Knowledge graph recommendation method combining multi-level collaborative information

Abstract

The present invention discloses a knowledge graph recommendation method combining multi-order collaborative information, comprising the following steps: 1) constructing a user-item graph based on the item click sequences of all users; 2) extracting multi-order collaborative information from the user-item graph; 3) converting the item set corresponding to the multi-order collaborative information into a corresponding entity set; 4) combining the multi-order collaborative information with the knowledge information and performing noise reduction on the high-order information through knowledge interactive propagation; 5) creating interaction between the local and high-order graphs, and using the interactive information to perform noise reduction on the high-order information, thereby improving the quality of each layer of knowledge information vector; 6) integrating the local and high-order multi-layer feature vectors of users and items in the knowledge interactive propagation to obtain the overall vector representation of the current user and item.

Description

Technical Field

[0001] The present invention relates to data mining technology, and in particular to a knowledge graph recommendation method combining multi-level collaborative information. Background Art

[0002] Traditional knowledge graph recommendation methods are usually based on knowledge graph embedding algorithms, such as TransE and TransR. However, these methods only focus on the first-order information in the knowledge graph and ignore long-distance high-order information. They are more suitable for in-graph applications, such as link prediction. With the advancement of deep learning technology, many deep neural network-based model studies are committed to enhancing the performance of knowledge graph recommendation tasks. Most current research works build architectures based on graph neural networks (GNNs), such as KGAT and KGCN. Although these models have demonstrated excellent performance in knowledge graph recommendation tasks, their method of encoding all neighboring order information into central nodes also brings the limitation of noise mixing. At the same time, current work only combines low-order collaborative information with knowledge graphs, ignoring the help of high-order collaborative information in modeling user interests in knowledge graph recommendation, which has great limitations. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a knowledge graph recommendation method that combines multi-level collaborative information to address the defects in the existing technology.

[0004] The technical solution adopted by the present invention to solve the technical problem is: a knowledge graph recommendation method combining multi-level collaborative information, comprising the following steps:

[0005] 1) Build a user-item graph based on all users’ item click sequences;

[0006] 2) Extract multi-level collaborative information from the user-item graph; for a given user-item bipartite graph Obtain multi-level collaborative information from the user and item perspectives respectively; low-level collaborative information represents the user's local interests or the item's local features, and high-level collaborative information represents the user's high-level interests or the item's high-level features;

[0007] 3) Aligning collections via item-entity Convert the item set corresponding to the multi-level collaborative information into the corresponding entity set;

[0008] User's initial local interest entity set The definition is as follows:

[0009]

[0010] User's initial high-level interest entity set The definition is as follows:

[0011]

[0012]

[0013]

[0014] Item initial local interest entity set The definition is as follows:

[0015]

[0016] Item initial high-level interest entity set The definition is as follows:

[0017]

[0018]

[0019] 4) Interactive dissemination of knowledge

[0020] Through interactive knowledge dissemination, multi-level collaborative information and knowledge information are combined and high-level information is de-noised;

[0021] 4.1) First, the initial multi-level collaborative information entity set of users and items is propagated in the knowledge graph. This combines collaborative information and knowledge information to obtain the local / high-level graph of users / items, and obtains a rich external entity set and triple set. The entity set and triple set of the local / high-level graph at level l are as follows:

[0022]

[0023]

[0024] Among them, the symbol Indicates u or v, symbol represents L or H, where l represents the number of entity layers of knowledge dissemination;

[0025] 4.2) Encode the multi-layer knowledge information after propagation and obtain the vector representation of each layer of knowledge information;

[0026] Define (h, r, t) as the i-th triplet of the l-th layer triple set, we can get the vector representation of the l-th layer as follows:

[0027]

[0028] in, is the vector representation of the local / high-order graph of the user / item at the lth layer, and the weight Obtained through an attention mechanism:

[0029]

[0030]

[0031] 5) Create interactions between local and high-order graphs and use the interaction information to reduce the noise of high-order information, thereby improving the quality of each layer of knowledge information vector;

[0032] The cross-order interaction is as follows: We first construct a d*d feature interaction matrix between the local and high-order vectors at the lth layer

[0033]

[0034] Therefore we incorporate every possible feature interaction We then update the local and high-order vectors through the feature interaction matrix to achieve the purpose of information noise reduction:

[0035]

[0036]

[0037] 6) Model prediction

[0038] Based on the local and high-order multi-layer feature vectors of users and items in the interactive propagation of knowledge, we integrate them to obtain the overall vector representation of the current user and item. First, we use four different aggregator schemes to obtain the vector representation of users and items in the local and high-order graphs respectively.

