A method and system for mining common neighbor relationships based on graph computing and massive data

Abstract

The application provides a method and system for mining common neighbor relationship of massive data based on graph calculation, wherein the method comprises the following steps: step 1, constructing a concept metadata system; step 2, collecting data from an information system to be mined, and obtaining instances of concepts according to the concept metadata system; step 3, uniformly encoding attributes in the instances; step 4, storing the encoding and the corresponding instances in a graph database in a distributed storage mode; and step 5, relationship mining, extracting instances from the distributed graph storage, discovering common neighbors between the instances according to the characteristics of the encoding, filtering according to preset rules, and generating recommended relationships; and the system is used for realizing the above method. The application can be widely applied to the fields of social network analysis and recommendation system optimization, and provides an efficient solution for relationship mining of massive data.

Description

Technical Field

[0001] This invention relates to a data mining method and system, and more particularly to a method and system for mining common neighbor relationships in massive datasets based on graph computation. Background Technology

[0002] Knowledge graphs describe concepts, entities, and their relationships in a structured form, expressing internet information in a way that more closely resembles human cognition. They provide a better ability to organize, manage, and understand the massive amounts of information on the internet. Knowledge graphs have revitalized internet semantic search and demonstrated powerful capabilities in intelligent question answering, becoming the infrastructure for intelligent knowledge applications on the internet. Along with big data and deep learning, knowledge graphs are one of the core driving forces propelling the development of the internet and artificial intelligence.

[0003] The construction of a knowledge graph involves two important steps: knowledge modeling and graph construction. Knowledge modeling aims to represent various types of knowledge in the real world into a structure that can be stored and computed by computers. Graph construction is divided into instance construction and relationship construction. Instance construction can migrate data from data sources to graph databases using a series of ETL tools. Compared to instance construction, relationship construction between instances is more challenging. Currently, there are roughly two mainstream approaches to relationship construction in the industry:

[0004] The first is the batch processing computation mode. This method requires traversing all nodes in the entire graph, using Spark / PySpark to process TB-level data, implementing duplicate value detection (entity disambiguation) with O(1) complexity through Bloom Filter, and using the Locality-Sensitive Hashing (LSH) algorithm to cluster similar entities (Jaccard similarity > 0.85). The batch processing computation mode has a time complexity of O(N²), which cannot support the mining of massive data (such as tens of millions of nodes) and cannot adapt to dynamic data updates (such as real-time addition of relationships in social networks), requiring repeated full computation.

[0005] Second is the index calculation mode. This method requires indexing the entire dataset, which is inefficient when dealing with large amounts of data. Enabling the index to quickly update in tandem with real-time dynamically updated data and non-high-cardinality attributes is also a significant challenge.

[0006] Therefore, how to quickly build relationships among massive instance data in graph databases and adapt to dynamic data updates has become an urgent problem to be solved in the graph construction process of knowledge graphs. Summary of the Invention

[0007] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method and system for mining common neighbor relationships based on graph computing and massive data, addressing the shortcomings of the existing technology.

[0008] To address the aforementioned technical problems, this invention discloses a method and system for mining common neighbor relationships based on graph computation and massive data, wherein the method includes the following steps:

[0009] Step 1: Construct a conceptual metadata system;

[0010] Step 2: Collect data from the information system to be mined, and obtain instances of the concept based on the conceptual metadata system constructed in Step 1;

[0011] Step 3: Unify encoding, and encode the attributes in the instance obtained in Step 2;

[0012] Step 4: Store the code and its corresponding instance in the graph database using a distributed storage method;

[0013] Step 5, Relationship Mining: Extract instances from distributed graph storage, discover common neighbors between instances based on encoding characteristics, and then filter rules according to preset rules to generate recommendation relationships.

[0014] Furthermore, the construction of the conceptual metadata system described in step 1 includes:

[0015] Step 1-1: Based on the table attributes, primary and foreign keys, and dependencies in the database of the information system to be mined, determine the main table and the relationships between tables.

