Point-of-interest recommendation method based on knowledge graph and graph attention convolutional network
By constructing an interest point recommendation method based on knowledge graphs and graph attention convolutional networks, the cold start problem and data sparsity problem of new user and new interest point recommendations are solved, the personalization and transparency of recommendations are enhanced, and more accurate location-related recommendations are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN UNIV OF SCI & TECH
- Filing Date
- 2023-10-30
- Publication Date
- 2026-08-04
AI Technical Summary
Existing point-of-interest recommendation systems suffer from a cold start problem with new users or new points of interest. Data sparsity leads to decreased recommendation accuracy, lacks personalization and timeliness, and fails to effectively capture geographical location and user social information, resulting in insufficient transparency.
We construct an interest point recommendation method based on knowledge graphs and graph attention convolutional networks. By using user-interest point interaction knowledge graphs, geographic preference matrices, and graph neural networks, we can mine users' hidden preferences and make recommendations by combining users' social networks and geographic location information.
It improves the accuracy and personalization of interest point recommendations, reduces the impact of data sparsity, addresses the cold start problem, enhances transparency and interpretability, and provides more accurate recommendation results.
Smart Images
Figure CN117349544B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of information technology, and relates to machine learning and recommendation systems, particularly an interest point recommendation method based on knowledge graphs and graph attention convolutional networks. Background Technology
[0002] With the rapid development of my country's economy and the continuous improvement of its technological level, the rapid development of information network technology, and the explosive growth of smartphones, location-based social networks (LBSNs), derived from the functions of smart applications, have gradually entered people's lives. Under this new type of social network, people visit and record geographical points of interest (POIs) offline on apps. As the number of recorded POIs increases, the massive increase in POI check-in data has led to information overload, making POI recommendation a research hotspot in both industry and academia. POI recommendation can not only help users in their daily lives, such as recommending tourist destinations and popular shops, but also provide strong support for data analysis such as holiday travel prediction and public transportation scheduling.
[0003] Traditional interest-based recommendation methods assume that user preferences are fixed and tend to make predictions from a static perspective, which has the following shortcomings;
[0004] Cold start problem: Traditional POI recommendation systems usually rely on users' historical behavior data for recommendations, so there is a cold start problem in the initial stage of users or emerging points of interest; for new users or new points of interest, the system cannot accurately understand the user's preferences, resulting in a decrease in the accuracy of recommendations.
[0005] Data sparsity: The user-interest matrix is often sparse, meaning that most users have only interacted with a few points of interest. This makes it difficult to accurately infer users' interests and behavioral patterns, thus affecting the accuracy of recommendations.
[0006] Lack of personalization: Traditional POI recommendation systems are mainly based on collaborative filtering or content-based methods, which lack personalization capabilities. These methods mainly focus on the similarity between users and points of interest, while ignoring users' personal characteristics, interests, preferences, and contextual information, resulting in a lack of personalization in the recommendation results.
[0007] Difficulty in handling timeliness and geographic location relevance: POI recommendations typically need to consider timeliness and geographic location relevance; however, traditional methods often fail to effectively capture these factors. For example, user interests may change over time, and traditional methods cannot update recommendation results in a timely manner.
[0008] Insufficient Explanation and Transparency: Traditional POI recommendation systems often lack explanation and transparency regarding the recommendation results, making it difficult for users to understand why they received certain recommendations. This may reduce users' trust and acceptance of the recommendation system.
[0009] In conclusion, how to accurately uncover users' hidden preferences and provide high-precision prediction functions are urgent problems that need to be solved. Summary of the Invention
[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide an interest point recommendation method based on knowledge graphs and graph attention convolutional networks. This method takes into account the user's geographical preferences and social network, and can effectively mine the user's hidden preferences, thereby improving the accuracy of interest point recommendation tasks.
[0011] The present invention solves the existing technical problems by adopting the following technical solution:
[0012] An interest point recommendation method based on knowledge graphs and graph attention convolutional networks includes the following steps:
[0013] Step 1: Construct a user-interest point interaction knowledge graph based on user interest point check-in data and user friend information data.
[0014] Step 2: Based on the geographic location data of points of interest and the user's check-in data at points of interest, calculate the user's geographic preference coefficient and generate the user's geographic preference matrix H;
[0015] Step 3: Initialize the graph neural network and integrate the user-interest point interaction knowledge graph. Each entity in the vector is parameterized as an embedding vector, resulting in the user vector e. u With interest point vector e v The user's geographical preference coefficient α v,e As a message relay;
[0016] Step 4: Interact with the knowledge graph based on user interests. Entity embedding and neighborhood entity embedding are performed on top of each other;
[0017] Step 5: Interact with the knowledge graph based on user interests. Perform convolution on the top layer to obtain the trained model and the updated user vector. With interest point vector
[0018] Step 6: Utilize the trained model and the updated user vectors With interest point vector Calculate the preference score y of user u for point of interest v. uvThis is combined with TOP-K sorting to generate a list of POI recommendations that users are interested in.
