multi-dimensional data related to cosmetics, including regulation data, raw material data, product data, and industry data; A graph construction module in communication with the data acquisition module for constructing a knowledge graph based on the preprocessed data; A multi-dimensional review module in communication with the graph construction module, based on the constructed knowledge graph, for compliance review through semantic matching and knowledge reasoning. The multi-dimensional review module includes a semantic matching unit, a knowledge reasoning unit, and a compliance review unit. The knowledge reasoning unit is used to integrate the rule reasoning results and the probability reasoning results of the knowledge graph to generate a unified compliance probability score and risk level. The compliance review unit is used to generate a compliance review report.

[0006] As preferred, the semantic matching unit performs semantic matching on the product information and the regulatory provisions based on a SimBERT model, captures semantic similarity of the text through contrastive learning optimization; the SimBERT model adopts a Transformer architecture, introduces a sparse attention module in a Transformer encoder layer, limits the attention calculation range through a mask; a phrase encoding layer is added between a word embedding layer and an attention layer, converts a word sequence into a mixed sequence of words and phrases, calculates semantic association between phrases through a phrase-level attention head; the attention layer outputs fusion features calculated through linear projection after concatenation of global attention, sparse attention and phrase-level attention.

[0007] As preferred, the fused features are calculated as: Output = Dropout(GELU(Concat(G inter ,S inter ,P inter )×W+b)); wherein, Output represents the fused features, G inter represents local information of the global attention fused with the sparse attention, S inter represents core information of the sparse attention fused with the phrase-level attention, P inter represents context information of the phrase-level attention fused with the global attention, Concat(G inter ,S inter ,P inte ) represents the concatenated high-dimensional features, W represents a weight matrix of the linear projection layer, b represents a bias vector of the linear projection layer, and GELU represents an activation function.

[0008] As preferred, the knowledge reasoning unit comprises a rule reasoning subunit, a probabilistic reasoning subunit and a reasoning integration subunit. The rule reasoning subunit constructs a review rule library based on a knowledge graph, and performs automatic compliance judgment through rule reasoning. The probabilistic reasoning subunit is configured to process fuzzy expressions in the regulatory provisions through a probabilistic reasoning model, and output a compliance probability. The reasoning integration subunit is configured to integrate the deterministic result of the rule reasoning and the fuzzy expression probability of the probabilistic reasoning, and generate a unified compliance probability score and a risk level.

[0009] As preferred, the review rule library is constructed based on the knowledge graph, and the automatic compliance judgment is performed through rule reasoning, comprising: Based on the regulatory provisions in the knowledge graph, the review rule library is constructed by classifying according to the compliance review dimensions, and the review rule library comprises raw material restriction rules, label requirement rules, claim specification rules, detection item rules and production process rules. Filtering applicable regulations and clauses from the regulation library according to the product information entered or uploaded by the user, the target market selected, and prioritizing them; Converting the filtered rules into Cypher query statements based on the reasoning engine, submitting them in batches to the Neo4j graph database to match the entity relationships in the knowledge graph, and performing logical judgments; Aggregating the reasoning results of all rules, classifying them into compliance items and non-compliance items, and labeling the corresponding regulation names, clause IDs, and constraint contents.

[0010] As a preferred embodiment, the processing of fuzzy expressions in the regulation clauses through the probabilistic reasoning model includes: Identifying common types of fuzzy expressions in cosmetic regulations, defining the modeling scope, and the fuzzy expression types include raw material usage amount ambiguity, claim expression ambiguity, detection requirement ambiguity, and applicable scope ambiguity; Based on the knowledge graph and historical compliance cases, extract the core features that affect the determination of fuzzy expressions; Based on the random forest classifier combined with the Bayesian network, a probabilistic reasoning model is constructed and encapsulated as an API service; The reasoning engine calls the probabilistic reasoning model through a network request and outputs the compliance probability while labeling the key influencing factors.

[0011] As a preferred embodiment, it also includes an application module that is communicatively connected with the data collection module, the graph construction module, and the multi-dimensional review module, and is used for visualizing the knowledge graph and risk level scoring, having retrieval and viewing functions, and real-time monitoring of regulation updates and system management.

