Knowledge perception recommendation method based on preference personalized aggregation
By constructing a collaborative attribute graph and using graph aggregation technology in knowledge-aware recommendation, combined with the selector mechanism, the problem of insufficient modeling of users' personalized preferences is solved, and a more accurate and personalized recommendation effect is achieved.
Patent Information
- Application Number
- CN202510358561.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-25
- Publication Date
- 2025-06-20
AI Technical Summary
The prior art is difficult to effectively model users' personalized preferences in knowledge-aware recommendations, resulting in insufficient recommendation performance.
By constructing a collaborative attribute map, combining user-item interaction history and knowledge graph, graph aggregation technology is used to build multi-faceted preference embedding representations for each user, and personalized graph aggregation is used to generate the user's final embedding representation.
Accurate modeling of users' personalized preferences in various aspects has been achieved, personalized and accurate recommendations have been improved, and user experience and economic value have been optimized.
Smart Images
Figure CN120179911A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of knowledge-aware recommendation, and in particular to a knowledge-aware recommendation method based on preference personalized aggregation. Background Art
[0002] Currently, the method based on graph neural network (GNN-based) is the mainstream method in the field of knowledge-aware recommendation. By overlapping multiple graph aggregation layers, the graph neural network can simultaneously capture high-order and low-order connection relationships on the graph, so it can better utilize the information of the knowledge graph to help improve the recommendation performance. In order to better utilize the knowledge graph information to construct personal preferences and improve the recommendation performance, existing methods have mainly made many improved designs in the graph aggregation mode, including:
[0003] By using various relationships on the knowledge graph to construct multiple user intents representing the entire recommendation scenario, and then participating the user intents in the graph aggregation process of user representation; although this method constructs multiple user intents at the same time, they are global-oriented and do not achieve personalized modeling for each user.
[0004] Through a heterogeneous propagation aggregation strategy, using the set of attribute nodes of the knowledge graph as the initial representation of users and items, and then performing different modes of propagation aggregation on them respectively; the modeling fineness of this method for user personal preference representation is still insufficient;
[0005] By constructing an attribute-level preference representation for users based on the attributes of the knowledge graph, and comparing and aligning the semantics of this representation with collaborative information, and finally using it in the scoring prediction stage; this method does not consider the diversity of user personal preferences, and lacks personalized modeling of user representation during graph aggregation. Summary of the Invention
[0006] Object of the Invention: The technical problem to be solved by the present invention is to provide a knowledge-aware recommendation method based on preference personalized aggregation in view of the deficiencies of the prior art.
[0007] To solve the above technical problem, the present invention discloses a knowledge-aware recommendation method based on preference personalized aggregation, including the steps of:
[0008] Step 1, input a user set, an item set, the interaction history of users and items, and a knowledge graph, and define a super node representing the preference of each user;
[0009] Step 2, construct a collaborative attribute graph based on the user-item interaction history, the knowledge graph, and the preference super node of each user;
[0010] Step 3: Centering on the preference super-node of each user on the collaborative attribute graph, obtain the preference embedding representation of each aspect of each user through graph aggregation;
[0011] Step 4: On the knowledge graph, obtain the item embedding representation of each layer through graph aggregation, and the sum of the item embedding representations of each layer is used as the final embedding representation of the item;
[0012] Step 5: On the user-item interaction graph, combine the selector and the preference embedding representation of each user, perform multi-layer personalized graph aggregation, and add the embedding representations obtained after graph aggregation of each layer of each user to obtain the final embedding representation of the user;
[0013] Step 6: Obtain the preference prediction score of each user for each item through the user embedding representation and the item embedding representation.
[0014] The user-item interaction graph described in Step 2 is a graph structure constructed based on historical interactions, which contains user nodes and item nodes, and the users and items with interaction records are connected by edges.