[0019] Furthermore, step 1 involves constructing a user-interest point interaction knowledge graph. The method is as follows: In the user interest point check-in data and user friend information data, if a user node is a friend of another user, then the relationship between these two nodes is set to 1; if a user node has a check-in relationship with an interest point node, then the relationship between these two nodes is set to 1. After traversing the data, a user-interest point interaction knowledge graph is obtained.
[0020] Furthermore, the specific implementation method of step 2 is as follows: In the geographical location data of points of interest and the user's check-in data of points of interest, cluster the latitude and longitude of each user's check-in points of interest to obtain the average latitude and longitude of the user's check-in; in the user's friend information data, calculate the geographical preference coefficient α between user friends according to the following formula. v,e :
[0021]
[0022] Where dis max Dis represents the maximum distance between the cluster centers of a user's check-in location and the cluster centers of other users' check-in locations. min Dis represents the minimum distance between the cluster centers of a user's check-in location and the cluster centers of other users' check-in locations. v,e Let v be the distance between user v and user e.
[0023] Furthermore, in step 3, when initializing the graph neural network, the vector dimension parameter value is set to 64, 128, or 256 dimensions; the number of convolutional layers is 1, 2, 3, 4, 5, 6, or 7.
[0024] Furthermore, the specific implementation method of step 4 is as follows: using a user-interest point interactive knowledge graph. Above, a hyperparameter K is defined for each node. For each node v, K neighbors are selected for computation. The neighborhood representation of node v is then denoted as... And it satisfies:
[0025]
[0026] After obtaining the representations of each node, the inner product of UV is used in conjunction with the sigmoid function to predict the click probability.
[0027] Furthermore, the hyperparameter K is 1, 2, 3, 4, 5, 6, 7 or 8.
[0028] Furthermore, the specific implementation method of step 5 is as follows:
[0029] Knowledge graph of user-interest interaction Above, the entity representation v and its neighborhood representation are... The aggregation is done by combining the vectors into a single vector.
[0030]
[0031] Where W and b are the transformation weights and biases, respectively, and σ is the sigmoid function;
[0032] Calculate the score between user u and a certain relation r using the inner product g(r,u).
[0033]
[0034] Characterizing the neighborhood information of node v using a linear combination of neighbor information
[0035]
[0036] The weight is the normalized score of all entities e and relations r in N(v) corresponding to node v with respect to the function g, calculated as follows:
[0037]
[0038] In the above formula, α v,e This represents the geographical preference coefficient.
[0039] Furthermore, the predicted score It is expressed as follows:
[0040]
[0041] In the above formula, These are the updated user vector and interest point vector, respectively. User vector transpose, nature
[0042] for and The inner product of.
[0043] The advantages and positive effects of this invention are:
[0044] 1. This invention constructs a knowledge graph based on LBSNs for point of interest recommendation, fully utilizing the auxiliary information of the check-in dataset to reduce the impact of data sparsity. It employs attention mechanisms and hierarchical structure information to capture the logical and hierarchical information in user check-in activities, aggregating logical and hierarchical information through an established layer-by-layer propagation mechanism to better represent user check-in preferences. Furthermore, it connects the user's vector representation with the vector representations of the multi-layer propagation to obtain the final user preference representation, which is then used for Top-K recommendation prediction. This improves upon the shortcomings of traditional point of interest recommendation methods that ignore user social and geographical attribute information, enhancing the accuracy of point of interest recommendation tasks and significantly reducing computational costs.
[0045] 2. This invention addresses the problem that existing point-of-interest recommendation algorithms ignore user friend information and implicit geographical preferences. Taking into account the user's own geographical preferences and social network, it mines the user's hidden preferences. It calculates the geographical preference coefficient between different users, mines the deep semantic information of the user's social information, and establishes a land preference matrix of user friend relationships based on the user's historical check-in records. It further aggregates spatial features to mine the user's implicit semantic information and social preference information.
