Interest point sequence recommendation system and recommendation method based on hyperbolic space rotation model
By employing a hyperbolic space rotation model in the recommendation of interest point sequences, constructing a spatiotemporal knowledge graph and performing weighted transition graph learning, and combining graph neural networks and recurrent neural networks, the shortcomings of existing methods in characterizing the hierarchical structure of user spatiotemporal data are addressed, and more efficient interest point recommendation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-21
AI Technical Summary
Existing interest point sequence recommendation methods cannot effectively characterize complex spatiotemporal relationship network structures, especially hierarchical structures, when modeling user spatiotemporal data. Furthermore, due to the characteristics of Euclidean space, they are limited in fully mining information from sparse user movement behavior data.
An interest point sequence recommendation method based on a hyperbolic space rotation model is adopted. By constructing a spatiotemporal knowledge graph, designing a hyperbolic space rotation model, performing weighted transition graph learning, and combining graph neural networks and recurrent neural networks, an interest point sequence recommendation model is constructed. The expressive power of hyperbolic space is used to characterize the hierarchical structure and multi-connection relationships of nodes.
It improves the accuracy of interest point sequence recommendation, can efficiently characterize the hierarchical structure in the relationship network, integrates multiple spatiotemporal factors, and enhances the recommendation effect of the model.
Smart Images
Figure CN121901516A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of user interest recommendation technology, specifically, it relates to an interest point sequence recommendation system and method based on a hyperbolic space rotation model. Background Technology
[0002] The Point of Interest (POI) sequence recommendation problem aims to recommend points of interest that a user is likely to visit in the near future, given the user's current location and time, as well as their historical visited locations. This problem is often defined as Next-Point-of-Interest (POI) Recommendation. Solving this problem typically requires considering multiple factors, such as the user's short- and long-term preferences, sequence transition patterns, and spatiotemporal factors.
[0003] Currently, although many methods have been proposed to solve the problem of interest point sequence recommendation, existing research still has limitations. Since check-in data requires voluntary and proactive upload and sharing by users, the data is extremely sparse, posing a significant challenge to interest point sequence recommendation. Therefore, to build an efficient recommendation model, it is necessary to fully explore the rich information hidden in spatiotemporal behavior. Through analysis of real datasets, this invention can discover very obvious hierarchical structure information in user movement behavior data. However, existing methods based on Euclidean space cannot effectively model complex spatiotemporal relationship network structures, especially hierarchical structures. This is mainly due to the inherent characteristics of Euclidean space itself. Compared to traditional Euclidean space, hyperbolic space has unique advantages: it has stronger expressive power, is more suitable for sparse data, and can efficiently characterize the hidden hierarchical structure in relationship networks. In recent years, hyperbolic space has been combined with technologies such as neural networks, achieving good results in multiple fields. Preliminary research also found that hyperbolic space can model complex user spatiotemporal interaction information using low-dimensional features. However, current research on hyperbolic space models for user spatiotemporal data modeling is in a very early stage and requires further exploration. Summary of the Invention
[0004] To address the aforementioned issues, this invention proposes an interest point sequence recommendation system and method based on a hyperbolic space rotation model. This system uses a rotation model in hyperbolic space to simultaneously model the hidden hierarchical structure and multi-connection relationships within user check-in data, more accurately characterizing the representation of each node. Compared to existing technologies, this method achieves superior results in specific fields such as text processing, image processing, and knowledge graphs.
[0005] This invention is achieved through the following technical solution: An interest point sequence recommendation method based on a hyperbolic space rotation model: The method specifically includes the following steps: Step 1: Obtain the dataset of points of interest and construct a spatiotemporal knowledge graph; Step 2: Based on the spatiotemporal knowledge graph, design a hyperbolic space rotation model and perform weighted transition graph learning; Step 3: Construct an interest point sequence recommendation model based on the hyperbolic space rotation model; Step 4: Output the recommendation results and determine whether convergence has occurred. If convergence has not occurred, repeat step 3. If convergence has occurred, the recommendation of interest points is complete.