[0015] In Step 1, the user-item interaction history is represented as
[0016]
[0017] where u represents the user and i represents the item, represents the user set, represents the item set;
[0018] The knowledge graph is a graph structure composed of entities and relationships. Entities correspond to the nodes of the graph, and relationships correspond to the edges, and are represented as:
[0019]
[0020] where both h and t represent entities, (h, r, t) represents a triple, h represents the head entity, t represents the tail entity, r represents the relationship, represents the set of all entities, represents the set of all relationships, and the entity set contains the item set is represented as
[0021] In Step 1, the preference super-node of each user is jointly composed of N sub-nodes, and each sub-node corresponds to a preference of a certain aspect of a user.
[0022] In Step 2, the specific construction method of the collaborative attribute graph is as follows:
[0023] Step 2-1: For each item, extract the non-item entities directly connected to it in the knowledge graph and the relationships between them, and form the attribute set of this item.
[0024]
[0025] Among them, the relationship-entity binary tuple (r, t) represents an attribute, and (i, r, t) represents a triple in the knowledge graph with the item node as the head entity node, where t represents the entity, r represents the relationship, and i represents the item. represents the set of all entities. represents the set of non-item entities in the knowledge graph, which is a subset of the knowledge graph entity set of.
[0026] Step 2-2: According to the user-item interaction history, obtain the set of items interacted with by each user u.
[0027]
[0028] Step 2-3: Combine the historical interaction item set of each user with the attribute set of the item to obtain the attribute set of the user. The calculation formula is as follows:
[0029]
[0030] Among them, ∪ represents the operation of taking the union.
[0031] Step 2-4: Add the preference hypernode p of the user to the relationship-entity binary tuple (r, t) in each user set u , to obtain (p u , r, t), and construct a star graph centered on the preference hypernode for each user, denoted as
[0032]
[0033] Step 2-5: Merge all the star graphs of the users together with the entity nodes as the connection, that is, if the star graphs of multiple users involve the same entity node, then merge multiple identical such entity nodes into one to form a collaborative attribute graph. The calculation formula is as follows:
[0034]
[0035] In Step 3, obtaining the preference embedding representation of each aspect of each user through graph aggregation is specifically as follows:
[0036]
[0037] Among them, represents the embedded representation of the preference of user u in the n-th aspect. There are a total of N aspects of preferences. represents a d-dimensional real vector. indicates that the embedded representation is a d-dimensional real vector; p u represents the preference supernode of user u, a k represents the user set the k-th relation-entity binary tuple in CAG represents the preference graph aggregation function. α(u, n, k) represents the attention weight corresponding to the k-th relation-entity binary tuple when constructing the preference of user u in the n-th aspect. This graph aggregation operation is achieved by weighted summation of the intermediate results obtained by inputting each relation-entity binary tuple to the graph aggregation function to obtain the final preference embedded representation. The method for calculating the attention weight α(u, n, k) is as follows:
[0038]
[0039] Among them, is a trainable weight, exp is the exponential function, and the attention weight α(u, n, k) is obtained by dividing the weight corresponding to the current k-th relation-entity binary tuple by the sum of the weights of all binary tuples.
[0040] The specific formula of the preference graph aggregation function is as follows:
[0041]
[0042] Among them, w CAG and b CAG are respectively a trainable weight and a bias term, e r is the embedded representation of relation r, is the embedded representation of entity node t. ⊙ is the element-wise multiplication operation, and ELU is the activation function. The meaning of this graph aggregation function is to multiply the embedded representations of the entity and the relation and then perform a non-linear transformation and add the embedded representation of the entity node to obtain a new embedded representation.
[0043] The specific calculation process of each layer and the final embedded representation of the item described in step 4 is as follows:
[0044] Step 4-1: On the knowledge graph, obtain the item embedded representation of each layer through graph aggregation
[0045]
[0046] Among them, the superscript (h) indicates that the embedding representation belongs to the h-th layer. represents the set of relationships and entities adjacent to item i on the knowledge graph, e v (h-1) represents the embedding representation of entity v at the (h - 1)-th layer. H is the total number of layers of graph aggregation. The embedding representation of each layer of the item is obtained by averaging the intermediate representations obtained by applying the graph aggregation function to all adjacent relationships and entities. represents the item graph aggregation function at the (h - 1)-th layer, which is specifically as follows:
[0047]
[0048] Among them, W (h-1) and b (h-1) represent the trainable weights and bias terms of graph aggregation at the (h - 1)-th layer. The meaning of this graph aggregation function is to multiply the embedding representations of entities and relationships, then perform a non-linear transformation, and subtract the embedding representation of the entity to obtain a new embedding representation.
