A next point of interest recommendation method and system based on high-order heterogeneous hypergraph and node mask self-supervision

Abstract

The application discloses a next interest point recommendation method and system based on high-order heterogeneous hypergraph and node mask self-supervision. The method comprises the following steps: preprocessing user original interaction data, filtering and constructing interest point semantic mapping according to time threshold segmentation; constructing a high-order heterogeneous hypergraph containing user, interest point, area and category nodes, generating each branch interest point embedding through branch gating, and randomly masking semantic nodes; inputting the masked representation into the heterogeneous hypergraph convolution network for multi-branch propagation, obtaining the final representation of the sample-level user and the candidate interest point through pooling and adaptive fusion; decoding and reconstructing the propagated semantic nodes, constructing the mask self-supervision loss with the pre-masked representation as supervision, and jointly optimizing the prediction loss; and calculating the score based on the final representation of the user and the interest point and outputting the recommendation result. The application models complex semantic association through high-order heterogeneous hypergraph, enhances the stability of the representation by using node mask self-supervision, and effectively improves the recommendation accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for recommending the next point of interest based on high-order heterogeneous hypergraphs and node mask self-supervision. Background Technology

[0002] With the rapid development of mobile internet and location-based services, users' check-in, access, and movement behaviors in real space are constantly accumulating, making next point of interest prediction an important research problem in the field of location recommendation. Existing methods mainly include sequence modeling and graph modeling: the former focuses on using behavioral sequence information to characterize the short-term evolution of users' interests, but has limited ability to jointly model long-term user interaction relationships, category semantic relationships, and spatial semantic relationships; graph modeling can propagate relationship information through graph structures, but most still remain at the level of pairwise relationship modeling, making it difficult to fully express the high-order associations formed by multiple points of interest sharing the same user, the same region, or the same category. In addition, existing methods usually make relatively shallow use of category semantics and region semantics, lacking effective constraints on the semantic node representations themselves. When there is noise or missing semantic information, the robustness and generalization ability of the model are prone to decline. Therefore, there is an urgent need for a next point of interest prediction method that can jointly model long-term user interaction information, spatial semantic information, category semantic information, and short-term transfer information, and enhance the robustness of semantic representation through a self-supervised mechanism. Summary of the Invention

[0003] This invention addresses the problems of insufficient interactive utilization of multi-source heterogeneous information and inadequate semantic node supervision by proposing a method and system for recommending the next point of interest based on high-order heterogeneous hypergraphs and node mask self-supervision.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] This invention proposes a next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision, including:

[0006] Preprocess the user's original interaction data (including check-in data) to obtain the user's behavior sequence sorted by time. According to the preset time threshold, the behavior sequence is divided into sessions. Sessions with a length less than the first preset threshold and users with a number of sessions less than the second preset threshold are removed. Training samples and test samples are generated. Construct the semantic mapping relationship between points of interest and their category attributes and region attributes.

[0007] A (higher-order) heterogeneous hypergraph is constructed, comprising user nodes, interest point nodes, region nodes, and category nodes. The hyperedge relationships include interest point-user interaction relationships, interest point-region relationships, interest point-category relationships, and interest point-interest point transition relationships, each corresponding to multiple semantic branches. The initial embedding of interest points is obtained through a gating mechanism to obtain the embedding representation of interest points in each branch. Random masking is performed on region nodes and category nodes, and the selected semantic node representations are replaced with learnable mask vectors to obtain the masked semantic node representations.

[0008] The embedded representations of interest points in each branch and the masked semantic node representations are fed into a heterogeneous hypergraph convolutional network for multi-branch propagation to obtain the interest point representations and semantic node representations of each branch. The corresponding prefix representations are extracted from the interest point representations of each branch according to the sample prefix sequence. The sample-level representations are obtained through pooling and adaptive fusion to obtain the final user representation at the sample level. The interest point representations of each branch are fused to obtain the final representation of the candidate interest points.

[0009] The propagated semantic node representation is reconstructed by the decoder. The corresponding semantic node representation before the mask is used as the supervision target. The reconstruction loss is calculated to construct the node mask self-supervised loss. The node mask self-supervised loss is jointly optimized with the main task prediction loss.

[0010] Using the test samples, a prediction score is calculated based on the final representation of the sample-level users and the final representation of the candidate interest points. The next interest point recommendation result is then output based on the prediction score.

[0011] Furthermore, the preprocessing of the user's original interaction data includes:

[0012] The behavior records of each user are arranged in ascending order of time; the behavior sequence is divided into multiple session sequences according to the preset time (interval) threshold between adjacent behavior records; sessions with a length less than the first preset threshold (preset minimum length threshold) and users with a number of effective sessions less than the second preset threshold (preset minimum number of sessions threshold) are removed; the effective sessions are divided into training session set and test session set according to time order, wherein the training samples construct behavior prefix sequences and target interest points by sliding window method, and the test samples construct behavior prefix sequences and target interest points by leave-one-out method.

[0013] Furthermore, the construction of the semantic mapping relationship between interest points and their category attributes and region attributes includes:

[0014] Establish an interest point-category mapping based on the category identifier corresponding to each interest point; perform GeoHash geocoding with preset precision to discretize the latitude and longitude of each interest point to obtain the region code, and establish an interest point-region mapping; establish index mappings for category nodes and region nodes respectively to form a set of category nodes, a set of region nodes and corresponding semantic mapping relationships.