[0016] Steps 1-2: Based on the main table and relationships, filter out candidate concepts to form a candidate concept set;

[0017] Steps 1-3: Define the attributes of the concept based on the fields contained in the table;

[0018] Steps 1-4: Determine the relationships between concepts through foreign key relationships in the table, and define object attributes based on these relationships;

[0019] Steps 1-5: Set constraints and select concepts from the candidate concept set; the selected concepts, their attributes, the relationships between concepts, and the object attributes of the concepts form a concept metadata system.

[0020] Furthermore, the data collection from the information system to be mined described in step 2 includes:

[0021] Full data import and incremental data import.

[0022] Furthermore, the full data import described in step 2 includes:

[0023] The data is accessed from the data source, i.e. the information system to be mined, and converted into instances of the concept in batches according to the concept metadata system built in step 1.

[0024] Furthermore, the incremental data transfer mentioned in step 2 includes:

[0025] Perform partial updates to the instance based on data changes in the data source.

[0026] Furthermore, the unified coding mentioned in step 3 includes:

[0027] Step 3-1, data cleaning, specifically includes: handling missing values, handling duplicate data, correcting erroneous data, handling outliers, data standardization and transformation, and data denoising;

[0028] Step 3-2: Filter high-cardinality attributes. Among the attributes in the instance, select the attributes whose probability of data duplication is less than the threshold as high-cardinality attributes.

[0029] Step 3-3: Unified encoding. Encode the high-cardinality attributes in the instance to obtain a string that uniquely corresponds to each high-cardinality attribute value.

[0030] Furthermore, step 4, which involves storing the encoding and corresponding instance in the graph database using a distributed storage method, includes:

[0031] Step 4-1: Add the code to be stored and its corresponding instance as a new node Δp in the graph database. The edge of this node is represented as (u,v); where u is the instance and v is the code.

[0032] Step 4-2: For each edge (u,v), use the graph routing algorithm to delete the neighbors of instance u;

[0033] Step 4-3: Store instance u as a new node Δp in the graph database, overwriting the original data;

[0034] Step 4-4: Store the edge (u,v) as a new edge in the graph database.

[0035] Furthermore, the relationship mining described in step 5 includes:

[0036] Step 5-1: Traverse the instances of the concept;

[0037] Step 5-2: Based on the graph routing method, query the adjacent instances that have common neighbors;

[0038] Step 5-3: Based on the starting point instance and the ending point instance, the starting point concept and the ending point concept are statistically obtained, the relationship is formed, and stored in the recommendation relationship table for subsequent manual review and use.

[0039] Furthermore, the threshold mentioned in step 3-2 is set to 60%.

[0040] This invention also proposes a system for mining association relationships based on graph computation and massive data common neighbors, used to implement the aforementioned method, including:

[0041] The module comprises a knowledge management module, a unified coding module, a relationship mining module, and a distributed graph storage module; among which,

[0042] The knowledge management module is used to construct a concept metadata system, collect data from the information system to be mined, and obtain instances of concepts based on the concept metadata system.

[0043] The unified encoding module encodes the attributes in the instances obtained from the knowledge management module;

[0044] The distributed graph storage module stores the code and its corresponding instance in the graph database using a distributed storage method;

[0045] The relationship mining module extracts instances from the distributed graph storage module, discovers common neighbors between instances based on the characteristics of the encoding, and then performs rule filtering based on preset rules to generate recommendation relationships.

[0046] Beneficial effects

[0047] 1. This invention cleans and uniformly encodes data and stores it in a distributed graph storage in the form of instance relationship triples. Through an efficient graph link algorithm, it quickly discovers instances with common neighbors (attribute values), avoids the complexity of cross-calculation of massive data, solves the space complexity problem of matching, and improves the parsing performance in matching calculation, thereby efficiently mining concept relationships.

[0048] 2. This invention uniformly encodes node data attribute values ​​during the data acquisition phase, and stores instance data along with the encoding into neighbor pairs in the form of relationships in the graph storage system. This significantly reduces the complexity of subsequent full-scale relationship mining and simplifies the processing of real-time incremental data. Compared to traditional offline computing modes, computational efficiency is improved by 3-5 times, memory consumption is reduced by more than 40%, and in dynamically updated scenarios, relationship routing discovery can be performed within seconds. Attached Figure Description

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0050] Figure 1 This is a schematic diagram of the system architecture of the present invention.