[0006] Furthermore, step 1 specifically includes: Step 1.1: Construct a spatiotemporal knowledge graph. Based on user check-in data, create a heterogeneous information network that includes users, points of interest, spatiotemporal relevance, and social relationships; define four types of relationships: access, time, space, and social relationships. The spatiotemporal knowledge graph can be represented by the following formula: ,in User and points of interest set , This represents the set of relationships between nodes. Represents all relationships in a spatiotemporal knowledge graph; There are four types of relationships: access ,time ,space and social relationships The access and time relationships are directed. Step 1.2: Design the embedding function; map nodes and edges to feature vectors to reflect the global hierarchical structure and local multi-element connectivity relationships; After the spatiotemporal knowledge graph is constructed, design an embedding function. Each node in the spatiotemporal knowledge graph and every edge Mapped to a 3D eigenvectors.
[0007] Furthermore, step 2 specifically includes: Step 2.1: Design the hyperbolic rotation model; First, map all nodes and edges of the spatiotemporal knowledge graph from step 1.2 to the same Poincaré sphere model. In a hyperbolic space, we use the hyperbolic space to model and preserve the global hierarchical structure information in the node graph; then we introduce rotation operations into the hyperbolic space and use rotation and transformation operations to characterize the multivariate relationships between nodes and model the local multivariate connection relationships. Step 2.2: Weight function design; Design a weight function. To capture the transition patterns between points of interest; Based on the hyperbolic rotation model, by defining points of interest and The similarity between two points of interest is calculated using temporal and spatial similarity functions, thus obtaining the points of interest in the transition map. and The weight s between them; this weight reflects the point of interest. and The spatiotemporal relationship between them; the specific formula is as follows:
[0008] in, Represents an exponential function. ( ) is a hyperbolic space distance function. The curvature of the temporal relationship between points of interest. The curvature of the distance relationship between points of interest. and Embedded representation of points of interest in hyperbolic space; Then, an interest point transition matrix is constructed based on the weighting function. ; Step 2.3: Construct a sparse transition graph to reduce space consumption; To reduce space consumption, only the weights of the nearest neighbors for each point of interest are calculated, and a sparse transition graph is constructed. ; Specifically, for each point of interest I only want to be closest to him The weights of each neighbor, i.e.:
[0009] in, Finally, the sparse matrix After standardization, an interest point transition map is obtained. : .
[0010] Furthermore, step 3 specifically includes: Construct an interest point sequence recommendation model; the interest point sequence recommendation model includes an embedding layer, a graph neural network layer, an aggregation layer, and a prediction layer; The goal of the embedding layer is to jointly learn representations of users and points of interest; users and points of interest are initially represented as one-hot vectors, first by assigning each user... Sign-in history The sequence is divided into multiple subsequences of equal length, and each subsequence is then used as input to the embedding layer. Each interest point is represented by a... One-hot vector representation of dimensionality; Furthermore, different users have different preferences, and each user is represented as a... The one-hot vectors of the user and interest points are transformed into corresponding low-dimensional dense representations through learning in the embedding layer, as shown below. and .
[0011] Furthermore, the graph neural network layer enriches the representation of interest points through a hyperbolic space rotation model; Using graph convolutional neural networks on interest point transfer maps Continue to improve the representation of each point of interest; An identity matrix was added to it. To obtain a new transition matrix :
[0012] Next, regularize the new transition graph:
[0013] in, express The out-degree diagonal matrix; in the transition graph The representations of each interest point's neighbors are aggregated to enrich the representation of each interest point, ultimately resulting in:
[0014] in This represents a low-dimensional dense representation of all interest points learned by the embedding layer. It is the updated interest point embedding representation.
[0015] Furthermore, the aggregation layer uses an RNN model to combine user preferences and spatiotemporal weighting operations; The aggregation layer consists of two modules: a loop module and an aggregation module; The loop module is used to capture sequence patterns in the user check-in sequence, and the aggregation module uses the spatiotemporal context in the check-in sequence to search for hidden states with high predictive power. The output of the graph neural network layer and the user preference embedding are input into the aggregation layer. First, a recurrent module is used to obtain all the hidden states. Then, the aggregation module uses the spatiotemporal context hidden in the check-in records to design a similarity function. Calculate the current hidden state With historical hidden state The correlation between them, that is:
[0016] in Reflects the periodicity of time. and Indicates adjacent sign-in records and The space and time interval between them; and These represent the time and space attenuation factors, respectively. Since the similarity function only focuses on the correlation between points of interest and ignores the user's preference for points of interest, user preferences are added to the output of the similarity function. Construct a user-interest preference graph based on the access relationship between users and points of interest. Then, the overall preference of each user for their points of interest is calculated:
[0017] in, It is a user preference matrix; Finally, for each user, a new similarity function is obtained that simultaneously considers user preferences and spatiotemporal correlation. :
[0018] in Indicates user Preference embedding, yes distance; Finally, the aggregation module obtains the final output value by weighting the output of the loop module:
[0019] in Similarity score .