[0012] As a preferred embodiment, the application module includes a knowledge query unit, a regulation update unit, a risk warning unit, and a system management unit; Knowledge query unit: used for providing knowledge graph visualization query function, users can obtain relevant knowledge information through retrieval and query, and support associated display and detail viewing of graph nodes; Regulation update unit: used for real-time monitoring of global cosmetic regulation update dynamics, classified display by region and regulation type, supporting user subscription of specific regulation update reminders, and providing update content comparison function; Risk warning unit: used for displaying product compliance risk level scoring and marking high-risk items, combining regulation update trends to predict future compliance risks, and supporting user viewing of risk reasons and response suggestions; System management unit: used for providing user management, data management, log management, and parameter configuration functions to ensure stable system operation.

[0013] As preferred, the security module is used to guarantee data security, encrypt sensitive data based on the AES encryption algorithm for storage and transmission, establish a data access control mechanism, assign different users' data access permissions based on the RBAC model, regularly conduct data backup and disaster recovery drills to prevent data loss; application security, implement interface authentication, attack prevention measures, log user operations for security audit and abnormal tracking; network security, configure firewall, intrusion detection system, intrusion prevention system, restrict illegal network access, and use VPN technology to ensure secure access for remote users.

[0014] As preferred, Prometheus and Grafana are used for system monitoring, monitoring indicators are set, including system indicators, application indicators and data indicators, and monitoring thresholds are set, and the system automatically triggers an alarm when the indicators exceed the thresholds.

[0015] The beneficial effects of the present application are: Improve compliance review efficiency and reduce labor costs: The system integrates data collection, knowledge modeling and intelligent reasoning into a full-process automation, replacing the traditional manual review of regulatory documents and the comparison of product information clause by clause, supporting multi-product batch compliance review, and the user only needs to upload product information and select the target market, and the system automatically completes multi-dimensional review and generates a report, improving review efficiency.

[0016] Improve compliance judgment accuracy and reduce omissions and misjudgments: Through the knowledge graph, the clauses, raw material restrictions and claim specifications of global multi-region regulations are structured and associated, solving the problem of omissions caused by scattered and overlapping regulations; based on the optimized SimBERT model, the semantic precise matching of product claims, label content and regulatory clauses is realized, avoiding misjudgments caused by synonym replacement and ambiguous expressions; based on rule-based reasoning for explicit clauses and probabilistic reasoning for ambiguous expressions, the compliance probability is quantified to improve the accuracy of judgment.

[0017] Dynamic adaptation to regulatory updates to reduce compliance risks: By monitoring the official websites of global regulatory agencies and authoritative databases through crawlers, automatically detecting regulatory updates and synchronizing them to the knowledge graph, users can subscribe to update reminders to avoid compliance risks caused by not following new regulations. After the regulations are updated, the system automatically analyzes the impact on the compliance of existing products and generates a comparison report to help enterprises quickly adjust product formulations, claims or labels to reduce compliance improvement costs. Integrate raw material risk levels, regulatory update frequencies, historical compliance cases and other data to provide early warning of product compliance risks and provide rectification suggestions to turn post-rectification into pre-prevention. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is a knowledge graph-based global cosmetic regulation intelligent compliance review system of the present application.

[0019] Figure 2 is a flow chart of the knowledge graph construction of the present application.

[0020] Figure 3 is a flow chart of the rule reasoning subunit of the present application.

[0021] Figure 4 is a flow chart of the probability reasoning subunit of the present application. DETAILED DESCRIPTION

[0022] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purposes, the specific embodiments, structures, features and effects according to the present application are described in detail below in combination with the drawings and preferred embodiments.

[0023] The system adopts a cloud-native deployment architecture, is based on Docker containerization technology and Kubernetes container orchestration technology, realizes elastic expansion and high-availability deployment of the system. Pack each functional module of the system into a Docker image, store it in an image repository, manage Docker containers through Kubernetes, realize automatic deployment, scaling, load balancing and fault recovery of containers. For example, when the compliance review request volume surges, Kubernetes automatically increases the number of container instances of the compliance review module to improve the system's concurrent processing capacity.

[0024] According to the user distribution, deploy system instances in multiple cloud regions, accelerate front-end resource access through CDN, and reduce network latency. Deploy MySQL database using master-slave replication architecture, master database is responsible for data writing, slave database is responsible for data reading, improve database concurrent processing capacity, Neo4j uses cluster deployment method, realizes data sharding storage and high availability.