[0051] Figure 2 This is a schematic diagram of the relationship between examples of the present invention.

[0052] Figure 3 This is a schematic diagram of the incremental update of the present invention.

[0053] Figure 4 This is a schematic diagram of instance relationship mining in this invention. Detailed Implementation

[0054] This invention provides a method and system for mining common neighbor relationships based on graph computing and massive data. The invention first utilizes a knowledge management system for data collection, and then cleans, extracts summaries, and uniformly encodes important attribute values ​​of instances using unified coding to generate one-hop neighbors. Next, a graph routing algorithm is used to extract common neighbors, mining relationships between instances. Finally, concept-level relationship mining is performed using methods such as rule filtering. The system is described below. Figure 1 As shown, it consists of knowledge management, unified coding, distributed graph storage, and relationship mining modules, and includes the following steps:

[0055] Step 1: Build a knowledge management system, including conceptual metadata system construction and data collection.

[0056] Knowledge management is responsible for defining concept metadata and configuring parameters for full data import and incremental data import, as well as storing and managing the mined relationship information.

[0057] Step 2: Unified Encoding

[0058] Uniform encoding is used to encode important attributes of the accessed instance data. A key characteristic of this encoding is that each attribute value corresponds one-to-one with a fixed-length string. If two attributes have the same encoded value, then the attribute values ​​are identical. The encoding algorithm first selects the high-cardinality attributes of the instance data, performs digest extraction and hash calculation on the attribute values, and finally generates a string of length 22, which is then stored in the graph database along with the instance data.

[0059] Step 3: Distributed Graph Storage

[0060] Distributed graph storage is used to store instance data of concepts and a unified encoding. Storage is hierarchical, based on nodes: high-frequency nodes use in-memory adjacency lists, while low-frequency nodes use compressed bitmaps stored on disk to reduce I / O overhead. The storage system employs sharding, redundantly storing data across multiple storage nodes.

[0061] Step 4: Relationship Mining

[0062] Relationship mining is used to uncover potential relationships in massive datasets. It first extracts instances from distributed graph storage, discovers common neighbors between instances based on the characteristics of unified encoding, and finally filters according to predefined rules to uncover conceptual relationships.

[0063] Example

[0064] A method for mining common neighbor relationships in massive datasets based on distributed graph computation and dynamic updates mainly includes four steps:

[0065] Step 1: Build a knowledge management system, including the construction of a concept metadata system and data collection.

[0066] Step 2: Standardize the encoding.

[0067] Step 3: Distributed graph storage.

[0068] Step 4: Relationship mining.

[0069] The specific plan for each step is as follows.

[0070] Step 1: Build a knowledge management system, including conceptual metadata system construction and data collection.

[0071] Step 1-1: Building a Knowledge Management System

[0072] Knowledge management is responsible for defining concept metadata and configuring parameters for full data import and incremental data import, as well as storing and managing the mined relationship information.

[0073] Steps 1-2: Construction of the Conceptual Metadata System

[0074] By examining table attributes, primary / foreign keys, and dependencies in the database / data standard, the mapping rules from the database model to the object-oriented entity model are obtained, thus directly yielding a candidate concept set. Then, the tuples in the database are analyzed to obtain more implicit semantic information and the "inheritance" relationships between concepts. Next, the candidate concepts are evaluated and refined to generate an initial concept model.

[0075] A typical example of converting a database model into an ontology concept:

[0076] (1) First, examine the table attributes, primary and foreign keys, and dependencies in the database. There are seven tables in the database: Infrastructure, Airport, Dock, Operation Relationship, Operating Unit, Equipment, and Location. The Infrastructure table is the main table for the Airport and Dock tables (i.e., it stores the common attributes of the airport and dock, such as ID, name, longitude, latitude, and operating unit). The Infrastructure table is associated with the Operating Unit table through the Operation Relationship table, and with the Location table through a foreign key. The Operating Unit table is also associated with the Equipment table through a foreign key.