[0015] Furthermore, the construction and branch embedding generation of the heterogeneous hypergraph includes:

[0016] A heterogeneous hypergraph is constructed, comprising four types of heterogeneous nodes: user nodes, interest point nodes, category nodes, and region nodes. Within this heterogeneous hypergraph, interest point-user interaction hyperedges, interest point-region relationship hyperedges, interest point-category relationship hyperedges, and interest point-interest point transition hyperedges are constructed, corresponding to the interaction association matrix, interest point-region association matrix, interest point-category association matrix, and interest point transition association matrix, respectively. The transition relationship hyperedges are constructed based on the pre-order to post-order access relationships of interest points within the training session and do not establish transition relationships across sessions. The initial interest point embeddings are used to generate interest point embedding representations corresponding to the interest point-user interaction branch, interest point-region relationship branch, interest point-category relationship branch, and interest point-interest point transition branch through a branch gating mechanism.

[0017] Furthermore, the random masking process performed on region nodes and category nodes includes:

[0018] According to the preset masking ratio, the set of node indices to be masked is randomly sampled from the set of region nodes and the set of category nodes respectively; the original node representation corresponding to the node to be masked is replaced with a learnable mask label vector; the original semantic node representation before masking is retained as the supervision target for subsequent reconstruction.

[0019] Furthermore, the step of feeding the embedded representations of interest points in each branch and the masked semantic node representations into a heterogeneous hypergraph convolutional network for multi-branch propagation includes:

[0020] Aggregation is performed on points of interest (POIs) to user nodes, POIs to category nodes, and POIs to region nodes. User nodes, category nodes, and region nodes fuse the aggregated POI representations with their own features, update their features, and then transmit them back to the POI nodes. In the transition branches, propagation is performed from source nodes to target nodes and from target nodes to source nodes. After multiple iterations, the outputs of each layer are aggregated to obtain the final POI representations for each branch and the final semantic node representations.

[0021] Furthermore, the step of obtaining the sample-level representation through pooling and then adaptively fusing it to obtain the final sample-level user representation includes:

[0022] The variable-length prefix sequence is padded and pooled based on the effective position mask to obtain the sample-level fixed-length representation of each branch; the long-term user preference representation in the interaction branch is fused with the corresponding prefix representation; adaptive weights are calculated for the point of interest-user interaction branch, point of interest-region relationship branch, point of interest-category relationship branch and point of interest-point transfer branch respectively, and the sample-level representation of each branch is weighted and fused based on the adaptive weights to obtain the final sample-level user representation.

[0023] Furthermore, the propagated semantic node representation is reconstructed through a decoder, using the semantic node representation before masking as the supervision target, and the reconstruction loss is calculated to construct the node mask self-supervised loss, including:

[0024] Extract the node vectors at the masked index positions from the propagated region node representation and category node representation; reconstruct the masked nodes using the region decoder and category decoder respectively; constrain the reconstruction results with the corresponding original node representations before masking to obtain the region reconstruction loss and category reconstruction loss; perform a weighted summation of the region reconstruction loss and category reconstruction loss to obtain the node mask self-supervised loss.

[0025] Another aspect of this invention proposes a next interest point recommendation system based on high-order heterogeneous hypergraphs and node mask self-supervision, comprising:

[0026] The data preprocessing module is used to preprocess the user's original interaction data to obtain the user's behavior sequence sorted by time, segment the behavior sequence into sessions according to a preset time threshold, remove sessions with a length less than a first preset threshold and users with a number of sessions less than a second preset threshold, generate training samples and test samples, and construct a semantic mapping relationship between interest points and their category attributes and region attributes.

[0027] The hypergraph construction and masking module is used to construct a heterogeneous hypergraph containing user nodes, interest point nodes, region nodes, and category nodes. The hyperedge relationships include interest point-user interaction relationships, interest point-region relationships, interest point-category relationships, and interest point-interest point transition relationships, each corresponding to multiple semantic branches. The initial embedding of interest points is obtained through a gating mechanism to obtain the embedding representation of interest points in each branch. Random masking is performed on region nodes and category nodes, and the selected semantic node representation is replaced with a learnable mask vector to obtain the masked semantic node representation.

[0028] The heterogeneous hypergraph convolution and representation generation module is used to feed the embedded representations of interest points in each branch and the masked semantic node representations into the heterogeneous hypergraph convolutional network for multi-branch propagation to obtain the interest point representations of each branch and the representations of each semantic node; using the sample prefix sequence of the training samples as an index, the corresponding prefix representation is extracted from the interest point representations of each branch, and the sample-level representation is obtained through pooling; the sample-level user final representation is obtained through adaptive fusion; and the interest point representations of each branch are fused to obtain the candidate interest point final representation.

[0029] The self-supervised and joint optimization module is used to reconstruct the semantic node representation after propagation through the decoder. The semantic node representation before the mask is used as the supervision target, and the reconstruction loss is calculated to construct the node mask self-supervised loss. The node mask self-supervised loss is jointly optimized with the main task prediction loss.

[0030] The recommendation output module is used to calculate a prediction score based on the final representation of the sample-level user and the final representation of the candidate interest point using the test sample, and output the recommendation result of the next interest point based on the prediction score.

[0031] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the next point of interest recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision as described above.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] First, a heterogeneous hypergraph convolution propagation mechanism for enhancing interest point representation was constructed. During heterogeneous hypergraph convolution, interest point messages are first aggregated and passed to user, category, and region nodes. These nodes then fuse their own representations with the aggregated messages before propagating them back to the interest points. This preserves more complete semantic node information while also enabling the interest point representation to learn information from other nodes, thereby enhancing the interest point representation.

[0034] Second, a node masking self-supervised mechanism is introduced to enhance the model's semantic learning capabilities. The random masking mechanism introduced for category nodes and region nodes, combined with decoding reconstruction for self-supervised learning, enables the model to recover the semantic representation of the masked nodes using the contextual information of the heterogeneous hypergraph, thus learning the information of the semantic nodes more fully.