[0039] Splicing Aggregator:

[0040] Pooling Aggregator:

[0041] Sum aggregator:

[0042] Average Aggregator:

[0043] We then concatenate the local and high-level representations to obtain complete user and item representations, and use the inner product to predict the likelihood of the user clicking on the item, obtaining the final output:

[0044]

[0045]

[0046]

[0047] The loss function is defined by cross entropy:

[0048]

[0049] in, represents the positive pairs of user-item interactions, The last item is the negative example pair of user-item interaction is the L2 regularization term.

[0050] According to the above solution, in step 2), the third-order information of the user and the second-order information of the item are used as high-order collaborative information.

[0051] The beneficial effects produced by the present invention are:

[0052] By combining multi-order collaborative information with knowledge information, constructing low-order and high-order sub-graphs, integrating more knowledge graph information, interacting between sub-graphs through cross-order interaction, and using interactive information to reduce the noise of high-order information, the accuracy of knowledge graph recommendations is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0054] Figure 1 It is a structural diagram of an embodiment of the present invention;

[0055] Figure 2 is a schematic diagram of collaborative interactive communication according to an embodiment of the present invention;

[0056] Figure 3 2 is a schematic diagram of cross-stage interaction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0058] like Figure 1 As shown in FIG, a knowledge graph recommendation method combining multi-level collaborative information includes the following steps:

[0059] 1) Build a user-item graph based on all users’ item click sequences;

[0060] 2) Extract multi-level collaborative information from the user-item graph; for a given user-item bipartite graph Obtain multi-level collaborative information from the user and item perspectives respectively; low-level collaborative information represents the user's local interests or the item's local features, and high-level collaborative information represents the user's high-level interests or the item's high-level features;

[0061] 3) Aligning collections via item-entity Convert the item set corresponding to the multi-level collaborative information into the corresponding entity set;

[0062] User's initial local interest entity set The definition is as follows:

[0063]

[0064] User's initial high-level interest entity set The definition is as follows:

[0065]

[0066]

[0067]

[0068] Item initial local interest entity set The definition is as follows:

[0069]

[0070] Item initial high-level interest entity set The definition is as follows:

[0071]

[0072]

[0073] 4) Interactive dissemination of knowledge

[0074] Through interactive knowledge dissemination, multi-level collaborative information and knowledge information are combined and high-level information is de-noised;

[0075] 4.1) First, the initial multi-level collaborative information entity set of users and items is propagated in the knowledge graph. This combines the collaborative information and knowledge information to obtain the local / higher-level graph of users / items. The external entity set and triple set of each layer in the local / higher-level graph are obtained. The obtained entity set and triple set of the first layer are as follows:

[0076]

[0077]

[0078] Among them, the symbol Indicates u or v, symbol represents L or H, where l represents the number of entity layers of knowledge dissemination;

[0079] 4.2) Encode the multi-layer knowledge information after propagation and obtain the vector representation of each layer of knowledge information;

[0080] Define (h, r, t) as the i-th triplet of the l-th layer triple set, we can get the vector representation of the l-th layer as follows:

[0081]

[0082] in, is the vector representation of the local / high-order graph of the user / item at the lth layer, and the weight Obtained through an attention mechanism:

[0083]

[0084]

[0085] 5) Create interactions between local and high-order graphs and use the interaction information to reduce the noise of high-order information, thereby improving the quality of each layer of knowledge information vector;

[0086] The cross-order interaction is as follows: We first construct a d*d feature interaction matrix between the local and high-order vectors at the lth layer

[0087]

[0088] Therefore we incorporate every possible feature interaction We then update the local and high-order vectors through the feature interaction matrix to achieve the purpose of information noise reduction:

[0089]

[0090]

[0091] 6) Model prediction

[0092] Based on the local and high-order multi-layer feature vectors of users and items in the interactive propagation of knowledge, we integrate them to obtain the overall vector representation of the current user and item. First, we use four different aggregator schemes to obtain the vector representation of users and items in the local and high-order graphs respectively.

[0093] Splicing Aggregator:

[0094] Pooling Aggregator:

[0095] Sum aggregator:

[0096] Average Aggregator:

[0097] We then concatenate the local and high-level representations to obtain complete user and item representations, and use the inner product to predict the likelihood of the user clicking on the item, obtaining the final output:

[0098]

[0099]

[0100]

[0101] The loss function is defined by cross entropy:

[0102]

[0103] in, represents the positive pairs of user-item interactions, The last item is the negative example pair of user-item interaction is the L2 regularization term.