[0077] (2) Select the candidate concept set. The infrastructure table in the database is the main table of the airport table and the port table. It is designed to avoid data redundancy and should not be classified as an independent entity object. The operation relationship table is designed to express the many-to-many relationship between infrastructure and operation units. It should also not be classified as an independent entity object. Therefore, five independent entity concepts were selected: airport, port, operation unit, equipment, and location.

[0078] (3) Define attributes. Based on each field in the table, the same attributes can be defined for the concepts. For example, the concept of "airport" can include fields such as "airport class", "runway type", and "runway length" from the airport table. Since the airport table inherits from the infrastructure table, it can also add all the field information from the infrastructure table: "ID", "name", "longitude", and "latitude" attributes.

[0079] (4) Define object attributes. This is key to connecting different concepts. By analyzing foreign key relationships, you can determine which concepts are related and define the corresponding object attributes. For example, if the "Airport" table is related to the "Operating Unit" table through a foreign key, then you can add an attribute "Operating Unit ID" to the "Airport" concept.

[0080] (5) Evaluation and Refinement. First, add constraints to certain attributes of the concepts to ensure data consistency and integrity. For example, the "longitude" and "latitude" attributes of the "airport" concept can be set to be mandatory, and the "runway" attribute can be set to be a positive number. Second, name the relationships between concepts. For example, the relationship between an airport and an operating unit can be named "operator", the relationship between an operating unit and its location can be named "base", and the relationship between an operating unit and its equipment can be named "core assets".

[0081] Through this step, the initial model obtained consists of five concepts (airport, dock, operating unit, equipment, and location) and three relationships (operator, location, and core assets).

[0082] Steps 1-3: Data Acquisition

[0083] Data acquisition primarily targets structured data (topics, relational databases, third-party APIs), and the main process is as follows:

[0084] Step 1-3-1: Select Data Source

[0085] Based on the type of data source being introduced, set the corresponding data source parameter configuration. After the configuration is complete, you can proceed to the next step of field mapping configuration.

[0086] Step 1-3-2: Configure the attribute mapping table

[0087] The mapping rule configuration reflects the mapping rules between the data source and the concept attributes, and also sets the routing rules for subsequent batch data extraction. The extraction will first transform the data queried from the data source according to the rules, based on the selected data source and attribute mapping table. Then, the values ​​will be filled into the corresponding attributes of the concept according to the attribute mapping table to generate new objects based on the concept attributes. Finally, all the processed data will be persisted to the graph database.

[0088] A typical attribute mapping configuration example: Airport concept attributes consist of two parts: attributes from the airport table (airport type, runway length, etc.) and attributes from the inherited infrastructure table (ID, name, longitude, latitude, operating unit, etc.). The resulting mapping rule configuration example is shown in Table 1.

[0089] Table 1 Example of attribute mapping

[0090]

[0091] Step 1-3-3: Configure Incremental Update Strategy

[0092] For data that can be captured for incremental updates (such as database CDC, topic subscriptions, etc.), incremental data can be captured by configuring the update source, thereby enabling incremental updates. For incremental data that cannot be captured or is difficult to capture, scheduled updates can be configured, and the system will perform periodic comparisons to mine incremental data, thus achieving incremental updates.

[0093] Steps 1-3-4: Data Extraction

[0094] After initial configuration, manual data extraction is required. For systems configured with an incremental update strategy, specific instances will be created based on the data in the database. For example, for the concept of "airport," each "airport" record stored in the database will generate a corresponding "airport" instance (such as Beijing Airport, Guangzhou Airport, etc.) after data extraction. For systems configured with an incremental update strategy, subsequent updates will be automatic based on changes in the data source.

[0095] Step 2: Unified Encoding

[0096] Step 2-1: Data Cleaning

[0097] Data cleaning is a crucial step in data preprocessing, aiming to improve data quality and ensure the accuracy of subsequent analysis. Its core logic includes: handling missing values, handling duplicate data, correcting erroneous data, handling outliers, data standardization and transformation, and data denoising. Through a systematic cleaning process, data becomes cleaner and more reliable, laying the foundation for subsequent analysis and modeling.