[0049] Step 4-2: Add together the embedding representations corresponding to each item and the embedding representations of a total of H + 1 layers of H-layer graph aggregation to obtain the final embedding representation of the item
[0050] The specific calculation process of the final user embedding representation described in Step 5 is as follows:
[0051] Step 5-1: On the user-item interaction graph, obtain the embedding representation of each layer of the user through graph aggregation
[0052]
[0053] Among them, is the weight generated by the selector to measure the matching degree of the preference of the item in a certain aspect of the user. This weight is multiplied by the relevant embedding representation, and the weighted sum of the results of multiplying the preference embedding representation and the item embedding representation based on the weight output by the selector is used to obtain the embedding representation of the user;
[0054] Step 5-2: Add together the embedding representation of the corresponding user and the embedding representations of a total of H + 1 layers of H-layer graph aggregation to form the final embedding representation of the user:
[0055] The selector is a multi-layer neural network. It takes the preference embedding representation of the user and the item embedding representation as inputs and outputs the weight to measure the matching degree of the preference of the item in a certain aspect of the user. The generation process of this weight is as follows:
[0056]
[0057] z l+1 = ELU(W l z l + b l ), l = 0, 1, 2, …, L - 1
[0058]
[0059] Among them, z0 represents the input of the selector, which is represented by the preference embedding and the embedding representation of item i at the h-th layer constitute, ‖ represents the concatenation operation, z l+1 represents the output of the (l + 1)-th layer of the network, W l and b l represent the training weights and bias terms of this l-th layer. The output of the L-th layer is obtained through iterative operations from layer 0 to layer L - 1, and finally, the output of the L-th layer is converted into a weight that measures the matching degree between the item and the user's preference in a certain aspect through the activation function ReLU
[0060] In step 6, the specific method for obtaining the preference prediction score of each user for each item is to do an inner product:
[0061]
[0062] Beneficial effects:
[0063] Technical level:
[0064] 1. By combining collaborative information, knowledge graphs, and user preference hypernodes to construct a collaborative attribute graph, and performing graph attention aggregation around the user preference hypernode, multiple different aspects of personalized preferences are constructed for each user;
[0065] 2. A selector mechanism is designed, which can evaluate the matching degree between an item and a specific user preference based on personalized preferences and item representations;
[0066] 3. In the process of aggregating user representations from the collaborative graph, by using the personalized preferences of different aspects of the user and the selector mechanism, the contribution degrees of the items to be aggregated are distinguished and a personalized transformation is formed for the item representations, more accurately and personalizedly modeling the user representations, making the final prediction results more accurate.
[0067] Application level:
[0068] 1. It has good transferability and can be applied in different recommendation scenarios (such as music recommendation, book recommendation, movie recommendation, etc.), automatically constructing preferences for different users in different scenarios.
[0069] 2. It can more accurately predict the items that users are interested in, optimizing the user experience and enhancing economic value. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0071] In view of the problem that the current knowledge-aware recommendation method is insufficient in modeling user personalized preferences, the present invention proposes a new method based on personalized preference aggregation, which realizes the modeling of users' multi-faceted personalized preferences, and proposes a selector mechanism that can personalize the aggregation process of user nodes in combination with personalized preferences, achieving more accurate and personalized modeling of user representations and improving the recommendation performance.
[0072] It first constructs a collaborative attribute graph by combining collaborative information, knowledge graph, and personal preference hypernodes, and aggregates with the hypernode as the center, realizing the modeling of multi-faceted personal preferences for each user. When obtaining the user representation by aggregation, the present invention proposes a selector mechanism, together with the previously obtained user personal preferences, realizing personalized aggregation of user nodes, more accurately modeling the user representation, and thus improving the recommendation performance.