[0104] Experiments demonstrate that our knowledge graph recommendation method achieves superior results compared to existing mainstream methods. The experiments were evaluated using three benchmark short sequence recommendation datasets: the MovieLens-1M dataset, the Book-Crossing dataset, and the Last.FM dataset. Details of the datasets are shown in Table 1.

[0105] Table 1 Corpus information

[0106]

[0107] The experimental section aims to evaluate the effectiveness of the proposed knowledge graph recommendation model on different datasets. Specifically, we use the AUC and F1 results of the model on three datasets as evaluation metrics. The experimental comparison results are given in Table 2.

[0108] Table 2 Comparative experimental results of short sequence recommendation

[0109]

[0110]

[0111] It is worth noting that, as can be observed in the table above, the method proposed in this invention consistently outperforms all of these baseline models across different metrics. This is because most of these models only combine first-order collaborative information with knowledge information, while ignoring the role of higher-order collaborative information. This invention proposes for the first time to combine higher-order collaborative information with knowledge information, and proposes modeling the interaction between multi-order collaborative information to reduce high-order noise. By combining multi-order collaborative information with knowledge information, the method proposed in this invention can more effectively learn the user's potential interests, thereby improving the accuracy of knowledge graph recommendations.

[0112] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A knowledge graph recommendation method combining multi-level collaborative information, characterized in that: The following steps are involved: 1) Build a user-item graph based on all users’ item click sequences; 2) Extract multi-level collaborative information from the user-item graph; for a given user-item bipartite graph Obtain multi-level collaborative information from the user and item perspectives respectively; low-level collaborative information represents the user's local interests or the item's local features, and high-level collaborative information represents the user's high-level interests or the item's high-level features; 3) Aligning collections via items and entities Convert the item set corresponding to the multi-level collaborative information into the corresponding entity set; User's initial local interest entity set The definition is as follows: User's initial high-level interest entity set The definition is as follows: Item initial local interest entity set The definition is as follows: Item initial high-level interest entity set The definition is as follows: 4) Interactive dissemination of knowledge Through interactive knowledge dissemination, multi-level collaborative information and knowledge information are combined and high-level information is de-noised; 4.1) First, the initial multi-level collaborative information entity set of users and items is propagated in the knowledge graph, thereby combining the collaborative information and knowledge information to obtain the local / high-level graph of users / items, and obtain the external entity set and triple set of each layer in the local / high-level graph; 4.2) Encode the multi-layer knowledge information after propagation and obtain the vector representation of each layer of knowledge information; Define (h, r, t) as the i-th triplet of the l-th layer triple set, and the vector representation of the l-th layer is as follows: in, is the vector representation of the lth layer of the local / high-order graph of the user / item, and the weight Obtained through an attention mechanism: 5) Create interactions between local and high-order graphs and use the interaction information to reduce the noise of high-order information, thereby improving the quality of each layer of knowledge information vector; 6) Model prediction Based on the local and high-order multi-layer feature vectors of users and items in the interactive propagation of knowledge, the overall vector representation of the current user and item is obtained by integration.

2. The knowledge graph recommendation method combining multi-level collaborative information according to claim 1 is characterized in that: In the step 2), the third-order information of the user and the second-order information of the item are used as high-order collaborative information.

3. The knowledge graph recommendation method combining multi-level collaborative information according to claim 1 is characterized in that: The external entity set and triple set of each layer obtained in step 4.1) are as follows: in, Indicates u or v, Represents L or H, where l represents the number of entity layers of knowledge dissemination.

4. The knowledge graph recommendation method combining multi-level collaborative information according to claim 1 is characterized in that: The cross-order interaction in step 5) is specifically as follows: First, a d*d feature interaction matrix is ​​constructed between the local and high-order vectors in the lth layer. Integrates every possible feature interaction The purpose of information noise reduction is achieved by updating local and high-order vectors through the feature interaction matrix:

5. The knowledge graph recommendation method combining multi-level collaborative information according to claim 1 is characterized in that: In step 6), the local and high-order multi-layer feature vectors of users and items in the interactive knowledge propagation are integrated to obtain the overall vector representation of the current user and item, as follows: Four different aggregators are used to obtain vector representations of users and items in local and high-order graphs respectively; the aggregators include: Splicing Aggregator: Pooling Aggregator: Sum aggregator: Average Aggregator: The local and high-level representations are concatenated to obtain complete user and item representations, and the inner product is used to predict the likelihood of the user clicking on the item to obtain the final output: The loss function is defined by cross entropy: in, represents the positive pairs of user-item interactions, The last item is the negative example pair of user-item interaction is the L2 regularization term.

Images