[0035] Third, the invention improves recommendation accuracy by fully utilizing heterogeneous information in the original interaction data. It simultaneously models user interaction information, category semantic information, spatial semantic information, and point-of-interest (POI) shift information, making it better suited for recommending the next POI in location-based service scenarios. Attached Figure Description

[0036] Figure 1 This is one of the flowcharts for a next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision according to an embodiment of the present invention;

[0037] Figure 2 A model framework diagram of a next interest point recommendation method based on high-order heterogeneous hypergraph and node mask self-supervision provided in an embodiment of the present invention;

[0038] Figure 3 This is the second flowchart of a method for recommending the next point of interest based on self-supervision of high-order heterogeneous hypergraphs and node masks according to an embodiment of the present invention.

[0039] Figure 4 This is a schematic diagram of the architecture of a next interest point recommendation system based on high-order heterogeneous hypergraphs and node mask self-supervision, according to an embodiment of the present invention. Detailed Implementation

[0040] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0041] like Figure 1 , Figure 2 As shown, a next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision includes:

[0042] S101: Preprocess the user's original interaction data to obtain the user's behavior sequence sorted by time, and segment the behavior sequence into sessions according to a preset time threshold ∆t, discarding sessions with a length less than the threshold. The number of sessions and the number of sessions are less than the threshold. For users, generate training and test samples; construct semantic mapping relationships between interest points and their category and region attributes.

[0043] S102: Construct a high-order heterogeneous hypergraph containing user nodes, interest point nodes, region nodes, and category nodes. The hyperedge relationships include interest point-user interaction relationships, interest point-region relationships, interest point-category relationships, and interest point-interest point transition relationships, each corresponding to multiple semantic branches (interest point-user interaction branch, interest point-region relationship branch, interest point-category relationship branch, and interest point-interest point transition branch); initial embedding of interest points. The embedding representation of interest points in each hyperedge branch is obtained through a gating mechanism. Random masking is performed on region nodes and category nodes, and the selected semantic node representation is replaced with a learnable mask vector to obtain the masked semantic node representation.

[0044] S103: The masked semantic node representation and the interest point embedding representation of each branch are input into a heterogeneous hypergraph convolutional network for multi-branch propagation to obtain the interest point representation of each branch and the representation of each semantic node. Using the sample prefix sequence of the training samples as an index, the corresponding prefix representation is extracted from the interest point representation of each branch and pooled to obtain the sample-level representation. The sample-level user final representation is obtained through adaptive weighted fusion. The interest point representations of each branch are fused to obtain the candidate interest point final representation.

[0045] S104: The semantic node representations after propagation are reconstructed by the decoder. The corresponding semantic node representation before the mask is used as the supervision target. The reconstruction loss is calculated to construct the node mask self-supervised loss. The node mask self-supervised loss is jointly optimized with the main task prediction loss.

[0046] S105: Using the test samples, calculate the candidate interest point prediction score based on the sample-level user final representation and the candidate interest point final representation, and output the next interest point recommendation result based on the prediction score.

[0047] Further, S101 includes:

[0048] S101.1: Preprocessing of raw behavior sequences to obtain historical user behavior data.

[0049] For any user (u∈U), their behavior sequence arranged in ascending order of time is represented as:

[0050]

[0051] in, This represents the number of records of user u's actions.

[0052] Based on the time interval threshold ∆t between adjacent behavior records, the user behavior sequence is divided into multiple session sequences, resulting in:

[0053]

[0054] in, This indicates the number of sessions corresponding to user u. This represents the i-th session. For sessions with lengths less than the threshold... Sessions are filtered, and sessions with fewer than a threshold are excluded. Users are filtered to obtain a valid set of users and a set of sessions.

[0055] S101.2: Constructing training and testing samples. Valid session set for user u. The sessions are divided into a training session set and a test session set according to time sequence, denoted as follows:

[0056] For any session: The training set uses a sliding window method to generate training samples. For any sample that satisfies... Given a position t, construct the behavior prefix sequence x and the target interest point y:

[0057]

[0058] The training sample set is obtained:

[0059]

[0060] The test set uses the leave-one-out method to generate test samples, constructing a behavior prefix sequence x and a target interest point y:

[0061]

[0062] This results in the test sample set:

[0063]

[0064] in and These represent the total number of training samples and the total number of test samples, respectively.

[0065] S101.3: Construct the category node set and index mapping.

[0066] Define a collection of categories:

[0067]

[0068] Define the category index mapping function:

[0069]

[0070] in This represents the number of categories.

[0071] The category node index for any point of interest p is:

[0072]

[0073] S101.4: Construct the set of region nodes and index mapping.

[0074] Given the region discretization precision parameter g, perform geocoding for each point of interest:

[0075]

[0076] Define a set of regions:

[0077]

[0078] Define the region index mapping function:

[0079]

[0080] in This refers to the number of regions.

[0081] For any point of interest The region node index is:

[0082]

[0083] This step serves as the basis for constructing category branches and region branches in the subsequent hypergraph.

[0084] Further, S102 includes:

[0085] S102.1: Heterogeneous Hypergraph Construction

[0086] Compared to ordinary graph structures that only model pairwise relationships between nodes, heterogeneous hypergraphs can represent higher-order associations between multiple points of interest and the same user, region, or category in the form of hyperedges, thus revealing the potential connections between points of interest more fully. To uniformly characterize higher-order heterogeneous information, this invention designs a multi-branch heterogeneous hypergraph:

[0087] Define heterogeneous hypergraphs as ,in and These represent nodes and superedges, respectively. It is the incidence matrix, used to describe the connectivity of the hypergraph. Nodes This includes a set of interest points (P), a set of users (U), a set of category nodes (C), and a set of region nodes (R). Hyperedge relationships include interaction hyperedges, category hyperedges, region hyperedges, and transition hyperedges. These correspond to four branches: interest point-user interaction branch, interest point-category relationship branch, interest point-region relationship branch, and interest point-interest point transition branch. The association matrix corresponding to these four branches is as follows: in .