[0020] Furthermore, the prediction layer combines the output of the aggregation layer and the user embedding to recommend the next point of interest; The aggregation layer will be applied at each time step in the prediction layer. Output and User Embedding The concatenated vector is input into a fully connected layer to generate the final result:
[0021] in, Learnable parameter matrix, Indicates a splicing operation; Finally, cross-entropy is used as the loss function:
[0022] in, Indicates the length of the user check-in sequence. and Representing tags respectively Other points of interest At the current location The predicted value, yes function.
[0023] A recommendation system based on a hyperbolic space rotation model for recommending interest point sequences, characterized in that: The recommendation system includes: a user interest point acquisition module, a hyperbolic space rotation module, an interest point recommendation module, and an output module; The user interest point collection module is used to acquire an interest point dataset and construct a spatiotemporal knowledge graph. The hyperbolic space rotation module is based on a spatiotemporal knowledge graph, which designs a hyperbolic space rotation model and performs weighted transition graph learning. The point of interest recommendation module is based on a hyperbolic space rotation model to construct a point of interest sequence recommendation model; The output module outputs the recommendation results and determines whether convergence has occurred. If convergence has not occurred, the interest point recommendation module is run again. If convergence has occurred, the recommendation of interest points is completed.
[0024] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.
[0025] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.
[0026] Beneficial effects of the invention Compared with the prior art, the present invention has the following beneficial effects: 1. Using hyperbolic space to model user movement sequences can efficiently characterize the hidden hierarchical structure in relational networks.
[0027] 2. Combining the rotation model proposed in knowledge graphs with hyperbolic space can effectively characterize the complex relationships between nodes while ensuring the hierarchical structure of nodes, and can further integrate various spatiotemporal factors.
[0028] 3. Combining the hyperbolic rotation model with the traditional sequence recommendation model further improves the recommendation accuracy of the model. Attached Figure Description
[0029] Figure 1 This is a schematic diagram of the training process of the interest point sequence recommendation model based on the hyperbolic space rotation model of the present invention.
[0030] Figure 2 This is a schematic diagram of an embodiment of a spatiotemporal knowledge graph.
[0031] Figure 3 This is a schematic diagram of a hyperbolic space rotation model.
[0032] Figure 4 The structure diagram of the sequence recommendation model. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0034] An interest point sequence recommendation system and method based on a hyperbolic space rotation model. The method specifically includes the following steps: Step 1: Obtain the dataset of points of interest and construct a spatiotemporal knowledge graph; Step 1.1: Construct a Spatial-Temporal Knowledge Graph (STKG). Based on user check-in data, create a heterogeneous information network that includes users, points of interest, spatiotemporal relevance, and social relationships; define four types of relationships: access, time, space, and social relationships. like Figure 2 As shown, STKG combines the traditional user-interest interaction graph with the spatiotemporal correlation between interest points and the friendship relationships between users.
[0035] The spatiotemporal knowledge graph can be represented by the following formula: ,in User and points of interest set , This represents the set of relationships between nodes. Represents all relationships in a spatiotemporal knowledge graph; There are four types of relationships: access ,time ,space and social relationships Among them, access relationships and time relationships are directed; the specific construction methods of these four types of relationships are as follows: Figure 2 : Step 1.2: Design the embedding function; Figure 2 The nodes and edges in the model are mapped as feature vectors to reflect the global hierarchical structure and local multi-element connectivity relationships; After the spatiotemporal knowledge graph is constructed, design an embedding function. Each node in the spatiotemporal knowledge graph and every edge Mapped to a 3D eigenvectors.