[0046] 3. The present invention is reasonably designed. It adopts technologies such as deep learning, graph neural networks and reinforcement learning, which can better handle the cold start problem, data sparsity and personalized needs, and make more accurate recommendations by combining geographical location and contextual information. At the same time, transparency and interpretability have also received more attention, enabling users to understand and trust the recommendation results. Attached Figure Description
[0047] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation
[0048] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0049] The design concept of this invention is as follows: Based on the field of point-of-interest (POI) recommendation, this invention utilizes knowledge graphs and graph attention convolutional network technology to construct a user preference attention matrix using user social network information, user-POI check-in information, and POI geographical information. This matrix participates in the convolution process in the form of an attention mechanism within the graph neural network. Finally, the vector representations passed from each layer are merged into the final user preference representation, which is then used for prediction to output a Top-K recommendation list, thereby completing the POI recommendation function.
[0050] Based on the above design concept, this invention proposes an interest point recommendation method based on knowledge graphs and graph attention convolutional networks, such as... Figure 1 As shown, it includes the following steps:
[0051] Step 1: Construct a user-interest point interaction knowledge graph based on user interest point check-in data and user friend information data.
[0052] In the user interest point check-in data and user friend information data, extract all triples corresponding to entities. Each triple (h, r, t) represents a relationship r between the head entity h and the tail entity t. For example, (Xiaoming, friend, Xiaoli) means Xiaoming's friend is Xiaoli, and (Xiaoming, check-in, People's Restaurant) means People's Restaurant is Xiaoming's interest point. Using user check-in records and the entity relationship model, if a user node is another user's friend, the relationship between these two nodes is set to 1; if a user node has a check-in relationship with an interest point node, the relationship between these two nodes is set to 1. After traversing the data, a user-interest point interaction knowledge graph is obtained.
[0053] Step 2: Based on the geographic location data of points of interest and the user's check-in data at points of interest, calculate the user's geographic preference coefficient and generate the user's geographic preference matrix H.
[0054] In the geolocation data of points of interest and the user's check-in data of points of interest, the latitude and longitude of each user's check-in are clustered to obtain the average latitude and longitude of the user's check-in. In the user's friend information data, the geographical preference coefficient between the user's friends is calculated, and this geographical preference coefficient α is... v,e The calculation formula is:
[0055]
[0056] Where dis max Dis represents the maximum distance between the cluster centers of a user's check-in location and the cluster centers of other users' check-in locations. min Dis represents the minimum distance between the cluster centers of a user's check-in location and the cluster centers of other users' check-in locations. v,e Let v be the distance between user v and user e.
[0057] Step 3: Initialize the graph neural network and integrate the user-interest point interaction knowledge graph. Each entity in the vector is parameterized as an embedding vector, resulting in the user vector e. u With interest point vector e v The user's geographical preference coefficient α v,e As a message relay.
[0058] In this step, when initializing the graph neural network, each entity in the knowledge graph is parameterized as an embedding vector to obtain the user vector and the interest point vector. In the selection of vector dimension parameter values, this invention considers three dimension options: 64-dimensional, 128-dimensional, and 256-dimensional.
[0059] Step 4: Interact with the knowledge graph based on user interests. Entity embedding and neighborhood entity embedding are performed on the top.
[0060] Knowledge graph of user-interest interaction For each node, a hyperparameter K is defined. For each node v, K neighbors are selected for computation. In this embodiment, the hyperparameter K is selected from {1, 2, 3, 4, 5, 6, 7, 8}.
[0061] The neighborhood of node v is denoted as follows: And it satisfies:
[0062]
[0063] After obtaining the representation of each node, the inner product of UV is used in conjunction with the sigmoid function to predict the click probability.
[0064] Step 5: Interact with the knowledge graph based on user interests. Perform convolution on the top layer to obtain the trained model and the updated user vector. With interest point vector
[0065] Knowledge graph of user-interest interaction Above, the entity representation v and its neighborhood representation are... Aggregate into a single vector. In this embodiment, the following aggregator is used for aggregation:
[0066]
[0067] Where W and b are the transformation weights and biases, respectively, and σ is the sigmoid function.
[0068] In this embodiment, the number of convolutional layers in the neural network is selected from {1, 2, 3, 4, 5, 6, 7}.
[0069] Define a function g: the inner product, used to calculate the score between user u and a certain relation r.
[0070]
[0071] The neighborhood information of node v is characterized by a linear combination of neighbor information, defined as:
[0072]
[0073] The weight π here is related to the relationship between nodes v and e, and the characteristics of the user u at this point. This weight is essentially the normalized score of all entities e and relations r in N(v) corresponding to v with respect to the function g, calculated as follows:
[0074]
[0075] Where α v,e This represents the geographical preference coefficient.