[0098] Step 2-2: Filter for high-basic attributes

[0099] In the fields of databases and data analysis, high-cardinality attributes typically refer to attributes with a large number of unique values ​​(e.g., the probability of data duplication is less than 60%), such as user ID, order number, and national ID number. Each value of these attributes is almost unique, resulting in a very high cardinality. In contrast, low-cardinality attributes, such as gender, category, season, and Boolean values, have only a small number of duplicate values.

[0100] When performing relationship mining, low-fundamental attributes should be avoided for association. For example, people should not be associated with each other through the gender field, because the gender field only has two values, "male / female". Business applications have shown that the associations built using low-fundamental attributes are not very practical.

[0101] Airport ID and name, port ID and name, location code, and hub location are selected as high-level attributes.

[0102] Steps 2-3: Unified Coding

[0103] Uniform encoding is used to encode important attributes of the accessed instance data. A key characteristic of this encoding is that each attribute value corresponds one-to-one with a fixed-length string. If two attributes have the same encoded value, then the attribute values ​​are identical. This system uses SHA-256 to encode attribute values. Taking "hello" as an example, the core steps are as follows:

[0104] Step 2-3-1: Input the string "hello" and fill the string into a 512-bit block.

[0105] Step 2-3-2: Initialize eight 32-bit hash constants.

[0106] Step 2-3-3: Compress the block 64 times, updating the intermediate value in each round.

[0107] Steps 2-3-4: Finally, the hash value of 64 characters is obtained by merging, as shown below.

[0108] 2cf24dba5fb0a30e26e83b2ac5b9e29e1b161e5c1fa7425e73043362938b9824

[0109] Step 3: Distributed Graph Storage

[0110] Incremental update process as follows Figure 3 As shown, when incrementally updating u, the field values ​​of u are synchronously encoded and stored in the graph library in the form of instance relationship triples.

[0111] For a newly added node Δp (such as a newly added residential instance: "Beijing Daxing"), the edge consisting of this node and the hash code of its key attributes is (u,v), where u is the newly added instance, such as "Beijing Daxing", and v is the unified code of the important attributes of the instance data, that is, the code calculated in step 2, such as the code of the high-basic attribute "location hub" of "Beijing Daxing".

[0112] For each edge (u,v), trigger the following operation:

[0113] Step 3-1: Use the graph routing algorithm to delete u's neighbor v1.

[0114] Step 3-2: Store u as an instance node in the graph library, overwriting the old data.

[0115] Step 3-3: Store (u,v) as a new edge in the graph library.

[0116] Step 4: Relationship Mining

[0117] Step 4-1: Traverse the instances

[0118] Retrieve all instances of concept A from a graph database, which can be traversed multiple times in pages. For example, a graph database such as NebulaGraph can be used.

[0119] Step 4-2: Mining Instance Relationships

[0120] Using a graph routing algorithm, starting from an instance u, a two-hop route lookup is performed to obtain the destination v. The mining logic is as follows: Figure 4 As shown, the query statement is as follows:

[0121] GO 2 STEPS FROM id(u) OVER * YIELD id(v),tags(v)

[0122] The query results are then stored in a relational database, as shown in Table 2.

[0123] Table 2. Schematic diagram of the results of mining instance relationships

[0124]

[0125] Through the above analysis, it was discovered that although Beijing Airport and Tianjin Port are located in different areas, they share the same central hub and can be linked through this hub. An example of this connection is shown in the diagram below. Figure 2 As shown.

[0126] Step 4-3: Statistical Conceptual Relationships

[0127] Based on the information in the table above, statistical analysis can be performed to determine the relationships between concepts. Combined with the ratio of the number of joint investigations to the total number of capabilities, a comprehensive score can be calculated for review by auditors.