[0036] Meanwhile, the present invention aims to learn embedding vectors that can simultaneously characterize the global hierarchical structure and local multi-connection relationships in the spatiotemporal knowledge graph.
[0037] Step 2: Based on the spatiotemporal knowledge graph, design a hyperbolic space rotation model and perform weighted transition graph learning; Based on the expectations of step 1, this invention designs a novel hyperbolic rotation model, such as... Figure 3 As shown, this means that all nodes and edges in the spatiotemporal knowledge graph are first mapped to the same... In a hyperbolic space, we can better model the global hierarchical structure of the graph. Then, we introduce rotation operations into the hyperbolic space to model local multi-element connectivity relationships.
[0038] Step 2.1: Design the hyperbolic rotation model; Currently, various hyperbolic space models have been proposed (such as the Poincaré sphere model and the Lorentz model), and these models are interconvertible. Since the Poincaré model is easier to understand and visualize, this invention uses the Poincaré sphere model as an example to describe hyperbolic space. First, map all nodes and edges of the spatiotemporal knowledge graph from step 1.2 to the same Poincaré sphere model. In a hyperbolic space, we use the hyperbolic space to model and preserve the global hierarchical structure information in the node graph; then we introduce rotation operations into the hyperbolic space, and use rotation and transformation operations to effectively characterize the multivariate relationships between nodes and model the local multivariate connection relationships.
[0039] Step 2.2: Weight function design; define temporal similarity and spatial similarity functions, calculate the similarity between interest points; construct the interest point transition matrix to reflect the spatiotemporal relationship between interest points; At this point, the embedding representation of each entity and each relation in the spatiotemporal knowledge graph has been learned. The next goal is to design a weight function. To capture the transition patterns between points of interest; Based on the hyperbolic rotation model, this invention can define points of interest. and The similarity between two points of interest is calculated using temporal and spatial similarity functions, thus obtaining the points of interest in the transition map. and The weight s between them; this weight reflects the point of interest. and The spatiotemporal relationship between them; the specific formula is as follows:
[0040] in, Represents an exponential function. ( ) is a hyperbolic space distance function. The curvature of the temporal relationship between points of interest. The curvature of the distance relationship between points of interest. and Embedded representation of points of interest in hyperbolic space; Subsequently, this invention constructs an interest point transition matrix based on a weighting function. Here, the complete solution is obtained. It is a very time-consuming and memory-intensive task.
[0041] Step 2.3: Construct a sparse transition graph to reduce space consumption; Only calculate the weights of the nearest neighbors for each point of interest; standardize the sparse matrix to obtain the point of interest transition graph; Therefore, in order to reduce space consumption, this invention constructs a sparse transfer graph. ; Specifically, for each point of interest I only want to be closest to him The weights of each neighbor, i.e.:
[0042] in, Finally, the sparse matrix... After standardization, an interest point transition map is obtained. : .
[0043] In step 2, this invention uses a hyperbolic space rotation model to learn a weighted transition graph. First, nodes and edges are mapped to hyperbolic space, and then rotation operations are introduced to model local multivariate connectivity relationships. Next, temporal and spatial similarity functions are defined to calculate the similarity between interest points, and an interest point transition matrix is constructed. Finally, to reduce space consumption, a sparse transition graph is constructed. Step 3: Recommendation of interest point sequences based on hyperbolic space rotation model; Construct an interest point sequence recommendation model; the interest point sequence recommendation model includes an embedding layer, a graph neural network layer, an aggregation layer, and a prediction layer; The embedding layer: learns dense representations of users and points of interest; The goal of the embedding layer is to jointly learn representations of users and points of interest (POIs); these representations encode information about users and POIs and are combined with subsequent modules to make the next POI recommendation; therefore, effectively learning vector representations is crucial. Users and POIs are initially represented as one-hot vectors. However, due to their sparsity, it is difficult for the model to capture user preferences using one-hot vectors. Therefore, this invention learns a low-dimensional dense representation for each user and POI.
[0044] Specifically, this invention first involves each user Sign-in history The sequence is divided into multiple subsequences of equal length, and each subsequence is then used as input to the embedding layer, such as... Figure 4 As shown in the bottom right corner. Each point of interest consists of one... A one-hot vector representation of dimensionality. Furthermore, different users have different preferences; this invention represents each user as a... The one-hot vectors of the user and interest points are transformed into corresponding low-dimensional dense representations through learning in the embedding layer, as shown below. and .
