WO2023179689A1

WO2023179689A1 - Knowledge graph-based recommendation method for internet of things

Info

Publication number: WO2023179689A1
Application number: PCT/CN2023/083172
Authority: WO
Inventors: 陈俊华; 刘然; 张珈铜; 洪承镐
Original assignee: 重庆邮电大学工业互联网研究院
Priority date: 2022-03-24
Filing date: 2023-03-22
Publication date: 2023-09-28
Also published as: CN114707077A

Abstract

The present invention relates to the field of intelligent recommendation, and specifically, to a knowledge graph-based recommendation method for the Internet of Things, comprising: acquiring a domain knowledge graph of a knowledge construction article from the Internet, and embedding the knowledge graph; after the knowledge graph is embedded, obtaining an article similarity matrix using a Euclidean similarity algorithm; storing user behavior data using the knowledge graph, to obtain a user behavior knowledge graph, and calculating a user-article association matrix and a user-article weight matrix; analyzing the article similarity matrix, the user-article association matrix, and the user-article weight matrix using a recommendation list generation algorithm, to obtain a Top-N recommendation list for a user. The present invention stores user behavior data in the form of a knowledge graph, solving the problem of sparseness of user behavior data, and by means of calculating the article similarity matrix by constructing the domain knowledge graph of articles, article recommendation effects are improved.

Description

A recommendation method based on knowledge graph for the Internet of Things

Technical field

The invention belongs to the field of intelligent recommendation, and specifically relates to a knowledge graph-based recommendation method for the Internet of Things.

Background technique

With the development of high-end manufacturing, enterprises in various fields need to carry out informatization and digital upgrades. Under this trend, it is necessary to use recommendation systems to help companies or individuals obtain information, especially in some industrial fields with high complexity and large amounts of information. In the IoT scenario, relying on the software and hardware support in the IoT, recommendation technology has a variety of application scenarios, including workflow recommendations, food recommendations, in-store shopping recommendations, and health monitoring.

In recommendation systems, collaborative filtering algorithm is one of the most important recommendation algorithms. Traditional collaborative filtering algorithms mainly rely on user behavior data on items to achieve personalized recommendations, because user behavior data on items can reflect the user's interest and preference for items. Generally speaking, user behavior data on items will be stored in the user-item interaction matrix, so the quality of the user-item interaction matrix determines the recommendation effect of the traditional collaborative filtering algorithm. In addition, in practical applications, the problem of sparse user behavior data is very prominent, which may result in items that have never been interacted with by users unable to be recommended.

At present, relevant scholars have proposed some solutions to the data sparsity problem in collaborative filtering. A common method to solve the sparsity problem is to map the high-dimensional user behavior matrix to a low-dimensional space, thereby reducing the impact of user data sparsity on the results of the recommendation algorithm. However, this dimensionality reduction method will discard some user behavior data and reduce the effectiveness of recommendations. Another important method to solve the sparsity problem is collaborative filtering of fused content, which solves the sparsity problem by finding additional information, enriching the portraits of items and users, and improving the accuracy of similarity calculation. But this requires higher costs to mine additional valid and reliable content information.

Contents of the invention

In order to solve the above problems, the present invention proposes a knowledge graph-based recommendation method for the Internet of Things, including:

S1. Obtain the relevant knowledge of the item from the Internet to construct the domain knowledge graph of the item;

S2. Embedding the domain knowledge graph of the item into the knowledge graph;

S3. Use the Euclidean similarity algorithm to calculate the domain knowledge graph after the knowledge graph is embedded, and obtain the item similarity matrix;

S4. Use the knowledge graph to store user behavior data, obtain the user behavior knowledge graph, and calculate the user-item association matrix and user-item weight matrix;

S5. Use the recommendation list generation algorithm to analyze the item similarity matrix, user-item association matrix and user-item weight matrix to obtain the user's Top-N recommendation list.

Furthermore, when using the knowledge graph to store user behavior data, distinguish the user-user relationship and the user-item relationship, and assign weights to both relationships. The weight value range is [0,1], and the weight value is used to represent the knowledge graph. The importance of the relationship between each entity, and the importance of the relationship becomes deeper as the weight value increases.

Furthermore, in the user behavior knowledge graph, users are represented as U = {U ₁ , U ₂ ,..., U _N }, and items are represented as K = {k ₁ , k ₂ ,..., k _M }. According to the user -The weight of item relationship and user-user relationship. Calculate the shortest path weight between items k _j , k _j ∈K and users U _i , U _i ∈U. The process is:

S11. Find all the communication paths from item k _j to user U _i based on the user behavior knowledge graph;

S12. Calculate the product of the weight values of all relationships in each connecting path, and record the product result as the path weight of the connecting path;

S13. Record the largest path weight among the path weights of all connected paths as the shortest path weight from item k _j to user U _i .