[0025] Referring to Figure 1 The present application provides a cosmetic global regulation intelligent compliance review system based on a knowledge graph, which comprises: A data acquisition module is used to acquire multi-dimensional data related to cosmetics and pre-process the data, including regulation data, raw material data, product data and industry data. Regulation data, raw material data, product data and industry data are collected through crawling, API docking, manual input and other methods and stored in an original data warehouse. Invalid characters, duplicate data and data with too many missing values in the original data are removed, and abnormal data is marked and corrected. The data format, unit and name are unified. Unstructured data is converted into semi-structured data, semi-structured data is converted into structured data, and pre-processed structured data is stored in a relational database, and semi-structured / unstructured data is stored in an object storage.

[0026] Knowledge graph construction module: Communicates with the data acquisition module and is used to construct a knowledge graph based on the preprocessed data; Please refer to Figure 2 As shown, the construction of a knowledge graph includes: Knowledge modeling: Define the ontology structure of the knowledge graph, and clarify the core entities, relationships and attributes; Based on the characteristics of the cosmetics regulatory field, Protégé was used to perform ontology modeling of a knowledge graph. Core entities include regulations, clauses, raw materials, products, brands, testing items, and risk levels; core relationships include regulations containing clauses, clauses restricting raw materials, products containing raw materials, corresponding risk levels of raw materials, and required testing items; entity attributes include the publication date / effective date / applicable region of the regulation, INCI name / CAS number / usage limit of the raw material, and product type / formula / manufacturing process.

[0027] Knowledge extraction: Knowledge is extracted from multi-source data by combining rule-based and deep learning methods; For structured data (such as Excel raw material catalogs and database tables), entities and attributes are directly extracted using rule matching algorithms, such as extraction based on the field mapping rule of "raw material name → CAS number → usage limit".

[0028] For semi-structured data (such as regulatory webpage forms and XML documents), XPath / CSS selector algorithms are used to locate key information and extract entities and relationships, such as extracting "clause number → restricted raw materials → limit value" from regulatory forms.

[0029] For unstructured data (such as regulatory texts and test reports), a remote supervision technique is used to construct a seed dataset based on existing knowledge rules. The seed dataset is used as a training sample, and the core entities are identified based on the CRF model. The identified entities are combined with text fragments, and the PCNN model is trained with the relation labels in the seed dataset as labels. This captures the contextual semantic features of entity pairs in the text, learns the mapping rules between entity pairs and relation types, and extracts complex relationships between entities from the text.

[0030] Specifically, unstructured text is cleaned, segmented, and tagged with parts of speech, while stop words are filtered to generate a standardized text corpus. Based on the characteristics of the cosmetics regulatory domain, core "entity-relationship" rule templates are compiled as the annotation basis for remote supervision. The compiled knowledge rules are used to automatically annotate the preprocessed text corpus, including locating key entities in the text according to the rule templates, annotating the correspondence between entities, and initially screening the automatic annotation results to form a seed dataset.

[0031] Using entity annotations from the seed dataset as training samples, a CRF (Conditional Random Field) model is trained. The model input consists of the segmented text sequence, part-of-speech tags, and contextual features; the model output is the entity type label for each word in the text. The trained CRF model is then used to process all preprocessed unstructured text, accurately identifying the core entities within the text.

[0032] The identified entities are combined with text fragments to generate training samples of "entity pairs and contextual text". Using the relation annotations in the seed dataset as labels, a PCNN (Piecewise Convolutional Neural Network) model is trained to capture the contextual semantic features of entity pairs in the text and learn the mapping rule of "entity pair → relation type". The trained PCNN model is then used to infer relations between all identified entity pairs in the text, outputting the accurate relations between entities.

[0033] Low-confidence predictions from the model are removed (e.g., associations with a prediction probability < 0.7), while high-risk annotations (e.g., relation inferences under fuzzy semantics) are manually validated. Synonymous entities are merged, and the predicted relations are mapped to standard relations defined by the knowledge graph, ultimately outputting structured triples of "entity 1-relation-entity 2".

[0034] Knowledge fusion: Extracted knowledge is disambiguated, attribute aligned, and knowledge conflict resolved, and integrated into unified knowledge triples to generate a knowledge graph; A similarity-based entity disambiguation algorithm is employed to calculate the name similarity and attribute similarity of entities from different data sources, merging duplicate entities. The attribute format across different data sources is standardized, and attributes with inconsistent units are converted. For attribute conflicts (such as inconsistent raw material limits from different sources), priority is assigned based on factors such as the publication date of regulations and their level of authority (e.g., official website of regulatory agencies > third-party databases), retaining highly authoritative and up-to-date attribute information and indicating the source of the conflict.