[0088] 1) Let the interest point embedding matrix be... For each branch, branch gating is first applied to obtain the input for that specific branch:

[0089]

[0090] in, and For the learnable gating parameters of the X branch. for Activation function. The branch gating coefficient matrix is ​​generated by the gating function and is used to characterize the retention weights of the basic interest points embedded in each feature dimension of the branch.

[0091]

[0092] in, This is an element-wise multiplication. This operation enables branch-specific input construction, enhancing the semantic diversity and expressive power of branches while sharing the basic representation of interest points.

[0093] Meanwhile, let the initial embeddings of the remaining heterogeneous nodes be respectively .

[0094] 2) Construct the association matrix. The rules for constructing the interaction (branching) association matrix are as follows:

[0095]

[0096] The association matrices corresponding to the interest point-category relationship branch and the interest point-region relationship branch (i.e., the interest point-region association matrix and the interest point-category association matrix) are constructed based on static semantic relationships. If the interest point p belongs to category c or region r, it is set to 1 in the corresponding matrix; otherwise, it is 0.

[0097] For transition branches, edges are only created within the same session; transition relationships are not established across sessions. The rules for constructing the interest point transition association matrix are as follows:

[0098]

[0099] 3) Normalization of the incidence matrix. To avoid representation bias caused by uneven distribution of node degrees, the branch matrix is ​​normalized. Normalization is performed so that the neighborhood information received by each node during message propagation is transformed from accumulation to an average aggregation after normalization based on the number of effective connections.

[0100] The side degree matrices of the interest point and the other heterogeneous nodes are as follows:

[0101]

[0102] Where 1 represents a column vector of all 1s, and the result is calculated for each row. The sum of all elements in the corresponding row. for The transpose of the matrix. (Through) Place the degree vector on the diagonal to obtain the degree matrix, and then use the following calculations to obtain the normalized incidence matrix.

[0103]

[0104] After processing, the normalized correlation matrix of each branch used for two-stage propagation is obtained.

[0105] S102.2: Semantic Node Masking. Category nodes and region nodes mainly carry the semantic attribute information of interest points, which plays an important auxiliary role in the recommendation results. However, their representations are easily affected by semantic noise and semantic perturbations during propagation. This invention designs a masking self-supervised learning mechanism to enable the model to learn more stable semantic representations from the hypergraph structure context.

[0106] A set of node indices to be masked is randomly selected from the set of category nodes according to a preset masking ratio, denoted as . The index is used to point to the position using the generated learnable vector. Instead, the learnable vector is randomly initialized by the parameter initialization module. Then, the masked category node representation is:

[0107]

[0108] in, This represents the category node representation after masking.

[0109] Similarly, for region nodes, the region node representation matrix after masking is obtained. The original representation before masking serves as the target for reconstruction supervision, while the semantic node representation after masking will be used as the input for subsequent heterogeneous hypergraph convolution.

[0110] Further, S103 includes:

[0111] S103.1: Construction of Heterogeneous Hypergraph Convolutional Networks. Let... The interest point of the l-th layer is represented as The remaining heterogeneous nodes are represented as For the interest point-region relationship branch and the interest point-category relationship branch, the input node representations are respectively initialized using masked vectors. , .

[0112] The propagation process at layer l is as follows:

[0113] The first step is to aggregate interest nodes to the superedge:

[0114]

[0115] in These are the learnable parameters of the model, responsible for the linear transformation at the point of interest.

[0116] The second step is the fusion of heterogeneous node features:

[0117]

[0118] in and For the model's learnable parameters, Responsible for linear transformations on the heterogeneous node side It is responsible for the transformation after splicing and merging.

[0119] The third step is to send back the points of interest and perform residual discard updates.

[0120]

[0121] After multiple iterations, the outputs of each layer are aggregated to obtain the final interest point representation of the branch. , , and the final representation of semantic nodes .

[0122] For the transition branch, let the input of the l-th layer be... Propagation occurs via a source node to target node and target node to source node method:

[0123]

[0124] After multi-level iteration and inter-level aggregation, we obtain Thus, we have obtained the final interest point representations for each branch, as well as the representations for each semantic node.

[0125] S103.2: Sample-level user and interest point representation generation. To construct the user representation of the current sample, it is first necessary to generate the representation based on the sample prefix. Extract the corresponding sample-level interest point representation from the final interest point representation.

[0126]

[0127] Where [x] represents the index operation. This represents the matrix of interest points from each branch, indexed by the interest point index of the sample prefix x. The extracted representation sequence.

[0128] Because the prefix lengths of different samples are inconsistent, the extracted sequences are variable-length sequences and cannot be directly used for unified fusion calculations. Therefore, this invention employs a masked pooling method: the vectors corresponding to the padding positions are masked, and mean aggregation is performed only on the interest point representations at valid positions. This effectively avoids the impact of differences in padding positions and sequence lengths on the representation results. The corresponding formula is as follows:

[0129]

[0130] After obtaining the interest point representations fused from the sample levels of each branch, they need to be mapped to sample-level user representations. The interaction branch contains user node representations; to simultaneously preserve both long-term stable user preferences and current prefix short-term intentions, an additional user vector is introduced. And merge it with the current sample prefix representation:

[0131]

[0132] Branch weights are adaptively generated from the interest point representations of each branch sample through a learnable gating function:

[0133]

[0134] By unifying and integrating user interaction, spatial semantics, category semantics, and transition semantics, and adaptively weighting and fusing each branch according to its weight, the final representation of the current sample user is obtained.

[0135]

[0136] The interest point representations of each branch are normalized and fused to obtain the final representation of the sample interest points.

[0137]

[0138] Further, S104 includes:

[0139] S104.1: Reconstruct the masked semantic nodes using a decoder

[0140]

[0141] in, Same as S102, representing the set of mask node indices. The representation decoder is used to map the propagated semantic node representation back to the original semantic embedding space in order to recover the reference representation of the masked node before masking.