Furthermore, the calculation process of the recommendation list generation algorithm includes:

S21. Obtain all items associated with any user through the user-item association matrix to form the user's item set;

S22. Select an item from the item set, add the item information to the user's temporary recommendation list, and obtain the q new items with the highest similarity to the item through the item similarity matrix; S23. Based on the user-item weight matrix , obtain the weight value of the shortest path from the item to the user in the user behavior knowledge graph;

S24. Multiply the similarity between q new items and the item and the weight of the shortest path from the item to the user to obtain the corresponding recommendation index, and add the information of the q new items to the user's temporary Recommended list;

S25. Delete the item from the item collection and return to step S22;

S26. After the calculation of all items in the item set is completed, arrange all the items in the temporary recommendation list in descending order according to the recommendation index, and export the top N item information to obtain the user's Top-N recommendation list.

Furthermore, the Scrapy crawler framework is used to obtain item-related knowledge from the Internet.

Furthermore, the Neo4j graph database is used to store user behavior data and obtain the user behavior knowledge graph.

Beneficial effects of the present invention:

The present invention proposes a knowledge graph-based recommendation method for the Internet of Things, which calculates the item similarity matrix by constructing a domain knowledge graph of items, so that the item similarity can better integrate the semantic relationships between items in the domain knowledge graph, thereby Improved the recommendation effect of the algorithm.

This invention stores user behavior data in the form of a knowledge graph, introduces user-user and user-item relationships, and assigns certain weights to calculate user-item association matrices and user-item weight matrices, thereby further improving user data in the user data. Mining the semantic relationship between users and items, obtaining better recommendation results, and ultimately solving the problem of user behavior data sparseness.

Description of the drawings

Figure 1 is a flow chart of the knowledge graph-based recommendation method for the Internet of Things of the present invention;

Figure 2 is a vehicle recommendation framework based on knowledge graph in an embodiment of the present invention;

Figure 3 is a vehicle domain semantic model in an embodiment of the present invention;

Figure 4 shows the TransD model training algorithm in the embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

The present invention proposes a knowledge graph-based recommendation method for the Internet of Things, as shown in Figure 1, including:

S2. Use the TransD model to embed the domain knowledge graph of the item into the knowledge graph;

Specifically, the Scrapy crawler framework is used to obtain expert knowledge in the field of this item from the Internet, including the price range, brand, origin, size, function, power, color and other characteristic dimensions of this type of item, and knowledge graph construction technology is used to construct this type of item. domain knowledge graph.

In one embodiment, based on the weight of the relationship, the concept of the shortest path weight between two entity nodes in the user behavior knowledge graph is proposed. The user node set in the user behavior knowledge graph is expressed as U={U ₁ , U ₂ , ...,U _N }, the item node is expressed as K={k ₁ ,k ₂ ,...,k _M }, calculate the distance between the item node k _j ,k _j ∈K and the user node U _i ,U _i ∈U The shortest path weight, the process is:

S11. Find all the connecting paths from item node k _j to user node U _i based on the user behavior knowledge graph;

S13. The greater the path weight, the more important the path is and the closer the relationship is. Therefore, all connected paths are The largest path weight among the path weights is recorded as the shortest path weight from item node k _j to user node U _i .

Specifically, the calculation process of the user-item association matrix is:

S101. Based on the above concept of shortest path weight, calculate the shortest path weight of user U _i and all items in the user behavior knowledge graph;

S102. Arrange all the calculated shortest path weights in descending order and obtain the items corresponding to the first p shortest path weights;

S103. Calculate the association between all users and items in the user behavior knowledge graph according to steps S101 and S102, and then record it in the form of a matrix.

Specifically, based on the user-item association matrix, the shortest path weights of users and items in the user behavior knowledge graph are stored to form a corresponding user-item weight matrix.

In one embodiment, the calculation process of the recommendation list generation algorithm includes:

S22. Select an item from the item set, add the item's information to the user's temporary recommendation list, and obtain the q new items with the highest similarity to the item through the item similarity matrix;

S23. Based on the user-item weight matrix, obtain the shortest path weight value of the item to the user in the user behavior knowledge graph;

S24. Multiply the similarity between the q new items and the item and the weight of the shortest path from the item to the user to obtain the corresponding recommendation index, and add the information of the q new items to the user's recommendations. list;

S25. Delete the item from the item collection and return to step S22;

In order to further elaborate on the content of the invention, the present invention will take the vehicle service recommendation application in the modern intelligent parking lot application scenario based on the Internet of Things as an embodiment. In this embodiment, the knowledge-based system for the Internet of Things Tupu’s vehicle service recommendation method, as shown in Figure 2, includes:

S31. Obtain vehicle-related knowledge from the Internet to construct a vehicle domain knowledge graph;

Specifically, the steps for constructing the vehicle domain knowledge graph are:

S301. In this embodiment, the Protégé ontology development tool is used to perform knowledge modeling on the vehicle domain knowledge graph, and the constructed vehicle domain semantic model is shown in Figure 3. Knowledge modeling requires the extraction of feature dimensions that describe vehicles. However, since there are many vehicle features, in order to reduce the construction cost of the domain knowledge graph, this embodiment only selects some important features for modeling, including car model, brand, car series, price range, appearance color, year model, vehicle level, Body structure, number of seats, gearbox, displacement, drive method, air intake method, manufacturer, energy source, production method and vehicle examples.

S302. Based on the constructed vehicle domain knowledge model, this embodiment uses the Scrapy crawler framework to collect knowledge data. In this embodiment, Autohome is selected as the data source, and structured data is crawled according to the vehicle domain knowledge graph;

S303. Extract knowledge from the data obtained by crawling. Since cars have a long history and there are many models and models, in order to simplify the knowledge extraction process and retain the main related information, for a car model, this embodiment only extracts the data in the last ten years. Era models, European cars are uniformly classified into "European series" without further subdivision, and the extracted appearance colors are roughly classified, for example, "Elegant White" is classified into "White";

S304. After the knowledge extraction is completed, a vehicle domain knowledge graph is constructed. In this embodiment, the Neo4j graph database is selected to store the vehicle domain knowledge graph.

S32. Embed the knowledge graph in the vehicle domain knowledge graph to facilitate subsequent similarity calculation; this embodiment uses the TransD model for training. If the training process converges or the algorithm reaches the maximum number of iterations, the training ends. At this time, entities with similar semantics in the vehicle domain knowledge map are mapped to corresponding positions in the vector space;

The TransD model training algorithm is shown in Figure 4. The main idea is that before entering the loop iterative training, in order to speed up the convergence speed and avoid overfitting, the algorithm first uses a random process to normalize entities and relationships, and uses units matrix initializes all projection matrices. Secondly, a small batch of triples is extracted from the training set, and the projection matrix and projection vector are calculated. again, A mini-batch of triples is taken out from the projected relation space, and negative sampling is performed to construct negative triples. Finally, the objective function is optimized through the stochastic gradient descent algorithm.

S33. Obtain the vehicle domain knowledge graph. After the knowledge graph embedding is completed, use the Euclidean similarity algorithm to obtain the vehicle similarity matrix;

Specifically, the similarity calculation process includes: mapping entities and relationships in the vehicle domain knowledge graph to a d-dimensional space, and embedding vehicle I _i into a d-dimensional vector I _i = (E _1i ,E _2i ,...E _pi ,...,E _di ) ^T , E _pi represents the value of vehicle I _i in the p-th dimension; Euclidean distance is used to measure the semantic similarity of vectors to accurately express the similarity between vehicles, because The Euclidean distance is greater than or equal to zero. In order to standardize the calculation, it is converted to between (0,1] through the following operation to obtain the Euclidean similarity:

The larger the calculation result is, the greater the semantic similarity between vehicles is. If its value is 1, it is considered that the two vehicles are very similar semantically; if its value is infinitely close to 0, it is deemed that there is no correlation between the two. Among them, d(I _i ,I _j ) represents the Euclidean distance between vehicles I _i and I _j . The smaller the value, the higher the semantic similarity between the two. The calculation formula is:

After calculating the similarities between all vehicles, they are finally displayed in a matrix:

S34. This embodiment selects the Neo4j graph database to store user behavior data, obtains the user behavior knowledge graph, and calculates the user-item association matrix and user-item weight matrix.

The user behavior knowledge graph retains a lot of complex semantic information between entities and relationships, including user-vehicle and user-user related information, and different relationships have different effects on the user's vehicle recommendation. There will also be varying degrees of impact. Therefore, this embodiment will assign different weight values to each semantic relationship in the user behavior knowledge graph to improve the accuracy of vehicle recommendation. The semantic relationship weight settings are as shown in Table 1:

Table 1 Relationship weight assignment

In Table 1, user-user relationships are defined as three types: family, friends, and colleagues, with weights of 1, 0.95, and 0.90 respectively. User-vehicle relationships are defined as ownership, rental, and driving relationships, with weights of 0.95 and 0.90 respectively. , 0.85 three weights.