[0128] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content and some or all of the steps in various embodiments regarding the method and system for mining common neighbor relationships based on graph computation and massive data. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0129] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0130] This invention provides an approach and system for mining common neighbor relationships based on graph computation and massive data. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for mining common neighbor relationships based on graph computation and massive data, characterized in that, Includes the following steps: Step 1: Construct a conceptual metadata system; Step 2: Collect data from the information system to be mined, and obtain instances of the concept based on the conceptual metadata system constructed in Step 1; Step 3: Unify encoding, and encode the attributes in the instance obtained in Step 2; Step 4: Store the code and its corresponding instance in the graph database using a distributed storage method; Step 5, Relationship Mining: Extract instances from distributed graph storage, discover common neighbors between instances based on encoding characteristics, and then filter rules according to preset rules to generate recommendation relationships; The characteristic of the encoding is that the same attribute value corresponds one-to-one with a string of fixed length. If the encoded values ​​of two attributes are the same, then the attribute values ​​are the same. The construction of the conceptual metadata system mentioned in step 1 includes: Step 1-1: Based on the table attributes, primary and foreign keys, and dependencies in the database of the information system to be mined, determine the main table and the relationships between tables. Steps 1-2: Based on the main table and relationships, filter out candidate concepts to form a candidate concept set; Steps 1-3: Define the attributes of the concept based on the fields contained in the table; Steps 1-4: Determine the relationships between concepts through foreign key relationships in the table, and define object attributes based on these relationships; Steps 1-5: Set constraints and select concepts from the candidate concept set; the selected concepts, their attributes, the relationships between concepts, and the object attributes of the concepts form a concept metadata system; Step 2, which involves data collection from the information system to be analyzed, includes: Full data import and incremental data import; The full data import mentioned in step 2 includes: Data is accessed from a data source, which is the information system to be mined, and is converted into instances of the concept in batches according to the conceptual metadata system constructed in step 1. The incremental data transfer mentioned in step 2 includes: Perform partial updates to the instance based on data changes in the data source; The unified coding mentioned in step 3 includes: Step 3-1, data cleaning, specifically includes: handling missing values, handling duplicate data, correcting erroneous data, handling outliers, data standardization and transformation, and data denoising; Step 3-2: Filter high-cardinality attributes. Among the attributes in the instance, select the attributes whose probability of data duplication is less than the threshold as high-cardinality attributes. Step 3-3: Unified encoding. Encode the high-cardinality attributes in the instance to obtain a string that uniquely corresponds to each high-cardinality attribute value. Step 4, which involves storing the encoding and its corresponding instance in the graph database using a distributed storage method, includes: Step 4-1: The code to be stored and its corresponding instance are added as a new node Δp in the graph database. The edge of this node is represented as (u,v); where u is the new instance and v is the code. Step 4-2: Use the graph routing algorithm to delete the neighbors of instance u; Step 4-3: Store instance u as a new node Δp in the graph database, overwriting the original data; Step 4-4: Store the edge (u,v) as a new edge in the graph database.

2. The method for mining common neighbor relationships based on graph computation and massive data as described in claim 1, characterized in that, The relationship mining described in step 5 includes: Step 5-1: Traverse the instances of the concept; Step 5-2: Based on the graph routing method, query the adjacent instances that have common neighbors; Step 5-3: Based on the starting point instance and the ending point instance, the starting point concept and the ending point concept are statistically obtained, the relationship is formed, and stored in the recommendation relationship table for subsequent manual review and use.

3. The method for mining common neighbor relationships based on graph computation and massive data according to claim 2, characterized in that, The threshold mentioned in step 3-2 is set to 60%.

4. A system for mining common neighbor relationships based on graph computation and massive data, characterized in that, To implement the method of any one of claims 1-3, comprising: The module comprises a knowledge management module, a unified coding module, a relationship mining module, and a distributed graph storage module; among them, The knowledge management module is used to construct a concept metadata system, collect data from the information system to be mined, and obtain instances of concepts based on the concept metadata system. The unified encoding module encodes the attributes in the instances obtained from the knowledge management module; The distributed graph storage module stores the code and its corresponding instance in the graph database using a distributed storage method; The relationship mining module extracts instances from the distributed graph storage module, discovers common neighbors between instances based on the characteristics of the encoding, and then performs rule filtering based on preset rules to generate recommendation relationships.

Images