[0035] Knowledge storage: A hybrid storage method combining graph databases and relational databases is used; The Neo4j database is used as the core graph database, and "entity-relationship-attribute" triples are imported to support efficient association queries and graph reasoning. At the same time, a MySQL database is used to store the knowledge graph's metadata, user permission data and other structured business data, and the data consistency between the two is ensured through a data synchronization mechanism.

[0036] Knowledge Updates: Establish a dynamic knowledge update mechanism that supports both automatic and manual updates.

[0037] By web crawling and monitoring regulatory updates on the websites of regulatory agencies in various countries, the system automatically triggers knowledge extraction and fusion processes to update relevant clauses and restrictions in the knowledge graph. For knowledge errors or omissions discovered manually, manual review and editing are supported to ensure the timeliness and accuracy of the knowledge graph.

[0038] The constructed cosmetics regulatory knowledge graph supports multi-dimensional retrieval of entities, relationships, and clauses, and visually displays node connections in graph form, such as a link display of "regulation → clause → restricted raw materials → risk level," helping users quickly locate target knowledge and replacing the inefficient method of manually flipping through regulatory documents. It receives regulatory update data in real time and automatically updates the clause content and restriction requirements in the graph to ensure that compliance review standards are consistent with the latest regulations. It supports storing knowledge by region and regulatory type, and can quickly adapt to the compliance review needs of different target markets (such as the EU, the US, and China). It integrates knowledge such as raw material risk levels, regulatory update frequency, and historical compliance cases to provide feature data for the risk warning module, supporting the training of machine learning models such as random forests.

[0039] Multi-dimensional review module: Communicates with the knowledge graph construction module. Based on the constructed knowledge graph, it conducts compliance review through semantic matching and knowledge reasoning. The multi-dimensional review module includes a semantic matching unit, a knowledge reasoning unit, and a compliance review unit. The knowledge reasoning unit is used to integrate the rule reasoning results and probability reasoning results of the knowledge graph to generate a unified compliance probability score and risk level. The compliance review unit is used to generate a compliance report.

[0040] The multi-dimensional review module includes a semantic matching unit, a knowledge reasoning unit, and a compliance review unit; Semantic matching unit: Based on the SimBERT model, it performs semantic matching between product information and regulatory clauses, and optimizes through comparative learning to capture the semantic similarity of text; The SimBERT model adopts the Transformer architecture, in the Transformer encoder layer. A sparse attention module is introduced to limit the scope of attention calculation through masking; a phrase encoding layer is added between the word embedding layer and the attention layer to convert the word sequence into a mixed sequence of words and phrases, and the semantic association between phrases is calculated through phrase-level attention heads; the output of the attention layer is a fusion feature calculated by concatenating global attention, sparse attention and phrase-level attention through linear projection.

[0041] SimBERT uses 12 attention heads by default, retains 4 global attention heads, and replaces the remaining 8 with sparse attention heads guided by local windows (each word only focuses on the k words before and after it) and key entities (additionally focusing on domain key entities appearing in the text, regardless of whether the entity is within the local window). By limiting the scope of attention calculation through masking, the amount of computation is reduced and the encoding efficiency is improved.

[0042] A phrase encoding layer is added between the word embedding layer and the attention layer, which adds four phrase-level attention heads (in parallel with the sparse / global head). This layer only calculates the attention between phrase vectors and ignores redundant associations between words within a phrase. The final input to the encoder layer is a fused vector of word embeddings and phrase encodings.

[0043] The fused features are calculated as follows: Output = Dropout(GELU(Concat(G)) inter ,S inter ,P inter )×W+b)); Where Output represents the fused features, G inter S represents the fusion of local information from global attention and sparse attention. inter P represents the core information of sparse attention fusion phrase-level attention. inter This indicates that phrase-level attention is fused with the contextual information of global attention. Concat(G) inter ,S inter ,P inte ) represents the high-dimensional features after concatenation, W represents the weight matrix of the linear projection layer, b represents the bias vector of the linear projection layer, and GELU represents the activation function.

[0044] To enhance the correlation between features and enable them to more comprehensively reflect the global, local, and phrase-level relationship logic, thus better adapting to the compliance requirements and product detail scenarios of cosmetics, the three attention features are first interactively processed and then concatenated, as shown in the formula: G inter =G + S·0.5; S inter =S + P·0.5; P inter =P + G·0.5; In the formula, G represents the global attention output feature, S represents the sparse attention output feature, and P represents the phrase-level attention output feature.