In order to facilitate the next step of calculation, the present invention proposes a vehicle selection strategy based on the shortest path weight, which mainly includes: calculating the shortest path weight w _ij of vehicle nodes connected to the user based on the user behavior knowledge graph; filtering the shortest path weight w _ij The vehicle nodes >0.55 are arranged in descending order, and the first p vehicle nodes are selected; the corresponding user-vehicle association matrix and user-vehicle weight matrix are generated based on the association results.

Specifically, the user-vehicle association matrix in this embodiment is shown in Table 2:

Table 2 User-vehicle correlation matrix

Among them, k _mp is the identification id of the vehicle node; when the number of vehicle nodes connected to the user is less than p, the I ₁ value of the matrix row is used to fill the remaining vacancies in the matrix, and corresponding in the user-vehicle weight matrix The weight w _ij is set to 0 to remove the influence of filled vehicle nodes.

Specifically, after obtaining the user-vehicle association matrix, it is necessary to record the shortest path weight of the vehicle I _j in the association matrix to the user U _i in the user behavior knowledge graph for subsequent generation of the recommendation list. The user-item weight matrix in this embodiment is shown in Table 3:

Table 3 User-vehicle weight matrix

Among them, w _mp is the shortest path weight from the corresponding vehicle I _p to the user U _m in the user-vehicle association matrix; similar to the processing method of the user-vehicle association matrix, when the number of vehicle nodes connected to the user is less than p, 0 is used Value fills the remaining gaps, indicating that this vehicle node is ignored.

S35. Use the recommendation list generation algorithm to analyze the vehicle similarity matrix, user-vehicle correlation matrix and user-vehicle weight matrix to obtain the user's Top-N recommendation list.

Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art will understand that various changes, modifications, and substitutions can be made to these embodiments without departing from the principles and spirit of the invention. and modifications, the scope of the invention is defined by the appended claims and their equivalents.

Claims

A knowledge graph-based recommendation method for the Internet of Things, which is characterized by including:

S1. Obtain the relevant knowledge of the item from the Internet to construct the domain knowledge graph of the item;

S2. Perform knowledge graph embedding on the domain knowledge graph of the item, that is, vectorize the entities and relationships in the domain knowledge graph;

S3. Use the Euclidean similarity algorithm to calculate the domain knowledge graph after the knowledge graph is embedded, and obtain the item similarity matrix;

S4. Use the knowledge graph to store user behavior data, obtain the user behavior knowledge graph, and calculate the user-item association matrix and user-item weight matrix;

S5. Use the recommendation list generation algorithm to analyze the item similarity matrix, user-item association matrix and user-item weight matrix to obtain the user's Top-N recommendation list.
A knowledge graph-based recommendation method for the Internet of Things according to claim 1, characterized in that when using the knowledge graph to store user behavior data, the user-user relationship and the user-item relationship are distinguished, and both relationships are evaluated. Give weight, the weight value range is [0,1].
A knowledge graph-based recommendation method for the Internet of Things according to claim 2, characterized in that users in the user behavior knowledge graph are represented as U={U 1 , U 2 ,..., U N }, and items Expressed as K={k 1 ,k 2 ,...,k M }, according to the weight of user-item relationship and user-user relationship, calculate items k j ,k j ∈K and users U i ,U i ∈U The shortest path weight between , the process is:

S11. Find all the communication paths from item k j to user U i based on the user behavior knowledge graph;

S12. Calculate the product of all relationship weight values in each connectivity path, and the product result is recorded as the path weight of the connectivity path;

S13. Record the largest path weight among the path weights of all connected paths as the shortest path weight from item k j to user U i .
A knowledge graph-based recommendation method for the Internet of Things according to claim 1, wherein The characteristic is that the calculation process of the recommendation list generation algorithm includes:

S21. Obtain all items associated with any user through the user-item association matrix to form the user's item set;

S22. Select an item from the item set, add the item's information to the user's temporary recommendation list, and obtain the q new items with the highest similarity to the item through the item similarity matrix;

S23. Based on the user-item weight matrix, obtain the shortest path weight value of the item to the user in the user behavior knowledge graph;

S24. Multiply the similarity between q new items and the item and the weight of the shortest path from the item to the user to obtain the corresponding recommendation index, and add the information of the q new items to the user's temporary Recommended list;

S25. Delete the item from the item collection and return to step S22;

S26. After the calculation of all items in the item set is completed, all items in the temporary recommendation list are arranged in descending order according to the recommendation index, and the top N item information is exported to obtain the user's Top-N recommendation list.
A knowledge graph-based recommendation method for the Internet of Things according to claim 1, characterized in that the Scrapy crawler framework is used to obtain relevant knowledge of items from the Internet.
A knowledge graph-based recommendation method for the Internet of Things according to claim 1, characterized in that a Neo4j graph database is used to store user behavior data to obtain a user behavior knowledge graph.