[0142] S104.2: Construct an auxiliary loss by recovering the reference representation of the masked node before masking using the propagated semantic node representation.

[0143]

[0144] in, This is a hyperparameter. To reconstruct the error function:

[0145]

[0146] in, This is a hyperparameter.

[0147] The self-supervised loss of semantic node masking is:

[0148]

[0149] S104.3: Using the real target interest point label y as the supervision signal, the main task loss is constructed using cross-entropy loss:

[0150]

[0151] The main task loss and the node mask self-supervised loss obtained from S104 are jointly used to construct the total loss function:

[0152]

[0153] Further, S105 includes:

[0154] Next point of interest prediction and joint optimization. Sample-level user representations obtained based on S103. With the final representation matrix of interest points Calculate the prediction score of this sample for all candidate interest points:

[0155]

[0156] By minimizing the total loss The model parameters are iteratively optimized to obtain the trained model parameters;

[0157] During the testing phase, candidate interest point prediction scores are obtained based on the test samples. Sort the results and output the next recommendation for the point of interest.

[0158] This invention constructs a high-order heterogeneous hypergraph to fully model the complex high-order semantic relationships between users and points of interest. Compared to ordinary graph structures that only model pairwise relationships between nodes, this invention constructs a heterogeneous hypergraph that includes user nodes, point of interest nodes, region nodes, and category nodes. It uses hyperedges to represent the high-order relationships between multiple points of interest and the same user, region, or category, thus revealing the potential connections between points of interest more fully.

[0159] By employing a multi-branch gating mechanism, the semantic diversity and expressive power of each branch are enhanced. This invention generates embedding representations for each hyperedge branch through branch gating of the initial embedding of interest points, enhancing the semantic diversity and expressive power of each branch while sharing the basic interest point representation. Subsequently, through adaptive weighted fusion of the sample-level representations of each branch, user interaction, spatial semantics, category semantics, and transition semantics are uniformly integrated to obtain a more comprehensive user representation.

[0160] This invention utilizes node mask self-supervised learning to enable the model to learn more stable semantic representations. Addressing the issue that category nodes and region nodes are easily affected by semantic noise and perturbations during propagation, this invention designs a mask self-supervised learning mechanism. This mechanism performs random masking on semantic nodes, using the similarity between the reconstructed representation after propagation and the original representation before masking as an auxiliary loss, thus enabling the model to learn more stable semantic representations from the hypergraph structure context.

[0161] This invention effectively addresses the aggregation problem of variable-length sequences using a masked pooling method. To address the issue of inconsistent prefix lengths among different samples preventing direct unified fusion, this invention employs a masked pooling method. This method masks the vectors corresponding to the padding positions and performs mean aggregation only on the interest point representations at valid positions, effectively avoiding the impact of padding position and sequence length differences on the representation results.

[0162] This invention improves recommendation performance through multi-task joint optimization. It jointly constructs a total loss function by combining the main task prediction loss and the node mask self-supervised loss. While optimizing the recommendation objective for the next point of interest, it enhances the model's ability to capture semantic information through a semantic node reconstruction task, achieving multi-task collaborative optimization and improving recommendation effectiveness.

[0163] As one possible implementation method, such as Figure 3 As shown, another method for recommending the next point of interest based on high-order heterogeneous hypergraphs and node mask self-supervision of the present invention includes:

[0164] S201: Dataset Preprocessing and Semantic Node Processing

[0165] This embodiment uses the TSMC2014-TKY dataset. First, the original check-in logs are read, aggregated by user identifier, and then sorted in ascending order by time. To reduce noise, the number of times each point of interest (POI) is accessed by different users is counted, low-frequency POIs are removed, and only POIs with at least 5 accessed users are retained. A session segmentation time interval threshold ∆t=24 is set, and records with adjacent check-in time differences exceeding this threshold are assigned to different sessions. Records shorter than [a certain length] are removed. Sessions with a count of 3 or less, and removing sessions with a valid count less than 3. =3 users. Based on this, the sessions are divided into training session set and test session set according to user time order, with the training ratio set to k=0.8. The training samples are constructed using a sliding window method to create prefix sequence-target interest points, and the test samples are constructed using a leave-one-out method to create the final prediction samples, thus forming the training sample set and the test sample set. The preprocessing results of the TSMC2014-TKY dataset are shown in Table 1.

[0166] Table 1. Preprocessing results of the TKY dataset

[0167]

[0168] For each valid point of interest (POI), its category identifier is read and a "POI-category" mapping relationship is established to form a set of category nodes; the latitude and longitude of each valid POI are geohash encoded, and the geocoding precision parameter is set to [value missing]. An interest point-region mapping relationship is established to form a set of region nodes. Subsequently, the category nodes and region nodes are indexed to obtain a semantic mapping table for subsequent branching graph construction. The number of semantic nodes in the semantic mapping table is shown in Table 2.

[0169] Table 2. Number of semantic nodes in the semantic mapping table

[0170]

[0171] S202: Heterogeneous Hypergraph Construction and Semantic Node Masking

[0172] User historical check-in data contains high-order shared relationships such as multiple points of interest being visited by the same user, multiple points of interest belonging to the same region, and multiple points of interest corresponding to the same category. Simply using a regular graph to connect nodes in pairs is insufficient to fully represent these diverse relationships, resulting in underutilization of information. Therefore, this invention integrates user nodes, point of interest nodes, category nodes, and region nodes into a heterogeneous hypergraph, and uses interaction hyperedges, category hyperedges, region hyperedges, and transition hyperedges to characterize different types of relationships, improving the ability to utilize high-order structural information in subsequent convolutional propagation stages. The specific steps are as follows:

[0173] S202.1: Branch Relationship Definition. Four branches are constructed: Interest Point-User Interaction Branch, Interest Point-Category Relationship Branch, Interest Point-Region Relationship Branch, and Interest Point Transfer Branch. The interaction and transfer branches are constructed solely based on the training session history, while the interest point-category and interest point-region relationship branches are constructed based on the static semantic mapping of interest points.