[0076] After decoding by a multi-layer graph convolutional network, this step yields the embedding vector of user u, representing neighbor information and geographical preferences. Embedding vector with interest point v Step 6: Based on the trained model and the updated user vectors With interest point vector Calculate user u's interest
[0077] Predicted score of point v It also combines TOP-K sorting to generate a list of POI recommendations that users are interested in.
[0078] In this step, the inner product of the final user feature vector representation and the interest point vector representation is used to obtain the prediction score of user u for interest point v. The predicted score It is expressed as follows:
[0079]
[0080] In the above formula, These are the updated user vector and interest point vector, respectively. User vector transpose, nature
[0081] for and The inner product of.
[0082] Finally, the loss function of the overall model of this invention is:
[0083]
[0084] in, It is the cross-entropy loss, where P is the negative sampling distribution, and T is the cross-entropy loss. u T is the number of negative samples for user u. In this invention, T u =|{v:y uv=1}|, and P follows a uniform distribution. The last term in the above formula is the L2 regularization term. When measuring, the number of interest points that a user u has not engaged in is much greater than the number of interest points that have engaged in engagement. For the class imbalance problem, negative sampling is used here to solve it.
[0085] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.
Claims
1. An interest point recommendation method based on knowledge graphs and graph attention convolutional networks, characterized in that: Includes the following steps: Step 1: Construct a user-interest point interaction knowledge graph based on user interest point check-in data and user friend information data. ; Step 2: Based on the geographic location data of points of interest and the user's check-in data at points of interest, calculate the user's geographic preference coefficient and generate a user geographic preference matrix. ; Step 3: Initialize the graph neural network and integrate the user-interest point interaction knowledge graph. Each entity in the vector is parameterized as an embedding vector to obtain the user vector. With interest point vector The user's geographical preference coefficient As a message relay; Step 4: Interact with the knowledge graph based on user interests. Entity embedding and neighborhood entity embedding are performed on top of each other; Step 5: Interact with the knowledge graph based on user interests. Perform convolution on the top layer to obtain the trained model and the updated user vector. With interest point vector ; Step 6: Utilize the trained model and the updated user vectors With interest point vector Calculate users Points of interest Preference score And combine the TOP-K ranking to generate a list of POI recommendations that users are interested in; Step 1 involves constructing a user-interest point interaction knowledge graph. The method is as follows: In the user interest point check-in data and user friend information data, if a user node is a friend of another user, then the relationship between these two nodes is set to 1; if a user node has a check-in relationship with an interest point node, then the relationship between these two nodes is set to 1. After traversing the data, a user-interest point interaction knowledge graph is obtained. ; The specific implementation method of step 2 is as follows: In the geographical location data of points of interest and the user's check-in data of points of interest, the latitude and longitude of each user's check-in points of interest are clustered to obtain the average latitude and longitude of the user's check-in. In user friend information data, the geographical preference coefficient between user friends is calculated using the following formula. : in This represents the maximum distance between the cluster centers of a user's check-in location and the cluster centers of other users' check-in locations. The minimum distance between the cluster centers of a user's check-in location and the cluster centers of other users' check-in locations. For users With users The distance; The specific implementation method of step 4 is as follows: using a user-interest point interactive knowledge graph. Define a hyperparameter for each node. For each node Select The node performs calculations for each neighbor, at which point the node... The neighborhood of is denoted as And satisfy: After obtaining the representation of each node, use The inner product of the two functions is used in conjunction with the sigmoid function to predict click probability. The specific implementation method of step 5 is as follows: Knowledge graph of user-interest interaction Above, represent entities and its neighborhood representation The aggregation is done by combining the vectors into a single vector. in and These are the transformation weights and biases, respectively, and 𝜎 is the sigmoid function; Using inner product Calculate users and a certain relationship The score between : Characterizing nodes using linear combinations of neighbor information Neighborhood information : The weight is the node's weight. corresponding All entities and relationships About functions The scores are then normalized and calculated as follows: In the above formula, Geographic preference coefficient; The predicted score It is expressed as follows: In the above formula, , These are the updated user vector and interest point vector, respectively. User vector transpose, The essence is and The inner product of.
2. The interest point recommendation method based on knowledge graph and graph attention convolutional network according to claim 1, characterized in that: In step 3, when initializing the graph neural network, the vector dimension parameter is set to 64, 128, or 256 dimensions; the number of convolutional layers is 1, 2, 3, 4, 5, 6, or 7.
3. The interest point recommendation method based on knowledge graph and graph attention convolutional network according to claim 1, characterized in that: The hyperparameters The digits are 1, 2, 3, 4, 5, 6, 7, or 8.