[0045] Specifically, product information and regulatory clauses from the cosmetics industry are collected as matching samples, labeled as matching / non-matching, and positive and negative sample pairs are constructed. These labeled samples are then input into the SimBERT model for fine-tuning and training. The fine-tuned SimBERT model transforms the preprocessed product information and regulatory clauses into high-dimensional semantic vectors, achieving a mapping from text to semantic features. SimBERT's Transformer encoder captures the contextual semantics of the text word by word, ultimately outputting vector forms of product information and regulatory clauses. The cosine similarity between the two semantic vectors is calculated to determine the semantic fit between the product information and the regulatory clauses. Based on the similarity calculation results, combined with preset thresholds and business rules, a matching conclusion is output. The matching results are deduplicated and filtered for low confidence (e.g., results with similarity between 0.7 and 0.75 may be prompted for manual review) to ensure the accuracy of the review results.

[0046] For example, considering the accuracy requirements of cosmetic compliance scenarios, a similarity threshold of 0.7 is preset. If the similarity is ≥0.7, it is judged as "semantic match successful", indicating that the product information complies with the constraints of the regulations (it may be compliant or non-compliant, and needs to be judged in conjunction with the type of regulations); if the similarity is <0.7, it is judged as "semantic mismatch", indicating that the product information is not directly related to the regulations.

[0047] If the clause is a prohibition clause, a successful match will be marked as "non-compliant" and associated with the corresponding legal clause as the basis; if the clause is a requirement clause, a successful match will be marked as "compliant"; a failed match will be marked as "non-compliant".

[0048] Knowledge Reasoning Unit: Used to integrate the rule-based reasoning results and probabilistic reasoning results of the knowledge graph to generate a unified compliance probability score and risk level; The knowledge reasoning unit includes a rule-based reasoning subunit, a probability-based reasoning subunit, and a reasoning integration subunit; The rule reasoning subunit constructs an audit rule base based on a knowledge graph and automatically judges compliance through rule reasoning; Please refer to Figure 3 As shown, the step of constructing a review rule base based on a knowledge graph and automatically judging compliance by combining rule reasoning includes: Based on the legal provisions in the knowledge graph, a review rule base is constructed according to the dimensions of compliance review. The review rule base includes rules for raw material restrictions, rules for labeling requirements, rules for claim specifications, rules for testing items, and rules for production processes. Rule reasoning is implemented using Neo4j's Cypher query language. Rules are represented in a ternary form, consisting of natural language description, Cypher query statements, and rule metadata, ensuring that rules can be directly mapped to the query logic of a graph database. The rule metadata is used for rule management, including information such as rule ID, applicable region, regulatory version, effective date, priority (regulatory authority level > effective date > applicable scenario), and rule type (raw material restriction type).

[0049] Based on the product information entered or uploaded by the user and the selected target market, applicable regulations and clauses are selected from the review rule base and sorted by priority. Based on the inference engine, the filtered rules are transformed into Cypher query statements, which are then submitted in batches to the Neo4j graph database to match entity relationships in the knowledge graph and execute logical judgments. For example, a prohibited rule: if a product contains mercury, it matches the rule in China's "Cosmetics Supervision and Administration Regulations" that prohibits the addition of mercury to cosmetics. The reasoning result is non-compliant, and the associated clause ID and content are linked.

[0050] Aggregate the reasoning results of all rules, categorize them into compliant and non-compliant items, and label the corresponding law name, clause ID, and constraint content.

[0051] The probabilistic reasoning subunit is used to process the ambiguous expressions in the legal provisions through a probabilistic reasoning model and output the compliance probability. Please refer to Figure 4 As shown, the process of using a probabilistic reasoning model to process ambiguous expressions in regulatory clauses and output compliance probabilities includes: This paper identifies common types of vague expressions in cosmetic regulations and clarifies the scope of modeling. These types of vague expressions include vague raw material usage, vague claims, vague testing requirements, and vague scope of application. Based on knowledge graphs and historical compliance cases, core features affecting the determination of fuzzy expressions are extracted. For example, in the fuzzy analysis of raw material usage, the core features extracted include product-related features, raw material-related features, regulatory-related features, and historical case features.

[0052] A probabilistic inference model is built based on a random forest classifier combined with a Bayesian network, and then packaged as an API service. Specifically, a Bayesian network is used to capture the causal dependencies between features in cosmetic compliance scenarios and to handle the uncertainty of feature associations. This involves employing domain knowledge and a scoring search method: first, core causal edges are determined based on compliance scenario experience; then, the network structure is optimized using the Bayesian Information Criterion (BIC) scoring method. For each sample, the model outputs a prior compliance probability. Then, a random forest is used to process multi-dimensional nonlinear features. Using the preprocessed feature matrix and compliance labels as input, 5-fold cross-validation is used to optimize hyperparameters. For each input sample, a posterior compliance probability is output, along with the feature importance. The weighted fusion of the prior and posterior compliance probabilities yields the precise compliance probability.