[0073] Embedding is the process of mapping high-dimensional data to a low-dimensional space, usually obtained through learning, aiming to preserve the semantic relationships and similarities between data. In machine learning and deep learning, embedding is a technique for converting data (such as words, images, or other entities) into a vector form. In a knowledge graph, entities and relationships are usually represented as points or vectors in a vector space. These vectors are called embedding representations. The actual embedding representation reflects the true meaning and relationships of entities and relationships in the graph, while the virtual embedding representation captures the hidden semantics and concepts between entities and relationships.
[0074] Example 1:
[0075] Corresponding to a music recommendation scenario:
[0076] Step 1: Input the user set, item set, user-item interaction history, and knowledge graph, and define a hypernode representing the preference for each user.
[0077] The user set contains 2,000 listeners.
[0078] The item set contains 4,000 songs of different types.
[0079] The user-item interaction history generates approximately 40,000 music listening records for these 2,000 users;
[0080] The knowledge graph is a knowledge graph about music. The entity set of the knowledge graph includes songs, singers, release time, and genre. The relationship set of the knowledge graph includes the song-singer correspondence, the song-genre correspondence, and the song-release time correspondence;
[0081] A total of 2,000 preference hypernodes are defined. Each hypernode corresponds to a user. It is stipulated that each hypernode consists of 3 sub-nodes, that is, it is considered that each user has 3 different aspects of preferences. At the current initial stage, these preferences are in a blank state and will learn specific meanings during the model training process.
[0082] Step 2: Construct a collaborative attribute graph based on the user-item interaction history, the knowledge graph, and the preference hypernodes of each user;
[0083] Taking a user A in the user set as an example: The main music genres in the music listening records of this user A are 90s R&B songs, Baroque music, and songs by singer X;
[0084] On the knowledge graph, first find the music involved in the music listening history of this user A, and then extract the non-song entities directly connected to these music and the relationships connecting them, including the relationship-entity pairs (song-release time correspondence, 90s), (song-singer correspondence, singer X), (song-genre, R&B), (song-genre correspondence, Baroque). These relationship-entity pairs constitute the attribute set of this user;
[0085] Connect each relationship-entity pair in the attribute set of this user A to the preference hypernode of this user according to the relationship in the pair to form a star graph of this user;
[0086] The star graphs of these 2,000 users are merged together to form a collaborative attribute graph.
[0087] Step 3: Taking the preference hypernode of each user as the center on the collaborative attribute graph, obtain the preference embedding representation of each aspect of each user through graph aggregation;
[0088] Taking the aforementioned user A as an example, the three preferences of this user are "likes 90s R&B songs", "likes Baroque music", and "likes songs by singer X". These three preferences respectively correspond to one sub-node in the preference hypernode of this user;
[0089] By performing graph aggregation operations on the collaborative attribute graph, the preference hyper-nodes that were originally in a blank state obtain information from the connected relation-entity pairs to represent the preferences of user A in 3 aspects;
[0090] Perform this operation for each user to obtain the preference embedding representation of each user.
[0091] Step 4: On the knowledge graph, obtain the item embedding representation of each layer through graph aggregation, and the sum of the item embedding representations of each layer is used as the final embedding representation of the item;
[0092] Perform graph aggregation operations on 4000 pieces of music respectively to obtain their final embedding representations.
[0093] Step 5: On the user-item interaction graph, combine the selector and the preference embedding representation of each user to perform multi-layer personalized graph aggregation, and add the embedding representations obtained after graph aggregation of each layer for each user to obtain the final embedding representation of the user;
[0094] Taking the aforementioned user A as an example, in step 3, the embedding representations corresponding to the three aspects of the user's preferences, "likes 90s R&B songs", "likes Baroque music", and "likes the songs of singer X", are obtained;
[0095] During the process of performing graph aggregation on this user on the user-item interaction graph, the preferences in the three aspects are input into the selector together with the songs in the historical interactions to obtain the matching degree weights between a certain aspect of the preference and a certain song.
[0096] When the preference "likes Baroque music" and a Baroque music piece are input into the selector, a relatively high weight score can be obtained.
[0097] When the preference "likes Baroque music" and a work of singer X are input into the selector, a relatively low weight score will be obtained.