[0174] S202.2: Construction of the association matrix. Let the set of interest points be P, the set of users be U, the set of category nodes be C, and the set of region nodes be R. Construct the following respectively:

[0175] 1) (Interest Point-User) Interaction Association Matrix If user u has visited point of interest p during the training session, then Otherwise, it is 0;

[0176] 2) Interest Point-Category Association Matrix If the category mapping of interest point p is c, then Otherwise, it is 0;

[0177] 3) Point of Interest-Region Association Matrix: If the region mapping of interest point p is r, then Otherwise, it is 0;

[0178] 4) Interest point transfer correlation matrix : A directed relationship from preorder to postorder within the session is adopted, and edges are not established across sessions. If the point of interest... Prior to in the conversation If it appears, then Otherwise, it is 0.

[0179] S202.3: Adjacency matrix normalization. To reduce the propagation bias caused by uneven node degree distribution, the degree of each branch matrix is ​​normalized. For any branch... Let the incidence matrix be . Its transpose is Constructing the interest point laterality matrix Compared with the lateral matrix of other heterogeneous nodes And obtain the normalized matrix. The same logic applies to the transition branches. After normalization, message propagation changes from accumulation to average aggregation based on the number of effective connections. The normalized branch matrices are then converted into sparse graph input tensors for subsequent two-stage propagation.

[0180] S202.4: Generate initial node embeddings, based on points of interest. The embedding is used as input to the branch gating to obtain the embedding representation of the interest points in each branch. To enable the model to learn more robust semantic representations without disrupting the propagation of the core interaction and temporal relationships, this invention performs masking operations on category nodes and region nodes.

[0181] S202.5: During the training phase, mask index sets are randomly sampled from the class node set and the region node set according to a preset ratio. and In this embodiment, the category mask ratio is 0.2, and the region mask ratio is 0.2.

[0182] S202.6: Replace the selected index position node representation with a learnable mask tag vector. or The replaced semantic node representation is denoted as... and This is used for subsequent hypergraph convolution. Original semantic node representation. , Used for subsequent mask self-supervised reconstruction.

[0183] S203: High-order heterogeneous hypergraph convolutional propagation

[0184] This step takes the interest point representation of each branch, the normalized branch matrix, and the masked semantic node representation as input, performs multi-branch propagation, and fuses the outputs of each branch to obtain sample-level user and interest point representations.

[0185] S203.1: Two-stage propagation through interaction, categories, and regional branches.

[0186] right Branches are executed according to a unified process: 1) Aggregation of interest points to superedges: Interest point messages are aggregated through... 1) Aggregate to other heterogeneous nodes; 2) Heterogeneous node feature fusion: fuse the aggregated message with the heterogeneous node's own representation to obtain the propagated heterogeneous node representation; 3) Hyperedge to interest point backhaul: via 4) Residual and random deactivation: Perform residual update and random deactivation on the interest point side and the heterogeneous node side respectively.

[0187] S203.2: Two-stage propagation of the transfer branch. The transfer branch propagates bidirectionally, utilizing... and Perform bidirectional propagation within the session, and update residuals and perform random deactivation.

[0188] S203.3: Inter-layer aggregation. For each branch, the outputs of each layer are aggregated to obtain the final interest point representation of the branch. In this example, the number of interaction branch layers is set to 2, the number of region branch layers is 2, the number of category branch layers is 1, and the number of transition branch layers is 4; the random inactivation rate is 0.3, and the interest point embedding dimension is 128. This aggregation mechanism is used to simultaneously preserve shallow local relationship information and deep structural semantic information.

[0189] S203.4: Sample-level representation generation. For any sample prefix... The corresponding interest point representations are extracted from the outputs of the four branches according to their indices. Since the prefix length *t* varies among different samples, masked pooling is performed on the prefix sequences of each branch. First, a unified fixed-dimensional representation is filled in, and then the effective positions are aggregated by mean to obtain a sample-level pooled representation. The interaction branch, based on the sample-level pooled representation, introduces and fuses user static vectors to form a sample-level user representation of "long-term stable preferences + current sample short-term intentions." The category, region, and transition branches retain the sample-level pooled representation as their corresponding branch sample-level user representation.

[0190] S203.5: Branch Weight Generation and Adaptive Fusion. Branch contributions are dynamically allocated across different samples to avoid information imbalance caused by fixed fusion weights. The user representations of each branch sample are input into a branch gating function to generate branch weights; then, the four semantic categories of interaction, category, region, and transition are adaptively weighted and fused according to these weights to obtain the final user representation for the current sample. .

[0191] S203.6: Construction of Interest Point Representations. The final interest point representations from the four branches are normalized and fused to obtain the final interest point representation matrix. .

[0192] S204: Self-supervised learning of semantic node masks

[0193] S204.1: Reconstruction of masked nodes. Based on the mask index set in S202. Extract the masked position vector from the semantic node representation after propagation from category branch and region branch. The data are fed into the category decoder and region decoder respectively, and the propagated semantic node representations are mapped back to the original semantic embedding space to obtain the corresponding reconstruction results. .

[0194] S204.2: Loss Construction and Joint Optimization. The semantic node representations retained before masking are used as the reconstruction supervision target. Region node reconstruction loss and category node reconstruction loss are calculated separately, and then weighted and summed according to their weight coefficients to obtain the semantic node mask self-supervised loss. This loss is then combined with the prediction loss to obtain the total loss. In this embodiment, the category reconstruction loss weight is 0.05, the region reconstruction loss weight is 0.05, and the similarity loss index parameter is 2.0.

[0195] S205: Next Point of Interest Prediction

[0196] This step is based on the user's final statement. Final representation of points of interest Interest points are scored and ranked. Based on the training samples, the model parameters are backpropagated and iteratively updated according to the total model loss. Then, the sample data from the test set is fed into the trained model to predict the next interest point.

[0197] The test was conducted on the TSMC2014-TKY dataset, with 30 training epochs, a batch size of 1024, and a learning rate of [missing information]. Weight decay With a learning rate scheduling coefficient of 0.1 and a random seed of 2026, the recommendation results are shown in Table 3.

[0198] Table 3. Examples of Recommendation Results in the TKY Dataset

[0199]

[0200] Based on the above embodiments, such as Figure 4 As shown, this invention also proposes a next interest point recommendation system based on high-order heterogeneous hypergraphs and node mask self-supervision, including:

[0201] The data preprocessing module is used to preprocess the user's original interaction data to obtain the user's behavior sequence sorted by time, segment the behavior sequence into sessions according to a preset time threshold, remove sessions with a length less than a first preset threshold and users with a number of sessions less than a second preset threshold, generate training samples and test samples, and construct a semantic mapping relationship between interest points and their category attributes and region attributes.

[0202] The hypergraph construction and masking module is used to construct a heterogeneous hypergraph containing user nodes, interest point nodes, region nodes, and category nodes. The hyperedge relationships include interest point-user interaction relationships, interest point-region relationships, interest point-category relationships, and interest point-interest point transition relationships, each corresponding to multiple semantic branches. The initial embedding of interest points is obtained through a gating mechanism to obtain the embedding representation of interest points in each branch. Random masking is performed on region nodes and category nodes, and the selected semantic node representation is replaced with a learnable mask vector to obtain the masked semantic node representation.

[0203] The heterogeneous hypergraph convolution and representation generation module is used to feed the embedded representations of interest points in each branch and the masked semantic node representations into the heterogeneous hypergraph convolutional network for multi-branch propagation to obtain the interest point representations of each branch and the representations of each semantic node; using the sample prefix sequence of the training samples as an index, the corresponding prefix representation is extracted from the interest point representations of each branch, and the sample-level representation is obtained through pooling; the sample-level user final representation is obtained through adaptive fusion; and the interest point representations of each branch are fused to obtain the candidate interest point final representation.

[0204] The self-supervised and joint optimization module is used to reconstruct the semantic node representation after propagation through the decoder. The semantic node representation before the mask is used as the supervision target, and the reconstruction loss is calculated to construct the node mask self-supervised loss. The node mask self-supervised loss is jointly optimized with the main task prediction loss.

[0205] The recommendation output module is used to calculate a prediction score based on the final representation of the sample-level user and the final representation of the candidate interest point using the test sample, and output the recommendation result of the next interest point based on the prediction score.

[0206] Based on the above embodiments, the present invention also proposes a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the next point of interest recommendation method based on high-order heterogeneous hypergraph and node mask self-supervision as described in any of the above embodiments.

[0207] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for recommending the next point of interest based on high-order heterogeneous hypergraphs and node mask self-supervision, characterized in that, include: The user's original interaction data is preprocessed to obtain the user's behavior sequence sorted by time. The behavior sequence is segmented into sessions according to a preset time threshold. Sessions with a length less than a first preset threshold and users with a number of sessions less than a second preset threshold are removed. Training samples and test samples are generated. A semantic mapping relationship between points of interest and their category attributes and region attributes is constructed. A heterogeneous hypergraph is constructed, comprising user nodes, interest point nodes, region nodes, and category nodes. The hyperedge relationships include interest point-user interaction relationships, interest point-region relationships, interest point-category relationships, and interest point-interest point transition relationships, each corresponding to multiple semantic branches. The initial embedding of interest points is obtained through a gating mechanism to achieve the embedding representation of interest points in each branch. Random masking is performed on region nodes and category nodes, and the selected semantic node representation is replaced with a learnable mask vector to obtain the masked semantic node representation. The embedded representations of interest points in each branch and the masked semantic node representations are fed into a heterogeneous hypergraph convolutional network for multi-branch propagation to obtain the interest point representations and semantic node representations of each branch. Using the sample prefix sequence of the training samples as an index, the corresponding prefix representations are extracted from the interest point representations of each branch. The sample-level representations are obtained through pooling and adaptive fusion to obtain the final user representation at the sample level. The interest point representations of each branch are fused to obtain the final representation of the candidate interest points. The propagated semantic node representation is reconstructed by the decoder. The corresponding semantic node representation before the mask is used as the supervision target. The reconstruction loss is calculated to construct the node mask self-supervised loss. The node mask self-supervised loss is jointly optimized with the main task prediction loss. Using the test samples, a prediction score is calculated based on the final representation of the sample-level users and the final representation of the candidate interest points. The next interest point recommendation result is then output based on the prediction score.

2. The next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision as described in claim 1, characterized in that, The preprocessing of the user's original interaction data includes: The behavior records of each user are arranged in ascending order of time; the behavior sequence is divided into multiple session sequences according to the preset time threshold between adjacent behavior records; sessions with a length less than the first preset threshold and users with a number of effective sessions less than the second preset threshold are removed; the effective sessions are divided into training session set and test session set according to time order, wherein the training samples construct behavior prefix sequences and target interest points by sliding window method, and the test samples construct behavior prefix sequences and target interest points by leave-one-out method.

3. The next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision as described in claim 1, characterized in that, The construction of the semantic mapping relationship between points of interest and their category and region attributes includes: Establish an interest point-category mapping based on the category identifier corresponding to each interest point; perform GeoHash geocoding with preset precision to discretize the latitude and longitude of each interest point to obtain the region code, and establish an interest point-region mapping; establish index mappings for category nodes and region nodes respectively to form a set of category nodes, a set of region nodes and corresponding semantic mapping relationships.

4. The next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision as described in claim 1, characterized in that, The construction and branch embedding generation of the heterogeneous hypergraph includes: A heterogeneous hypergraph is constructed, comprising four types of heterogeneous nodes: user nodes, interest point nodes, category nodes, and region nodes. Within this heterogeneous hypergraph, interest point-user interaction hyperedges, interest point-region relationship hyperedges, interest point-category relationship hyperedges, and interest point-interest point transition hyperedges are constructed, corresponding to the interaction association matrix, interest point-region association matrix, interest point-category association matrix, and interest point transition association matrix, respectively. The transition relationship hyperedges are constructed based on the pre-order to post-order access relationships of interest points within the training session and do not establish transition relationships across sessions. The initial interest point embeddings are used to generate interest point embedding representations corresponding to the interest point-user interaction branch, interest point-region relationship branch, interest point-category relationship branch, and interest point-interest point transition branch through a branch gating mechanism.

5. The next interest point recommendation method based on high-order heterogeneous hypergraph and node mask self-supervision as described in claim 1, characterized in that, The random masking process for region nodes and category nodes includes: According to the preset masking ratio, the set of node indices to be masked is randomly sampled from the set of region nodes and the set of category nodes respectively; the original node representation corresponding to the node to be masked is replaced with a learnable mask label vector; the original semantic node representation before masking is retained as the supervision target for subsequent reconstruction.

6. The next interest point recommendation method based on high-order heterogeneous hypergraph and node mask self-supervision as described in claim 1, characterized in that, The step of feeding the embedded representations of interest points in each branch and the masked semantic node representations into a heterogeneous hypergraph convolutional network for multi-branch propagation includes: Aggregation is performed on points of interest (POIs) to user nodes, POIs to category nodes, and POIs to region nodes. User nodes, category nodes, and region nodes fuse the aggregated POI representations with their own features, update their features, and then transmit them back to the POI nodes. In the transition branches, propagation is performed from source nodes to target nodes and from target nodes to source nodes. After multiple iterations, the outputs of each layer are aggregated to obtain the final POI representations for each branch and the final semantic node representations.

7. The next interest point recommendation method based on high-order heterogeneous hypergraph and node mask self-supervision as described in claim 4, characterized in that, The process of obtaining sample-level representations through pooling and then adaptively fusing them to obtain the final sample-level user representation includes: The variable-length prefix sequence is padded and pooled based on the effective position mask to obtain the sample-level fixed-length representation of each branch; the long-term user preference representation in the interaction branch is fused with the corresponding prefix representation; adaptive weights are calculated for the point of interest-user interaction branch, point of interest-region relationship branch, point of interest-category relationship branch and point of interest-point transfer branch respectively, and the sample-level representation of each branch is weighted and fused based on the adaptive weights to obtain the final sample-level user representation.

8. The next interest point recommendation method based on high-order heterogeneous hypergraphs and node mask self-supervision as described in claim 1, characterized in that, The propagated semantic node representation is reconstructed through a decoder, using the semantic node representation before the mask as the supervision target. The reconstruction loss is calculated to construct the node mask self-supervised loss, including: Extract the node vectors at the masked index positions from the propagated region node representation and category node representation; reconstruct the masked nodes using the region decoder and category decoder respectively; constrain the reconstruction results with the corresponding original node representations before masking to obtain the region reconstruction loss and category reconstruction loss; perform a weighted summation of the region reconstruction loss and category reconstruction loss to obtain the node mask self-supervised loss.

9. A next interest point recommendation system based on high-order heterogeneous hypergraphs and node mask self-supervision, characterized in that, include: The data preprocessing module is used to preprocess the user's original interaction data to obtain the user's behavior sequence sorted by time, segment the behavior sequence into sessions according to a preset time threshold, remove sessions with a length less than a first preset threshold and users with a number of sessions less than a second preset threshold, generate training samples and test samples, and construct a semantic mapping relationship between interest points and their category attributes and region attributes. The hypergraph construction and masking module is used to construct a heterogeneous hypergraph containing user nodes, interest point nodes, region nodes, and category nodes. The hyperedge relationships include interest point-user interaction relationships, interest point-region relationships, interest point-category relationships, and interest point-interest point transition relationships, each corresponding to multiple semantic branches. The initial embedding of interest points is obtained through a gating mechanism to obtain the embedding representation of interest points in each branch. Random masking is performed on region nodes and category nodes, and the selected semantic node representation is replaced with a learnable mask vector to obtain the masked semantic node representation. The heterogeneous hypergraph convolution and representation generation module is used to feed the embedded representations of interest points in each branch and the masked semantic node representations into the heterogeneous hypergraph convolutional network for multi-branch propagation to obtain the interest point representations of each branch and the representations of each semantic node; using the sample prefix sequence of the training samples as an index, the corresponding prefix representation is extracted from the interest point representations of each branch, and the sample-level representation is obtained through pooling; the sample-level user final representation is obtained through adaptive fusion; and the interest point representations of each branch are fused to obtain the candidate interest point final representation. The self-supervised and joint optimization module is used to reconstruct the semantic node representation after propagation through the decoder. The semantic node representation before the mask is used as the supervision target, and the reconstruction loss is calculated to construct the node mask self-supervised loss. The node mask self-supervised loss is jointly optimized with the main task prediction loss. The recommendation output module is used to calculate a prediction score based on the final representation of the sample-level user and the final representation of the candidate interest point using the test sample, and output the recommendation result of the next interest point based on the prediction score.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for recommending the next point of interest based on self-supervision using high-order heterogeneous hypergraphs and node masks as described in any one of claims 1 to 8.

Images