[0045] The graph neural network layer enriches the representation of interest points by learning the interest point transfer map through the hyperbolic space rotation model. To accurately reflect the features of each interest point, simply using the low-dimensional dense representation learned by the embedding layer is insufficient. Therefore, this invention uses a graph convolutional neural network (GCN) in the interest point (weighted) transition graph. The representation of each interest point is further improved. Inspired by the success of LightGCN, this invention focuses solely on the core functionality of GCN: neighbor aggregation. Note that due to the weighted transition graph... It does not reflect the importance of each point of interest to oneself. To solve this problem, the present invention... An identity matrix was added to it. To obtain a new transition matrix :
[0046] Next, regularize the new transition graph:
[0047] in, express The out-degree diagonal matrix. This invention relates to the transition graph. The representations of each interest point's neighbors are aggregated to enrich the representation of each interest point, such as... Figure 4As shown in the lower left corner. Ultimately, the present invention yields:
[0048] in This represents a low-dimensional dense representation of all interest points learned by the embedding layer. It is the updated interest point embedding representation.
[0049] The aggregation layer uses an RNN model to obtain the hidden state output by performing spatiotemporal weighting on the historical hidden states and combining it with user preferences. The aggregation layer comprises two modules: a recurrent module and an aggregation module. The recurrent module captures sequence patterns in the user check-in sequence, while the aggregation module utilizes the spatiotemporal context of the check-in sequence to search for hidden states with high predictive power. Specifically, after the output of the graph neural network layer and the user preference embedding are input into the aggregation layer, a recurrent neural network model is first used to obtain all hidden states. However, directly using these hidden states for recommendation cannot fully utilize the potential temporal periodicity and spatial context in the check-in sequence. For example, users generally exhibit regular check-in cycles and tend to visit nearby points of interest. Inspired by Flashback, this invention explicitly utilizes the spatiotemporal context hidden in the check-in records to design a similarity function. Calculate the current hidden state With historical hidden state The correlation between them, that is:
[0050] in Reflects the periodicity of time. and Indicates adjacent sign-in records and The space and time interval between them. and These represent the time and spatial decay factors, respectively. Since the similarity function only focuses on the correlation between points of interest and ignores user preferences for those points, this invention adds user preferences to the output of the similarity function.
[0051] like Figure 4 As shown in the top left corner, a user-interest preference graph is constructed based on the access relationship between users and points of interest. Then, this invention can determine each user's overall preference for points of interest:
[0052] in, It is a user preference matrix.
[0053] Ultimately, for each user, this invention can derive a new similarity function that simultaneously considers user preferences and spatiotemporal correlation. :
[0054] in Indicates user Preference embedding, yes Distance. Finally, the aggregation module obtains the final output value by weighting the output of the loop module:
[0055] in Similarity score .
[0056] The prediction layer combines the output of the aggregation layer and the user embedding to recommend the user's next point of interest. Its overall architecture is as follows: Figure 4 As shown.
[0057] In the prediction layer, the present invention aggregates the layers at each time step. Output and User Embedding The concatenated vector is input into a fully connected layer to generate the final result:
[0058] in, Learnable parameter matrix, This indicates a concatenation operation. Finally, this invention uses cross-entropy as the loss function:
[0059] in, Indicates the length of the user check-in sequence. and Representing tags respectively Other points of interest At the current location The predicted value, yes function.
[0060] Step 3 involves building an interest point sequence recommendation model, including an embedding layer, a graph neural network layer, an aggregation layer, and a prediction layer. The embedding layer is used to learn dense representations of users and interest points, and the graph neural network layer enriches the representation of interest points through a hyperbolic space rotation model. The aggregation layer contains a recurrent module and an aggregation module, which are used to capture sequence patterns in user check-in sequences and search for hidden states with high predictive power. Step 4: Output the recommendation results and determine whether convergence has occurred. If convergence has not occurred, repeat step 3. If convergence has occurred, the recommendation of interest points is complete.
[0061] In step 4, the model's performance metrics need to be monitored to determine whether the model has converged. Generally, the model can be considered converged when the performance metrics show very little or no improvement over several consecutive iterations.
[0062] If the model has converged, it outputs the final recommendation results. These results can be a list of points of interest that the user might be interested in, or a probability distribution predicting user behavior.
[0063] Iterative optimization: If the model does not converge, the model parameters or training strategy need to be adjusted, and then return to step 3 to retrain the model until satisfactory performance is achieved.
[0064] Complete the recommendation process: Once the model converges and the performance metrics meet expectations, the recommendation system is complete and can begin providing users with interest-based recommendation services.
[0065] A recommendation system for interest point sequences based on a hyperbolic space rotation model, characterized in that: The recommendation system includes: a user interest point acquisition module, a hyperbolic space rotation module, an interest point recommendation module, and an output module; The user interest point collection module is used to acquire an interest point dataset and construct a spatiotemporal knowledge graph. The hyperbolic space rotation module is based on a spatiotemporal knowledge graph, which designs a hyperbolic space rotation model and performs weighted transition graph learning. The point of interest recommendation module is based on a hyperbolic space rotation model to construct a point of interest sequence recommendation model; The output module outputs the recommendation results and determines whether convergence has occurred. If convergence has not occurred, the interest point recommendation module is run again. If convergence has occurred, the recommendation of interest points is completed.
[0066] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.
[0067] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.
[0068] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.
[0069] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium such as a floppy disk, hard disk, magnetic tape; an optical medium such as a high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.
[0070] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.
[0071] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.
[0072] The above provides a detailed description of the interest point sequence recommendation system and method based on a hyperbolic space rotation model proposed in this invention, and elucidates the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. An interest point sequence recommendation method based on a hyperbolic space rotation model, characterized in that: The method specifically includes the following steps: Step 1: Obtain the dataset of points of interest and construct a spatiotemporal knowledge graph; Step 2: Based on the spatiotemporal knowledge graph, design a hyperbolic space rotation model and perform weighted transition graph learning; Step 3: Construct an interest point sequence recommendation model based on the hyperbolic space rotation model; Step 4: Output the recommendation results and determine whether convergence has occurred. If convergence has not occurred, repeat step 3. If convergence has occurred, the recommendation of interest points is complete.
2. The recommended method according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Construct a spatiotemporal knowledge graph. Based on user check-in data, create a heterogeneous information network that includes users, points of interest, spatiotemporal relevance, and social relationships; define four types of relationships: access, time, space, and social relationships. The spatiotemporal knowledge graph can be represented by the following formula: ,in User and points of interest set , This represents the set of relationships between nodes. Represents all relations in a spatiotemporal knowledge graph; There are four types of relationships: access ,time ,space and social relationships The access and time relationships are directed. Step 1.2: Design the embedding function; map nodes and edges to feature vectors to reflect the global hierarchical structure and local multi-element connectivity relationships; After the spatiotemporal knowledge graph is constructed, design an embedding function. Each node in the spatiotemporal knowledge graph and every edge Mapped to a 3D eigenvectors.
3. The recommended method according to claim 2, characterized in that: Step 2 specifically includes: Step 2.1: Design the hyperbolic rotation model; First, map all nodes and edges of the spatiotemporal knowledge graph from step 1.2 to the same Poincaré sphere model. In a hyperbolic space, we use the hyperbolic space to model and preserve the global hierarchical structure information in the node graph; then we introduce rotation operations into the hyperbolic space and use rotation and transformation operations to characterize the multivariate relationships between nodes and model the local multivariate connection relationships. Step 2.2: Weight function design; Design a weight function. To capture the transition patterns between points of interest; Based on the hyperbolic rotation model, by defining points of interest and The similarity between two points of interest is calculated using temporal and spatial similarity functions, thus obtaining the points of interest in the transition map. and The weight s between them; this weight reflects the point of interest. and The spatiotemporal relationship between them; the specific formula is as follows: in, Represents an exponential function. ( ) is a hyperbolic space distance function. The curvature of the temporal relationship between points of interest. The curvature of the distance relationship between points of interest. and Embedded representation of points of interest in hyperbolic space; Then, an interest point transition matrix is constructed based on the weighting function. ; Step 2.3: Construct a sparse transition graph to reduce space consumption; To reduce space consumption, only the weights of the nearest neighbors for each point of interest are calculated, and a sparse transition graph is constructed. ; Specifically, for each point of interest I only want to be closest to him The weights of each neighbor, i.e.: in, Finally, the sparse matrix After standardization, the interest point transfer map is obtained. : 。 4. The recommended method according to claim 3, characterized in that: Step 3 specifically includes: Construct an interest point sequence recommendation model; the interest point sequence recommendation model includes an embedding layer, a graph neural network layer, an aggregation layer, and a prediction layer; The goal of the embedding layer is to jointly learn representations of users and points of interest; users and points of interest are initially represented as one-hot vectors, first by assigning each user... Sign-in history The sequence is divided into multiple subsequences of equal length, and each subsequence is then used as input to the embedding layer. Each interest point is represented by a... One-hot vector representation of dimensionality; Furthermore, different users have different preferences, and each user is represented as a... The one-hot vectors of the user and interest points are transformed into corresponding low-dimensional dense representations through learning in the embedding layer, as shown below. and .
5. The recommended method according to claim 4, characterized in that: The graph neural network layer enriches the representation of interest points through a hyperbolic space rotation model; Using graph convolutional neural networks on interest point transfer maps Continue to improve the representation of each point of interest; An identity matrix was added to it. To obtain a new transition matrix : Next, regularize the new transition graph: in, express The out-degree diagonal matrix; in the transition graph The representations of each interest point's neighbors are aggregated to enrich the representation of each interest point, ultimately resulting in: in This represents a low-dimensional dense representation of all interest points learned by the embedding layer. It is the updated interest point embedding representation.
6. The recommended method according to claim 5, characterized in that: The aggregation layer uses an RNN model that combines user preferences and spatiotemporal weighting operations; The aggregation layer consists of two modules: a loop module and an aggregation module; The loop module is used to capture sequence patterns in the user check-in sequence, and the aggregation module uses the spatiotemporal context in the check-in sequence to search for hidden states with high predictive power. After the output of the graph neural network layer and the user preference embedding are input into the aggregation layer, a recurrent module is first used to obtain all the hidden states. Then, the aggregation module uses the spatiotemporal context hidden in the check-in records to design a similarity function. Calculate the current hidden state With historical hidden state The correlation between them, that is: in Reflects the periodicity of time. and Indicates adjacent sign-in records and The space and time interval between them; and These represent the time and space attenuation factors, respectively. Since the similarity function only focuses on the correlation between points of interest and ignores the user's preference for points of interest, user preferences are added to the output of the similarity function. Construct a user-interest preference graph based on the access relationship between users and points of interest. Then, the overall preference of each user for their points of interest is calculated: in, It is a user preference matrix; Finally, for each user, a new similarity function is obtained that simultaneously considers user preferences and spatiotemporal correlation. : in Indicates user Preference embedding, yes distance; Finally, the aggregation module obtains the final output value by weighting the output of the loop module: in Similarity score .
7. The recommended method according to claim 6, characterized in that: The prediction layer combines the output of the aggregation layer and the user embedding to recommend the next point of interest. The aggregation layer will be applied at each time step in the prediction layer. Output and User Embedding The concatenated vector is input into a fully connected layer to generate the final result: in, Learnable parameter matrix, Indicates a splicing operation; Finally, cross-entropy is used as the loss function: in, Indicates the length of the user check-in sequence. and Representing tags respectively Other points of interest At the current location The predicted value, yes function.
8. A recommendation system for performing the interest point sequence recommendation method based on a hyperbolic space rotation model as described in any one of claims 1 to 7, characterized in that: The recommendation system includes: a user interest point acquisition module, a hyperbolic space rotation module, an interest point recommendation module, and an output module; The user interest point collection module is used to acquire an interest point dataset and construct a spatiotemporal knowledge graph. The hyperbolic space rotation module is based on a spatiotemporal knowledge graph, which designs a hyperbolic space rotation model and performs weighted transition graph learning. The point of interest recommendation module is based on a hyperbolic space rotation model to construct a point of interest sequence recommendation model; The output module outputs the recommendation results and determines whether convergence has occurred. If convergence has not occurred, the interest point recommendation module is run again. If convergence has occurred, the recommendation of interest points is completed.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.