[0098] Apply these weights to the process of graph aggregation to make the aggregation process personalized for this user;
[0099] Perform such personalized graph aggregation operations for each user to obtain the final embedding representations of 2000 users.
[0100] Step 6: Obtain the predicted score of each user for each item through the user embedding representation and the item embedding representation.
[0101] After the previous steps, the final embeddings of 2000 users and 4000 pieces of music are obtained, and then taking the inner product of them can obtain the scores of each user for each piece of music;
[0102] During the training process, based on the user's music listening records, the predicted scores are made closer to the actual listening records, thereby optimizing various trainable parameters of the model;
[0103] The trained model can be used to predict the scores of each user for other unheard music, and a song list is formed by sorting according to the predicted scores and recommended to each user;
[0104] Taking the aforementioned user A as an example, among the list of unheard music recommended by the model for him, the top five songs are: Song 1 by Singer X, R&B Song 1 from the 1990s, R&B Song 2 from the 1990s, Baroque Music 1, and Song 2 by Singer X. This recommendation result conforms to the user's preferences in three aspects.
[0105] Example 2:
[0106] Corresponding to a movie recommendation scenario:
[0107] Step 1: Input the user set, item set, user-item interaction history, and knowledge graph, and define a hypernode representing the preferences for each user;
[0108] The user set contains 3,000 movie fans;
[0109] The item set contains 10,000 movies of different types;
[0110] The user-item interaction history is approximately 80,000 positive feedback records generated by these 3,000 movie fans;
[0111] The knowledge graph is a knowledge graph about movies. Its entity set contains movies, directors, release times, genres, and actors. Its relationship set contains movie-director correspondence, movie-actor correspondence, movie-genre correspondence, and movie-release time correspondence;
[0112] A total of 3,000 preference hypernodes are defined, each corresponding to a user. It is stipulated that each hypernode consists of 4 child nodes, that is, it is considered that each user has 4 different aspects of preferences. At the current stage, these preferences are in a blank state and will learn specific meanings during the model training process.
[0113] Step 2: Construct a collaborative attribute graph based on the user-item interaction history, knowledge graph, and the preference hypernodes of each user;
[0114] Taking a movie fan B in the user set as an example: The main types of movies in the positive feedback records of this movie fan are movies by Director M, movies starring Actor N, science fiction movies, and documentaries;
[0115] On the knowledge graph, first find the movies involved in the positive reviews of user B, and then extract the non-movie entities directly connected to these movies and the relationships connecting them, including (movie - director correspondence, director M), (movie - actor correspondence, actor N), (movie - genre correspondence, science fiction), (movie genre, documentary). These relationship-entity pairs form the user's attribute set;
[0116] Connect each relationship-entity pair in the user set to the user's preference hypernode according to the relationship in the pair to form the user's star graph;
[0117] The star graphs of these two thousand users are merged together to form the collaborative attribute graph.
[0118] Step 3: On the collaborative attribute graph, centered on each user's preference hypernode, obtain the preference embedding representation of each aspect of each user through graph aggregation;
[0119] Taking the aforementioned user B as an example, the four preferences of this user are "likes movies directed by director M", "likes movies starring actor N", "likes science fiction movies", and "likes documentaries". These 4 preferences respectively correspond to a sub-node in the user's preference hypernode;
[0120] By performing graph aggregation operations on the collaborative attribute graph, the preference hypernode that was originally in a blank state obtains information from the connected relationship-entity pairs to represent the 4 aspects of the user's preferences;
[0121] Perform this operation on each user to obtain the preference embedding representation of each user.
[0122] Step 4: On the knowledge graph, obtain the item embedding representation of each layer through graph aggregation, and the sum of the item embedding representations of each layer is used as the final item embedding representation;
[0123] For 10,000 movies, perform graph aggregation operations respectively to obtain their final embedding representations.