[0053] The inference engine calls the probabilistic inference model through network requests, outputs the compliance probability, and labels the key influencing factors.

[0054] The reasoning integration subunit is used to integrate the deterministic results of rule-based reasoning with the fuzzy probability representations of probabilistic reasoning to generate a unified compliance probability score and risk level.

[0055] Specifically, rule-based reasoning directly maps compliance to probability 1 and non-compliance to probability 0; probabilistic reasoning uses compliance probabilities between 0 and 1. Based on the probability score, five risk levels are defined, including: Low risk [0.9, 1.0], fully compliant with regulatory requirements, no rectification required, and the listing process can proceed normally; Medium to low risk [0.7, 0.9), highly compliant, with a minor possibility of non-compliance. It is recommended to review the raw material content / claims. No major rectification is required. Medium risk [0.5, 0.7], adjust raw material content / modify claims, and resubmit for review; Medium to high risk [0.3, 0.5], must be rectified (such as replacing raw materials, supplementing test reports), and re-examined after rectification; High risk [0.0, 0.3], prohibited from market launch, product formula / claims system to be redesigned.

[0056] Compliance review unit: Used to generate compliance review reports based on semantic matching results and knowledge reasoning results, and push them to users.

[0057] The scattered semantic matching results, knowledge reasoning results, and basic product information are standardized and integrated. The semantic matching results are matched one-to-one with the knowledge reasoning results. Non-compliant items are sorted from high to low risk level (high-risk items are displayed first), while compliant items are classified according to review dimensions (raw material compliance, label compliance, claim compliance, etc.) to facilitate quick browsing for users.

[0058] The report includes: basic product information, such as product name, product type, target market, review time, and report number; a compliance review overview, such as the number of compliant items, the number of non-compliant items, risk level distribution, and overall compliance probability score; detailed review results, listed separately for compliant and non-compliant items, each including review dimensions, specific content, judgment basis, and risk level (non-compliant items only); risk warning details, with high / medium-high risk items highlighted, explaining the reasons for the risk and the possible consequences; rectification suggestions, providing specific and actionable suggestions for each non-compliant item; and a regulatory basis appendix, listing all relevant regulatory names, clause IDs, original text of the clauses, and effective dates, with support for viewing the knowledge graph relationships.

[0059] The compliance report supports multiple output formats to meet different user scenarios, and pushes reports according to user preset preferences to ensure timely receipt. It also supports multi-channel push notifications. The system provides related operational functions for compliance reports, such as: resubmitting for review – after modifying product information according to rectification suggestions, users can directly initiate a second review through the "Resubmit for Review" button on the report page, and the system automatically links to historical reports; providing feedback – users can raise objections to the report results, which triggers a manual review process, and the review results will be notified to the user via push notification; and tracing the review history – the system records all review reports for this product (sorted by time), allowing users to view and compare historical reports.

[0060] It also includes an application module, which is connected to the data acquisition module, the knowledge graph construction module, and the multi-dimensional review module. This module is used to visualize the knowledge graph and risk level scoring, and has search and viewing functions, as well as real-time monitoring of regulatory updates and system management.

[0061] The application module includes a knowledge query unit, a regulation update unit, a risk warning unit, and a system management unit; Knowledge Query Unit: This unit provides a visual query function for the knowledge graph, allowing users to obtain relevant knowledge information through retrieval and query methods. It also supports the display of related graph nodes and detailed viewing of their information. Regulatory Update Unit: Used to monitor global cosmetic regulatory updates in real time, categorized and displayed by region and regulatory type, supports users to subscribe to update reminders for specific regulations, and provides a function to compare updated content; The system selects official websites of major global cosmetic regulatory agencies (such as the US FDA, EU ECHA, China NMPA, Japan's Ministry of Health, Labour and Welfare, and Korea's Ministry of Food and Drug Safety) and authoritative regulatory databases (such as LexisNexis and Westlaw) as monitoring data sources. It employs web crawling combined with webpage change detection technologies (such as hash value comparison and text similarity comparison) to periodically crawl the webpage content of the monitored data sources and compare it with historical versions to detect any regulatory updates. When a regulatory update is detected, the system automatically extracts the updated content (such as updated clauses, effective dates, and scope of application), triggering a knowledge graph update process to update relevant knowledge. Simultaneously, the updated content is categorized and organized by region and regulatory type, generating regulatory update notifications and pushing them to users subscribed to the relevant regulations via system messages, emails, and SMS. Users can view update details, comparison results with older versions, and an analysis of the update's impact on the compliance of existing products through the system.

[0062] Risk warning unit: Used to display the product compliance risk level score and mark high-risk items, predict future compliance risks based on regulatory update trends, and support users to view the causes of risks and corresponding response suggestions; When a user submits a product for compliance review, in addition to outputting the review results, the system also calculates the product's risk level through a risk model, marks high-risk items (such as the use of high-risk prohibited or restricted raw materials, the adoption of absolute claims that are prone to violations, etc.), and pushes risk warnings to the user through system messages, emails, and other means, and provides risk response suggestions (such as replacing raw materials, modifying claims, supplementing test reports, etc.).

[0063] System Management Unit: Provides functions for user management (role assignment, access control), data management (data backup, data import and export), log management (operation logs, review logs, system logs), and parameter configuration (target market configuration, algorithm parameter configuration) to ensure stable system operation.

[0064] It also includes a security module, which is used to ensure data security by encrypting and storing sensitive data based on the AES encryption algorithm, establishing a data access control mechanism, allocating data access permissions for different users based on the RBAC model, and regularly performing data backups and disaster recovery drills to prevent data loss; application security by implementing interface authentication and anti-attack measures, logging user operations for easy security auditing and anomaly tracking; and network security by configuring firewalls, intrusion detection systems, and intrusion prevention systems to restrict unauthorized network access and using VPN technology to ensure secure access for remote users.

[0065] Prometheus and Grafana are used for system monitoring. Monitoring metrics include: system metrics such as server CPU utilization, memory utilization, disk utilization, and network bandwidth utilization; application metrics such as module response time, API call success rate, concurrent requests, and error rate; and data metrics such as knowledge graph data volume, data update frequency, number of review reports generated, and user activity. Monitoring thresholds are set. When a metric exceeds the threshold (e.g., server CPU utilization ≥ 80%, API call success rate ≤ 95%), the system automatically triggers an alarm (via email, SMS, WeChat Work, etc.), allowing maintenance personnel to address the issue promptly.

[0066] A scheduled backup strategy is employed, performing a full backup of the MySQL database, Neo4j graph database, and object storage every morning at midnight, and incremental backups every hour. Backup data is stored at an off-site backup center using encrypted storage to prevent data leakage. A disaster recovery plan is developed; in the event of a major system failure (such as server downtime or data loss), the appropriate recovery process is initiated based on the failure type. For example, in the event of database data loss, recovery is performed using off-site backup data, with a recovery time objective (RTO) of ≤4 hours and a recovery point objective (RPO) of ≤1 hour.

[0067] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A knowledge graph-based intelligent compliance review system for global cosmetic regulations, characterized in that: include: Data acquisition module: used to collect multi-dimensional data related to cosmetics and preprocess it, including regulatory data, raw material data, product data and industry data; Knowledge graph construction module: Communicates with the data acquisition module and is used to construct a knowledge graph based on the preprocessed data; Multi-dimensional review module: Communicates with the knowledge graph construction module. Based on the constructed knowledge graph, it conducts compliance review through semantic matching and knowledge reasoning. The multi-dimensional review module includes a semantic matching unit, a knowledge reasoning unit, and a compliance review unit. The knowledge reasoning unit is used to integrate the rule reasoning results and probability reasoning results of the knowledge graph to generate a unified compliance probability score and risk level. The compliance review unit is used to generate a compliance review report.

2. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 1, characterized in that, The semantic matching unit performs semantic matching between product information and regulatory clauses based on the SimBERT model. It optimizes through contrastive learning to capture the semantic similarity of the text. The SimBERT model adopts the Transformer architecture, introducing a sparse attention module in the Transformer encoder layer and limiting the scope of attention calculation through masking. A phrase encoding layer is added between the word embedding layer and the attention layer to convert the word sequence into a mixed sequence of words and phrases. The semantic association between phrases is calculated through phrase-level attention heads. The output of the attention layer is a fusion feature obtained by concatenating global attention, sparse attention, and phrase-level attention and calculating it through linear projection.

3. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 2, characterized in that, The fused features are calculated as follows: Output=Dropout(GELU(Concat(G inter ,S inter ,P inter )×W+b)); Where Output represents the fused features, G inter S represents the fusion of local information from global attention and sparse attention. inter P represents the core information of sparse attention fusion phrase-level attention. inter This indicates that phrase-level attention is fused with the contextual information of global attention. Concat(G) inter ,S inter ,P inte ) represents the high-dimensional features after concatenation, W represents the weight matrix of the linear projection layer, b represents the bias vector of the linear projection layer, and GELU represents the activation function.

4. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 1, characterized in that, The knowledge reasoning unit includes a rule-based reasoning subunit, a probability-based reasoning subunit, and a reasoning integration subunit; The rule reasoning subunit constructs an audit rule base based on a knowledge graph and automatically judges compliance through rule reasoning; The probabilistic reasoning subunit is used to process the ambiguous expressions in the legal provisions through a probabilistic reasoning model and output the compliance probability. The reasoning integration subunit is used to integrate the deterministic results of rule-based reasoning with the fuzzy probability representations of probabilistic reasoning to generate a unified compliance probability score and risk level.

5. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 4, characterized in that, The method of constructing a review rule base based on knowledge graphs and automatically judging compliance by combining rule reasoning includes: Based on the legal provisions in the knowledge graph, a review rule base is constructed according to the dimensions of compliance review. The review rule base includes rules for raw material restrictions, rules for labeling requirements, rules for claim specifications, rules for testing items, and rules for production processes. Based on the product information entered or uploaded by the user and the selected target market, applicable regulations and clauses are selected from the review rule base and sorted by priority. Based on the inference engine, the filtered rules are transformed into Cypher query statements, which are then submitted in batches to the Neo4j graph database to match entity relationships in the knowledge graph and execute logical judgments. Aggregate the reasoning results of all rules, categorize them into compliant and non-compliant items, and label the corresponding law name, clause ID, and constraint content.

6. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 4, characterized in that, The process of using a probabilistic reasoning model to process ambiguous expressions in regulatory clauses and output compliance probabilities includes: This paper identifies common types of vague expressions in cosmetic regulations and clarifies the scope of modeling. These types of vague expressions include vague raw material usage, vague claims, vague testing requirements, and vague scope of application. Based on knowledge graphs and historical compliance cases, core features affecting the determination of fuzzy expressions are extracted. A probabilistic inference model is built based on a random forest classifier combined with a Bayesian network, and then packaged as an API service. The inference engine calls the probabilistic inference model through network requests, outputs the compliance probability, and labels the key influencing factors.

7. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 1, characterized in that, It also includes an application module, which is connected to the data acquisition module, the knowledge graph construction module, and the multi-dimensional review module. This module is used to visualize the knowledge graph and risk level scoring, has search and viewing functions, and supports real-time monitoring of regulatory updates and system management.

8. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 7, characterized in that, The application module includes a knowledge query unit, a regulation update unit, a risk warning unit, and a system management unit; Knowledge Query Unit: This unit provides a visual query function for the knowledge graph, allowing users to obtain relevant knowledge information through retrieval and query methods. It also supports the display of related graph nodes and detailed viewing of their information. Regulatory Update Unit: Used to monitor global cosmetic regulatory updates in real time, categorized and displayed by region and regulatory type, supports users to subscribe to update reminders for specific regulations, and provides a function to compare updated content; Risk warning unit: Used to display the product compliance risk level score and mark high-risk items, predict future compliance risks based on regulatory update trends, and support users to view the causes of risks and corresponding response suggestions; System Management Unit: Provides user management, data management, log management, and parameter configuration functions to ensure stable system operation.

9. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 7, characterized in that, It also includes a security module, which is used to ensure data security by encrypting and storing sensitive data based on the AES encryption algorithm, establishing a data access control mechanism, allocating data access permissions for different users based on the RBAC model, and regularly performing data backup and disaster recovery drills to prevent data loss. Application security includes implementing interface authentication and attack prevention measures, logging user operations to facilitate security auditing and anomaly tracking; network security includes configuring firewalls, intrusion detection systems, and intrusion prevention systems to restrict unauthorized network access, and using VPN technology to ensure secure access for remote users.

10. The knowledge graph-based intelligent compliance review system for global cosmetic regulations as described in claim 9, characterized in that, The system is monitored using Prometheus and Grafana. Monitoring metrics include system metrics, application metrics, and data metrics. Monitoring thresholds are set, and the system automatically triggers alarms when a metric exceeds the threshold.

Images