[0124] Step 5: On the user-item interaction graph, combine the selector and the preference embedding representation of each user to perform multi-layer personalized graph aggregation, and add the embedding representations obtained after graph aggregation of each layer of each user to obtain the final embedding representation of the user;
[0125] Taking the aforementioned user as an example, in step 3, the embedding representations corresponding to the 4 aspects of the user's preferences "likes movies directed by director M", "likes movies starring actor N", "likes science fiction movies", and "likes documentaries" are obtained;
[0126] In the process of graph aggregation for the user on the user-item interaction graph, the preferences in four aspects are respectively input into the selector together with the movies in their positive review records to obtain the matching degree weights between a certain aspect preference and a certain movie.
[0127] When the preference "likes movies directed by director M" and a movie directed by director M are input into the selector, a relatively high weight score can be obtained.
[0128] When the preference "likes movies directed by director M" and a science fiction movie not directed by director M are input into the selector, a relatively low weight score will be obtained.
[0129] These weights are used in the process of graph aggregation to make the aggregation process personalized for the user.
[0130] Such personalized graph aggregation operations are performed for each user to obtain the final embedded representations of 3000 users.
[0131] Step 6: Obtain the predicted rating of each user for each item through the user embedded representation and the item embedded representation.
[0132] After the previous steps, the final embeddings of 3000 users and 10000 movies are obtained, and then the inner product of them can be used to obtain the rating of each user for each movie.
[0133] During the training process, based on the positive review records of the users, the predicted ratings are made closer to the situation of the positive review records, so as to optimize various trainable parameters of the model.
[0134] The trained model can be used to predict the scores of each user for other movies that have not been watched, and a movie list is formed according to the predicted scores and recommended to each user.
[0135] Taking the aforementioned user B as an example, among the movies that have not been watched recommended by the model for him, the top five movies are: Movie 1 starring actor N, Movie 2 starring actor N, Documentary 1, Science Fiction Movie 1, and Movie 1 directed by director M. This recommendation result conforms to the preferences of this user in four aspects.
[0136] The present invention provides a knowledge-aware recommendation method based on preference personalized aggregation. There are many methods and ways to specifically implement this technical solution. The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented by using the prior art.
Claims
1. A knowledge-aware recommendation method based on preference personalized aggregation, characterized by: Includes steps: Step 1: Input the user set, item set, user and item interaction history, and knowledge graph, and define a super node for each user to represent his / her preference; Step 2: Construct a collaborative attribute graph based on the user-item interaction history, knowledge graph, and each user's preference supernode; Step 3: On the collaborative attribute graph, with each user's preference supernode as the center, obtain the preference embedding representation of each aspect of each user through graph aggregation; Step 4: On the knowledge graph, the item embedding representation of each layer is obtained through graph aggregation, and the item embedding representation of each layer is added together as the final embedding representation of the item; Step 5: On the user-item interaction graph, combine the selector and each user’s preference embedding representation to perform multi-layer personalized graph aggregation, and add the embedding representations obtained after each user’s graph aggregation at each layer to obtain the user’s final embedding representation; Step 6: Obtain each user's predicted preference score for each item through user embedding representation and item embedding representation.
2. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 1 is characterized by: The user-item interaction graph described in step 2 is a graph structure constructed based on historical interactions, including user nodes and item nodes, and users and items with interaction records are connected by edges.
3. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 2 is characterized by: In step 1, the user-item interaction history is expressed as Among them, u represents the user, i represents the item, Represents a user set, Represents a set of items; The knowledge graph It is a graph structure composed of entities and relationships. Entities correspond to the nodes of the graph, and relationships correspond to the edges of the graph, which can be expressed as: Among them, (h, r, t) represents a triple, h represents the head entity, t represents the tail entity, r represents the relationship, Represents the set of all entities, Represents the collection of all relationships, entity set Includes item set Expressed as 4. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 3 is characterized by: In step 1, the preference supernode of each user is composed of N subnodes, and each subnode corresponds to a certain aspect of preference of a user.
5. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 4 is characterized by: In step 2, the specific construction method of the collaborative attribute graph is as follows: Step 2-1: For each item, extract the non-item entities directly connected to it on the knowledge graph and the relationships between them to form the attribute set of this item. Among them, the relationship-entity tuple (r, t) represents the attribute, (i, r, t) represents the triple on the knowledge graph with the item node as the head entity node, t represents the entity, r represents the relationship, and i represents the item. represents the set of all entities, Represents the collection of non-object entities on the knowledge graph, which is the knowledge graph entity set A subset of Step 2-2: Based on the user-item interaction history, obtain the set of items that each user u has interacted with Step 2-3: Combine each user’s historical interaction item set A collection of properties related to items Get the user's attribute set The calculation formula is as follows: Among them, ∪ represents the operation of taking the union; Step 2-4: Add the user's preference supernode p to each relationship-entity tuple (r, t) in the user set u , and get the triple (p u ,r,t), a star graph is constructed for each user, centered on its preferred supernode It is expressed as: Step 2-5: Merge all users' star graphs together with entity nodes. That is, if multiple users' star graphs involve the same entity node, merge multiple identical entity nodes into one to form a collaborative attribute graph. The calculation formula is as follows:
6. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 5 is characterized by: In step 3, the preference embedding representation of each aspect of each user is obtained through graph aggregation, as follows: in, The embedding representation of the nth aspect preference of user u, there are N aspects of preference in total, represents a d-dimensional real vector, Denotes that the embedding is represented as a d-dimensional real vector; p u represents the preference supernode of user u, a k Represents a user collection The kth relation-entity pair in CAG represents the preference graph aggregation function, α(u,n,k) represents the attention weight corresponding to the kth relationship-entity pair when constructing the nth aspect preference of user u, and the method for calculating the attention weight α(u,n,k) is as follows: in, represents the trainable weight, exp is the exponential function, and the attention weight α(u,n,k) is equal to the weight corresponding to the current k-th relation-entity pair divided by the sum of the weights of all relation-entity pairs; The specific formula of the preference graph aggregation function is as follows: Among them, W CAG and b CAG are the trainable weights and bias terms, e r is the embedding representation of relation r, is the embedding representation of the entity node t, ⊙ is the element-by-element multiplication operation, and ELU is the activation function.
7. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 6 is characterized by: The specific calculation process of each layer of the object and the final embedding representation described in step 4 is as follows: Step 4-1: On the knowledge graph, obtain the item embedding representation of each layer through graph aggregation The superscript (h) indicates that the embedding belongs to the hth layer. represents the set of adjacent relationships and entities of item i on the knowledge graph, e r is the embedding representation of relation r, e v (h-1) represents the embedding representation of entity v at the h-1 layer, H is the total number of layers of graph aggregation, Represents the item graph aggregation function of the h-1th layer. The item graph aggregation function is as follows: Among them, W (h-1) and b (h-1) Represents the trainable weights and biases of the h-1th layer graph aggregation; Step 4-2: Embedding representation corresponding to each item And H layers of graph aggregation, a total of H+1 layers of embedding representations are added together to get the final embedding representation of the item 8. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 7 is characterized by: The specific calculation process of the final user embedding representation described in step 5 is as follows: Step 5-1: On the user-item interaction graph, obtain the user embedding representation of each layer through graph aggregation in, It is the weight generated by the selector that measures how well the item matches a certain aspect of the user's preferences; Step 5-2: Embed the corresponding user And the H-layer graph aggregation, a total of H+1 layers of embedding representations are added together to form the final embedding representation of the user:
9. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 8, characterized in that: The selector is a multi-layer neural network that takes the user's preference embedding representation and the item embedding representation as input and outputs a weight that measures the degree of match between the item and the user's preference in a certain aspect. The weight generation process is as follows: With l+1 =ELU(IN l With l +b l ),l=0,1,2,…,L-1 Among them, z0 represents the input of the selector, which is represented by the preference embedding and the embedding representation of item i at layer h Composition, ‖ represents the connection operation, z l+1 represents the output of the l+1th layer of the network, W l and b l Represents the training weights and bias items of the lth layer. The output of the Lth layer is obtained by iterative operation from 0 to L-1 layers. Finally, the output of the Lth layer is converted into a weight that measures the degree of match between the item and a certain aspect of the user's preference through the activation function ReLU.
10. The knowledge-aware recommendation method based on preference personalized aggregation according to claim 9, characterized in that: In step 6, the preference prediction score of each user for each item is obtained. The specific method is to do the inner product: