Project recommendation method for online tourism

Abstract

The invention discloses a project recommendation method for online tourism, which comprises the following steps: generating three types of hyperedges and adjacent matrixes thereof based on an interactive behavior sequence of each user on an online tourism platform, including session hyperedges, category hyperedges and user history hyperedges, respectively inputting the adjacent matrixes of the three types of hyperedges into a graph convolutional network, outputting the adjacency matrix, the adjacency matrix and the item embedding corresponding to the adjacency matrix, the item embedding, the item embedding and the intermediate representation of the item embedding; the project is embedded and input into the short-term interest model, and the short-term interest model outputs short-term interest representation; project embedding and project embedding are fused and then input into the long-term interest model, and the short-term interest model outputs long-term interest representation; performing fusion based on the short-term interest representation, the long-term interest representation and the item embedded intermediate representation to obtain a final user interest representation; the next interest item of the user and the preference item category of the future time period are predicted based on the user interest representation, and the generalization ability of the model is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of personalized recommendation technology, and more specifically, this invention relates to a method for recommending items for online travel. Background Technology

[0002] In recent years, with the rapid development of mobile internet, social media, and e-commerce platforms, users' online behavior has become increasingly fragmented and instantaneous, generating a large amount of short-cycle interaction data. Against this backdrop, traditional recommendation systems, which rely on long-term historical user data, have shown significant shortcomings in scenarios with high real-time requirements or where users are anonymous.

[0003] Session-based recommendation (SBR) has emerged to predict a user's most likely next item of interest based solely on their clicks, browsing, and other behavioral sequences within the current session, without relying on long-term user profiles. SBR shows broad application prospects in e-commerce, news, and short videos. However, in these practical applications, users often exhibit significant consistency in category preferences and trends in interest evolution across multiple sessions. These relatively stable preferences do not contradict the goals of SBR but rather complement short-term interests, especially in non-cold-start scenarios, enhancing the model's semantic generalization ability and recommendation accuracy. Therefore, integrating current session and user historical behavior modeling is becoming a crucial development direction for SBR.

[0004] To enhance the ability of models to model sequential relationships within a session, many methods have introduced Recurrent Neural Network (RNN) structures, such as GRU4Rec and NARM, which encode user behavior sequences using GRU or LSTM, achieving good results in modeling short-term interests. However, RNNs themselves suffer from problems such as vanishing gradients and insufficient ability to model long-distance dependencies, making it difficult to capture the dynamic features of long-term user interest evolution. In recent years, Graph Neural Networks (GNNs) have been widely used in SBR tasks, improving the ability to express user interests by constructing session graphs to model the jump relationships between items. Representative methods such as SR-GNN, GC-SAN, and TAGNN have performed well on multiple datasets. However, most of these methods only focus on the sequential structure within the current session, ignoring higher-level information such as cross-session semantics and category preferences. Summary of the Invention

[0005] This invention provides a method for recommending items for online travel, aiming to improve at least one of the above-mentioned problems.

[0006] This invention is implemented as follows: a method for recommending items for online travel, the method being as follows:

[0007] (1) Generate session hyperedges, category hyperedges and user history hyperedges based on the interaction behavior sequence of each user on the online travel platform, and construct the connection matrix corresponding to the session hyperedges, category hyperedges and user history hyperedges;

[0008] (2) The adjacency matrix of the session hyperedge Adjacency matrix of category superedges And the adjacency matrix of user history superedges The inputs to the hypergraph convolutional neural network are respectively, and the hypergraph convolutional neural network outputs the adjacency matrix respectively. Adjacency matrix and adjacency matrix Corresponding project embedding Project Embedding Project Embedding and project embedding The middle representation ;

[0009] (3) Embed the project Input a short-term interest model, and the short-term interest model outputs a short-term interest representation. ;

[0010] (4) Embed the project Embedded with project After fusion, the input is the long-term interest model, and the output is the long-term interest representation from the short-term interest model. ;

[0011] (5) Based on short-term interest representation Long-term interest representation and project embedding The middle representation By integrating these elements, the final representation of user interests can be obtained. ;

[0012] (6) User interest representation Predict the user's next interest item and the preferred item categories for future time periods.

[0013] Furthermore, the session hyperedge set Construction method:

[0014] Each visit by a user on an online travel platform will be divided into a session, which contains several items. All the corresponding nodes of each item appearing in a session will be connected by an edge to form a session hyperedge, which will be added to the session hyperedge set. middle.

[0015] Furthermore, the category superedge set Construction method:

[0016] Based on the mapping table between projects and project categories, each project corresponds to a project category. Nodes belonging to the same category in different time periods are connected by an edge to form a category hyperedge, which is then added to the category hyperedge set. middle.

[0017] Furthermore, the user's historical super-edge set Construction method:

[0018] Based on all users' historical sessions, the click frequency of items across all historical interactions is counted. Items with click frequencies exceeding a set threshold are added to the long-term interest item set. Items in the long-term interest item set are considered long-term interest items. Nodes corresponding to all items in the long-term interest item set are connected by an edge to form a user history hyperedge, which is then added to the user history hyperedge set. .

[0019] Furthermore, the hypergraph convolutional neural network consists of three hypergraph convolutional layers, which are the adjacency matrices of the class hyperedges. After inputting into the hypergraph convolutional neural network, the output of the second hypergraph convolutional layer is the item embedding. The middle representation .

[0020] Furthermore, a gated recurrent unit is used as the short-term interest model, and a Transformer model is used as the long-term interest model.

[0021] Furthermore, the adjacency matrix Each column corresponds to a set of session superedges. A session hyperedge in the context of a session is defined such that the element value of the item appearing in the corresponding column is set to 1, otherwise it is set to 0.

[0022] Adjacency matrix Each column corresponds to a set of superedges for that category. A category superedge in the table is used to set the element value of the corresponding column of the item appearing in the category superedge to 1, otherwise it is set to 0;

[0023] Items appearing in the user's history superedge in the adjacency matrix Set the value of the element in the corresponding column to 1, otherwise set it to 0.

[0024] Furthermore, the method for predicting a user's next interest item is as follows:

[0025] Based on user interest representation Predicting the user's next interest item as a project Click preference score Select items with high click preference scores as candidate items. ;

[0026] Output candidate items using softmax normalization. Click probability of click preference score The item with the highest click rate is selected as the user's next item of interest;

[0027] Among them, click preference score The specific calculation formula is as follows:

[0028] ;

[0029] in, For the project Learnable embedding vectors.

[0030] Furthermore, user preferred item categories in future time periods. Predicted probability The specific details are as follows:

[0031] ;

[0032] in, , These are the weight parameters and bias parameters of the linear projection layer.

[0033] This invention represents items within the same session, cross-session items of the same category, and frequently interacted items as multi-type hyperedges. It captures high-order adjacency relationships through a hierarchical hypergraph convolutional neural network, and combines GRU and Transformer to model short-term interests and long-term evolutionary preferences respectively. It also achieves dynamic fusion of multi-scale interests through a task-aware gating mechanism, introduces category prediction as an auxiliary task, and conducts joint training with the main recommendation task to enhance the model's generalization ability. Attached Figure Description

[0034] Figure 1 This is a flowchart of a project recommendation method for online tourism provided in an embodiment of the present invention;

[0035] Figure 2 The ablation experiments provided for embodiments of the present invention include, (a) performance evaluation of each ablation variant on HR@10, and (b) performance evaluation of each ablation variant on MRR@10.

[0036] Figure 3The sensitivity experiment of the loss balance factor λ between the recommendation task and the auxiliary task provided in the embodiment of the present invention is shown in which (a) is the change of different datasets on HR@10, (b) is the change of different datasets on MRR@10, and (c) is the change of different datasets on Accuracy@10.

[0037] Figure 4 The experiment on the impact of the number of convolutional layers provided in the embodiments of the present invention includes (a) the recommendation performance of different datasets on HR@10 and (b) the recommendation performance of different datasets on MRR@10.

[0038] Figure 5 The experiment on the impact of Transformer depth provided in the embodiments of the present invention includes (a) the performance of different datasets on HR@10 and (b) the performance of different datasets on MRR@10.

[0039] Figure 6 The experiment on the impact of Transformer embedding dimension provided in the embodiments of the present invention shows that (a) is the performance of different datasets on HR@10, and (b) is the performance of different datasets on MRR@10.

[0040] Figure 7 The experiment on the impact of GRU embedding dimension provided in the embodiments of the present invention, wherein (a) is the performance of different datasets on HR@10, and (b) is the performance of different datasets on MRR@10;

[0041] Figure 8 The experiment on the impact of the number of attention heads in the Transformer provided in the embodiments of the present invention includes (a) the performance of different datasets on HR@10 and (b) the performance of different datasets on MRR@10. Detailed Implementation

[0042] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0043] This invention proposes an interest evolution conversation recommendation model (MTIESR) that integrates multi-task learning. Through structural modeling enhancement, temporal modeling and multi-task collaborative optimization, it improves the conversation recommendation performance from multiple levels. (1) A multi-level hypergraph structure is designed to integrate three types of semantic relations: in-conversation item co-occurrence, cross-conversation category consistency and user historical high-frequency behavior preferences, to comprehensively model the user's structured interests. In order to model short-term and long-term preferences at the same time, a parallel structure is constructed, and finally, multi-scale interest representations are integrated through a gating mechanism. (2) A multi-task learning framework is proposed to jointly optimize the next item recommendation and the next category prediction tasks. Among them, category prediction, as an auxiliary task, can enhance semantic understanding and improve generalization ability, and improve recommendation accuracy through category trend adjustment.

[0044] Figure 1 The flowchart of the project recommendation method for online tourism provided in this embodiment of the invention is as follows:

[0045] (1) Generate session hyperedges, category hyperedges and user history hyperedges based on the interaction behavior sequence of each user on the online travel platform, and construct the connection matrix corresponding to the session hyperedges, category hyperedges and user history hyperedges.

[0046] User set ,in, This represents the nth user in the user set. The total number of users in a user set, or project set. ,in, This represents the m-th item in the project set. This indicates the number of projects in the project set, and the project category set. ,in, This represents the p-th project category in the project category set. This indicates the number of project categories in the project category set. (User) The sequence of interactive behaviors on online travel platforms is a time-ordered set of conversations. ,in, Represents a set of sessions The t-th session, session ,in, Representing a session The i-th item that appears in the list belongs to the item set and the item category to which the item belongs belongs to the item category set. The goal of this invention is to predict the next item that the user is most likely to click and the item category that the user may prefer in the next time period.

[0047] To effectively capture the multi-scale interest association structure of users, this invention constructs a heterogeneous hypergraph that integrates three types of semantic relations. The heterogeneous hypergraph includes three types of hyperedges, including: a set of conversation hyperedges. Category superedge set and user history super-edge set The specific construction process is as follows:

[0048] Session hyperedge set Construction method: Based on the sequence of all user interaction behaviors on the online travel platform, each visit is divided into a session. Each session contains several items (such as flights, hotels, attractions). All the corresponding nodes of the items appearing in each session are connected by an edge to form a session hyperedge, which is then added to the session hyperedge set. In the diagram, each session superedge represents a target item that appears once in a user session.

[0049] Category superedge set Construction method: Based on the mapping table between projects and project categories, each project corresponds to a unique project category (such as "flights", "hotels", "attractions"). Nodes corresponding to projects belonging to the same category in different time periods are connected by an edge to form a category hyperedge, which is then added to the category hyperedge set. In this approach, category hyperedges are used to characterize the consistency of user categories across different sessions. This method can capture users' long-term category interest preferences, avoiding the problem in traditional methods where category information is merely used as an additional feature and cannot be used in a structured way, thereby improving the model's ability to model cross-session interest evolution.

[0050] User history super-edge set Construction method: Based on the historical sessions of all users, the click frequency of items in all historical interactions is counted. Items with a click frequency exceeding a set threshold are placed into the long-term interest item set. Items in the long-term interest item set are long-term interest items. The nodes corresponding to all items in the long-term interest item set are connected by an edge to form a user history hyperedge, which is then added to the user history hyperedge set. This method can reflect users' long-term stable preferences and alleviate the cold start and data sparsity problems through high-frequency interest modeling.

[0051] Set of session hyperedges Category superedge set and user history super-edge set Add to superedge set In the middle, set the superedge set Weight of each superedge Weight The weights are set based on the click frequency of the corresponding hyperedge. Higher frequency results in higher weights. Softmax normalization is performed on all hyperedge weights to obtain standardized weights. By normalizing the weights of all hyperedges, the influence of different types of hyperedges in the feature propagation process can be balanced, and the excessive dominance of a certain type of hyperedge in structural modeling can be avoided.

[0052] In this embodiment of the invention, based on the session hyperedge set Construct the adjacency matrix of the session hyperedge Adjacency matrix Each column corresponds to a set of session superedges. A session hyperedge in the context of a collection of hyperedges sets the value of an item appearing in that hyperedge to 1 in the corresponding column, otherwise to 0; based on the category hyperedge set. Construct the adjacency matrix of category superedges Adjacency matrix Each column corresponds to a set of superedges for that category. A category hyperedge in the data structure is used to set the element value of the corresponding column of an item appearing in that category hyperedge to 1, otherwise to 0; based on the user's historical hyperedge set. Construct the adjacency matrix of user history superedges Items appearing in the user's history superedges are in the adjacency matrix Set the value of the element in the corresponding column to 1, otherwise set it to 0.

[0053] (2) The adjacency matrix of the session hyperedge Adjacency matrix of category superedges And the adjacency matrix of user history superedges The inputs to the hypergraph convolutional neural network are respectively, and the hypergraph convolutional neural network outputs the adjacency matrix respectively. Adjacency matrix and adjacency matrix Corresponding project embedding Project Embedding Project Embedding and project embedding The middle representation ;

[0054] In this embodiment of the invention, the hypergraph convolutional neural network consists of three hypergraph convolutional layers, which convert the adjacency matrix of category hyperedges. After inputting into the hypergraph convolutional neural network, the output of the second hypergraph convolutional layer is the item embedding. The middle representation To fully utilize the multi-semantic relationships inherent in the hypergraph structure, this invention employs hypergraph convolution operations for feature propagation and high-order aggregation of item nodes. By utilizing adaptive hypergraph convolutional layers, it can capture the high-order relationships between users and items in the hypergraph. Hypergraphs differ significantly from traditional graphs because edges in a hypergraph can connect multiple nodes, thus enabling the exploration and aggregation of features to reveal shared preferences between users and items. The edges of a hypergraph are also called hyperedges. Assuming the initial embedding of an item is... , The embedding dimension of the project is represented by the hypergraph structure, which is represented by the adjacency matrix. The output of the l-th hypergraph convolutional layer is... It is expressed as follows:

[0055] ;

[0056] in, This represents the output of the (l-1)th hypergraph convolutional layer. For the first Learnable parameters of each hypergraph convolutional layer and These represent the degree matrices of nodes and hyperedges, respectively. The activation function is LeakyReLU. This represents the adjacency matrix of the input. hour, This represents the weight corresponding to the session hyperedge, with the input being the adjacency matrix. hour, This represents the weights corresponding to the category hyperedges, given an adjacency matrix as input. hour, This indicates the weight corresponding to the user's historical superedge.

[0057] (3) Project Embedding Input a short-term interest model, and the short-term interest model outputs a short-term interest representation. ;

[0058] The current session sequence directly reflects the user's immediate goals, exhibiting clear sequential dependencies and local interest jumps. In real-world recommendation scenarios, users often browse a particular type of item in a concentrated manner, accompanied by rapid interest shifts. Therefore, the model needs to balance sequential modeling capabilities with preference transfer representation capabilities. To this end, this invention employs item embedding based on hypergraph convolution output. As input sequences, short-term interest models are modeled using gated recurrent units (GRUs). GRUs effectively model state transitions through gating mechanisms and perform well in capturing the interest dynamics of conversation sequences. Item embeddings are then extracted from the output of the hypergraph convolutional neural network. Extract the t-th item that appears in the current session sequentially. The convolution representation, The data is sequentially input into the gated recurrent unit (GRU). The state update mechanism of the GRU is as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] in, This represents the hidden state at time step k and time step (k-1). , These represent the update gate and reset gate at the k-th time step, respectively. For element-wise multiplication, This represents the candidate hidden state at the k-th time step. This represents the activation function. , This represents the learnable matrix of the update gate; , The learnable matrix representing the reset gate; , The learnable matrix of the hidden layer represents the short-term interest of the hidden state output at the last time step. The GRU model can capture changes in interest based on the order of behavior in the current session, alleviate the gradient vanishing problem through a gating mechanism, and effectively remember short-term click intentions.

[0064] (4) Embed the project Embedded with project After fusion, the input is the long-term interest model, and the output is the long-term interest representation from the short-term interest model. ;

[0065] This invention views relatively stable preferences as users' long-term interests, and modeling these long-term interests is significant for understanding their stable preferences and behavioral evolution trends. In real-world recommendation scenarios, users may click on items belonging to the same category or semantically related multiple times across different sessions; these cross-session clicks constitute their long-term interests. However, these interests do not always appear consecutively and are often affected by factors such as semantic jumps and time decay. Therefore, modeling users' long-term interests requires not only incorporating historical behavioral information but also capturing non-local dependencies and preference transfer capabilities. Based on this, this invention employs the Transformer model in its long-term interest model to fully leverage its advantages in non-sequential modeling and global dependency modeling. It receives item embeddings formed by category hyperedges. Embedded with user history hyper-edge project The data is then concatenated and used as input for a long-term interest model. The final output is a vector representing the user's long-term interests.

[0066] Each historical action has a fixed position (i.e., the order of occurrence) in the user's behavior trajectory, and a learnable position embedding matrix is ​​introduced. , This indicates the maximum length of the session, and each long-term interest item corresponds to an embedding vector. This includes the corresponding location information, and the specific calculation is as follows:

[0067] ;

[0068] ;

[0069] in, Let represent the embedding representation of the i-th long-term interest item in the set of long-term interest items. , This indicates the number of long-term interest items in the long-term interest item set. This represents a long-term interest item embedding with location-encoded information.

[0070] A multi-head self-attention mechanism is used to model global dependencies between historical behaviors in different attention subspaces. For the input long-term interest item embeddings... First, it is mapped to query, key, and value vectors using three sets of projection matrices. The attention weights between each position are calculated using a scaled dot product.

[0071] ;

[0072] in, This represents the data embedding dimension. To improve the model's expressive diversity, H attention heads are introduced to learn different attention patterns in parallel, with each attention head outputting attention weights. After concatenation, multi-head attention weights are obtained through linear mapping. To mitigate gradient vanishing and stabilize deep network training, each submodule employs residual connections and layer normalization strategies. This applies to the output of the multi-head attention module. It is expressed as follows:

[0073] ;

[0074] in, Representation layer normalization. This structure allows the long-term interest model to maintain a consistent input-output distribution when learning deep features, thereby improving training efficiency and generalization ability. After completing the multi-head self-attention representation, the feedforward neural network... It consists of two linear transformations, with a non-linear activation function used for mapping in between, and the output of the feedforward neural network... It is expressed as follows:

[0075] ;

[0076] in, , These represent the weight matrices of the first and last layers of the feedforward neural network, respectively. , To represent the bias parameters of the first and last layers of the feedforward neural network, ReLU is used as the activation function. Subsequently, residual connections and layer normalization are applied to output the long-term interest representation. , means as follows:

[0077] .

[0078] (5) Based on short-term interest representation Long-term interest representation and project embedding The middle representation By integrating these elements, the final representation of user interests can be obtained. ;

[0079] This invention designs two tasks in parallel: item prediction and category prediction. It introduces a task-aware gating mechanism on top of the traditional multi-task framework to achieve deep coupling between the auxiliary task and representation learning and interest evolution modeling. Specifically, in short-term interest representation... Representation of long-term interests Introducing project embedding during the integration process The middle representation It is used as a gating signal, enabling the category task to influence the formation of the interest representation, as follows:

[0080] ;

[0081] ;

[0082] in, , Indicates learnable parameters, Indicating short-term interest Long-term interest representation Intermediate representation of category prediction splicing, To integrate weights, This represents the final expression of user interests.

[0083] (6) User interest representation Predict the user's next interest item and the preferred item categories for future time periods.

[0084] Based on user interest representation Simultaneously, two outputs are generated: a probability distribution prediction of the next item of interest and a prediction of the user's preferred item category in the next stage. The main task is to predict the next item that the user is most likely to click in the current session. The user's click preference score for the next item is calculated using the following formula:

[0085] ;

[0086] in, This indicates the user's next interest item as a project. Click preference score For the project Learnable embedding vectors are used to select items with high click preference scores as candidate items. .

[0087] Output candidate items using softmax normalization. The click probability distribution of the click preference score is represented as follows:

[0088] ;

[0089] in, This indicates the number of candidate items to be output.

[0090] In addition to the main task, an auxiliary item category prediction task is introduced to predict the distribution of item categories that users may prefer in the next stage. This invention directly incorporates item category embedding into the gating mechanism, where category nodes participate in the fusion of short-term and long-term interests in the forward propagation, while being constrained by the category prediction loss during backpropagation, thus forming a "task-aware" dynamic gating. A linear projection layer is used to map the user representation to the item category space, as shown below:

[0091] ;

[0092] in, Indicates the user's preferred item categories for a future time period. The predicted probability, , These are the weight parameters and bias parameters of the linear projection layer.

[0093] This invention combines the loss functions of the two tasks mentioned above into a comprehensive loss function. And introduce weighting coefficients The impact of controlling project category prediction tasks on the overall loss function. Specifically as follows:

[0094] ;

[0095] in, Indicates the predicted loss of the interest item. This indicates the predicted loss based on the preferred item category.

[0096] Predicted loss of interest items Using cross-entropy loss, the specific representation is as follows:

[0097] ;

[0098] in, This represents the predicted probability of the next item of interest, as output during the training process. This indicates that the label vector is formed using one-hot encoding. .

[0099] Predicted loss by preference item category Using cross-entropy loss, the specific representation is as follows:

[0100] ;

[0101] in, The preferred item categories for future time periods represented by the output of the training process. Predicted probability This indicates that the label vector is formed using one-hot encoding. .

[0102] Due to project embedding The middle representation The input is directly fed into the task-aware gating system. The magnitude of λ not only affects the auxiliary task itself but also determines the constraint strength of category prediction on the gating parameter updates during backpropagation, thereby dynamically adjusting the fusion method of short-term and long-term interests. This makes the category prediction task not only an additional supervisory signal in this model but also a key component that directly affects the gating mechanism.

[0103] Three real-world datasets with varying sizes, number of items, categories, and click intervals were used in the experiment. Cosmetics is a dataset from a mid-sized online cosmetics store. Due to its large size, only data from October was used. Multi-Category is a dataset from a large multi-category online store. Fliggy released the UBF dataset, a large-scale, desensitized user behavior dataset designed to improve personalized travel recommendations. For these three datasets, the last 10% of the data was used as the test set, and the remainder as the training set. To ensure fairness, following previous work, items and categories appearing less than five times, and sessions shorter than 5 or longer than 50 characters, were filtered out. Items appearing in the test set but not in the training set were also filtered out. Statistics for the datasets, including click count, number of training sessions, number of test sessions, number of items, and number of categories, are detailed in Table 1.

[0104] Table 1 Dataset Analysis

[0105]

[0106] We used commonly used metrics such as Hit Ratio (HR@K), Mean Reciprocal Rank (MRR@K), and Accuracy (ACC@K) to evaluate my recommendation list, setting K to 10 or 20. HR@K and MRR@K are used to evaluate the model's performance in the next item recommendation task, while ACC@K is used to evaluate the model's accuracy in the next category prediction task.

[0107] Hit Ratio (HR@K) is a metric used to evaluate whether a recommendation system successfully recommends a target item among the top K candidate results. The HR@K value indicates the proportion of samples in the test set whose target item appears in the top K positions of the corresponding recommendation list. If the target item appears in any position in the list, it is considered a "hit," regardless of its specific ranking. Therefore, this metric focuses more on the coverage of recommendations than on ranking quality; a higher HR@K indicates that the model more fully covers user interests and the better the recommendation accuracy.

[0108] The Mean Reverse Rank (MRR@K) is a recommendation metric based on rank position, which better reflects the accuracy of the algorithm's predictions. MRR considers the rank of positive samples in the recommendation list; the higher the rank of a positive sample, the greater its contribution to the MRR. A higher MRR value indicates more accurate predictions when the predicted result appears in the first few positions of the recommendation list.

[0109] Accuracy@K evaluates the model's accuracy in predicting the user's next category of interest in an auxiliary task. Class prediction is treated as a multi-class classification task, determining in each sample whether the true target class appears in the top K class probability rankings.

[0110] All experiments were conducted on a single NVIDIA RTX 3060 GPU. The proposed MTIESR model was implemented using the PyTorch framework, and a consistent training configuration was maintained across all datasets to ensure fair comparisons. Model training used the Adam optimizer with an initial learning rate of 0.001, a training batch size of 256, and an evaluation batch size of 1024. The embedding vector dimension was set to 256 based on parameter sensitivity experiments, and the number of attention heads was set to 6. To prevent overfitting, L2 regularization (λ = 1e-6) and a Dropout mechanism were added to the embedding layer and each submodule, with the Dropout ratio set to 0.1.

[0111] During model training, if the validation set performance does not improve within 10 consecutive epochs, an early stopping strategy is triggered. Key structural parameters were optimized using a grid search on the validation set, with default settings of L=2, T=3, and λ=0.7. Their impact will be further explored in the parameter sensitivity analysis section. All experiments were repeated under the same data preprocessing procedure and random seed control to ensure the stability and reproducibility of the comparisons.

[0112] This invention compares the MTISER model with five baseline models. These baseline models cover a variety of modeling approaches in conversational recommendation, including graph neural networks (GNNs), hypergraph structures, and multi-task recommendation methods. A brief overview of these baseline models is provided below:

[0113] SR-GNN—Applies Graph Neural Networks (GNNs) to conversation recommendation, capturing complex transition relationships between items by constructing a conversation graph and modeling user interests using Gated Graph Neural Networks (GGNNs). It proposes a dynamic fusion mechanism between global preferences and current interests based on the conversation graph. Case4SR—Proposes a conversation recommendation framework based on contrastive learning, generating positive and negative sample pairs through data augmentation and jointly optimizing with self-supervised tasks and the main recommendation task. It is the first to introduce contrastive learning into conversation recommendation to alleviate the data sparsity problem. CM-HGNN—Based on Heterogeneous Graph Neural Networks, it models multimodal information such as items, categories, and user behavior in a conversation as a heterogeneous graph, proposing a cross-modal attention mechanism to fuse multi-source features. IEGT—Proposes a graph Transformer architecture for enhanced temporal modeling, explicitly introducing time interval information between items, and modeling changes in user preferences through positional offset and time-aware multi-head attention modules. STEM—Proposes a multi-task recommendation framework based on task embedding, introducing learnable Task Embedding to guide shared feature extraction and task branch optimization. By using a task-aware mechanism to enable information flow and expression sharing between tasks, the performance of multi-behavior recommendation is significantly improved, and it is suitable for joint modeling of multiple objectives such as clicks, likes, and conversions.

[0114] The baseline comparison results are shown in Table 2, which presents the recommendation performance of the proposed MTIESR model and five baseline models on the Cosmetics, Multi-category, and UBF datasets. The focus is on the two core metrics, P@K and MRR@K. The best results are bolded, and the second-best results are underlined. It can be observed that MTIESR achieves optimal results on all datasets and metrics, demonstrating comprehensive and robust recommendation capabilities.

[0115] Table 2 Performance comparison of evaluation metrics across all datasets

[0116]

[0117] First, from the perspective of recommendation methods based on graph structure modeling, SR-GNN constructs a directed click graph within a session and uses the GNN structure to capture local sequential dependencies, making it suitable for modeling shorter sequences. However, it lacks the ability to model temporal dynamics and category context, thus performing poorly on datasets like Cosmetics that are dominated by short-term preferences, and exhibiting low HR@20 on long-behavioral datasets like UBF, making it difficult to capture higher-order interest transfers.

[0118] On the other hand, Case4SR enhances the expressive power of conversation representation by using metapath embedding and attention mechanisms, and introduces category features for weighting. However, since its category modeling is based on concatenation rather than structural learning, although it performs slightly better on multi-category datasets, it cannot compete with structural modeling methods on datasets with deeper semantic levels such as UBF. Its HR@20 is only 60.54%, which is significantly lower than MTIESR.

[0119] CM_HGNN is a hypergraph neural network approach that models various behaviors as hyperedges, capturing more complex high-order relationships between items compared to SR-GNN. It achieved moderate performance on Cosmetics. However, due to its hyperedge design's lack of differentiation between interest granularity and the absence of explicit introduction of category and temporal factors, it lacks generalization ability on multi-category and UBF (Unbounded Component Filtering), with MRR@20 significantly lower than MTIESR.

[0120] In the structural enhancement model, IEGT combines graph neural structures with time-aware mechanisms, modeling time intervals and order through edge weights, making the model more accurate in capturing users' historical preferences. In multi-category scenarios, its HR@20 reaches 68.96% and MRR@10 reaches 38.88%, leading all baselines. However, IEGT lacks category guidance and semantic hierarchical structure modeling, resulting in lower ranking performance than MTIESR in UBF scenarios with imbalanced multi-interest semantic distributions.

[0121] STEM further integrates graph modeling and sequence modeling, enhancing its ability to represent short-term and some long-term interests through hypergraph convolution and temporal feature awareness mechanisms. On the Cosmetics dataset, its HR@10 reaches 62.47% and MRR@20 is 27.16%, outperforming most baseline models. However, STEM does not explicitly introduce class relationships or long-term historical modeling paths, and when faced with datasets like UBF that require long-term interest tracking, its MRR@20 still lags behind MTIESR.

[0122] In summary, MTIESR outperforms existing baselines on all three datasets, with particularly significant advantages on datasets with complex categories and more dynamic interest evolution, such as Multi-category and UBF, validating the effectiveness of the proposed task-aware gating and multi-task optimization strategies.

[0123] To further analyze the impact of each key module on the overall performance of the model, several ablation variants were designed based on the complete MTIESR model, including: retaining only the short-term interest modeling module (GRU-only), retaining only the long-term interest modeling module (Transformer-only), removing the auxiliary task category prediction (without CatTask), removing the gating mechanism for category input (Gate-input-ablation, where the gating only relies on short-term and long-term interests and no longer reads category representations), and blocking the gating mechanism for the inverse constraint of category tasks (Gate-stop-grad, where the gating reads the forward information of the category representation but does not allow the category loss to be backpropagated to the shared layer). The performance of each variant on the Cosmetics, Multi-category, and UBF datasets was evaluated using HR@10 and MRR@10. The experimental results are shown in Figure 2. The complete MTIESR model achieved the best performance on all datasets, indicating that each sub-module has a synergistic gain effect in improving the ability to model and rank user interests. When the Transformer branch is removed and only GRU is retained (GRU-only), the model cannot model the interest transfer and structured preference aggregation path across time slices. Because it loses the structural propagation capability between category hyperedges and historically high-frequency items, the model relies solely on click order within a local session to generate interest representations, failing to perceive users' long-term stable preferences. Therefore, this variant exhibits the most significant degradation on datasets with large behavioral and category spans, such as UBF and Multi-category datasets.

[0124] While the Transformer-only variant retains the ability to model the structure of historical interests, the removal of the GRU module makes it difficult for the model to accurately capture fine-grained sequential preferences and interest transitions within the current session. This is particularly evident on datasets like Cosmetics, which feature short-term behaviors and high-frequency jumps, resulting in a significant performance drop. This demonstrates that local modeling of short-term interests is crucial for improving recommendation accuracy.

[0125] Removing the category prediction task (without CatTask) deprives the model of semantic-level supervision, preventing it from learning the interest trend prediction ability of "the category to which the target item belongs" through auxiliary loss. This not only weakens generalization ability but also reduces the model's ability to capture the structure of category consistency behavior, as evidenced by a more significant decrease in MRR@10, indicating a weakening of the ranking signal.

[0126] In further comparative experiments of task-aware gating mechanisms, Gate-input-ablation (removing the gating input of class representations) leads to performance degradation, approaching that of removing the class prediction task, indicating that class representation as a forward signal is crucial for gating to make reasonable weights when short-term and long-term interests conflict. Gate-stop-grad (blocking backward supervision of class tasks) performs between Gate-input-ablation and the complete model, indicating that backward constraints on class tasks are also an important factor in improving the quality of shared representations. Overall, the effectiveness of task-aware gating stems from the synergistic effect of forward supply of class representations and backward supervision of class tasks.

[0127] To maximize the demonstration of model performance, this paper investigates the following hyperparameters while ensuring that other hyperparameters are in their optimal state: (1) balancing factor λ; (2) number of hypergraph convolution layers L; (3) number of Transformer layers T; (4) Transformer embedding dimension; (5) GRU embedding dimension; and (6) number of attention heads H. Through experiments, the optimal range of values ​​for these parameters is determined, and their specific impact on model performance is understood.

[0128] In the multi-task learning framework, the weighting of the loss between the recommendation task and the auxiliary task is controlled by a balancing factor λ. To gain a deeper understanding of the combined impact of λ on the main and auxiliary tasks, we conducted sensitivity experiments on three datasets, setting λ to {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0}, and recording the trends of three metrics: HR@10, MRR@10, and Accuracy@10. The experimental results show... Figure 3 As shown. From Figure 3It can be seen that as λ increases from 0.1 to 0.5–0.6, both HR@10 and MRR@10 on the three datasets show a continuous upward trend, reaching a peak in this range; when λ further increases (e.g., 0.8 and above), the ranking performance begins to decline. This indicates that moderately introducing semantic supervision from the category prediction task can significantly improve recommendation performance, because task-aware gating relies on category representation when fusing long-term and short-term interests, and the auxiliary task provides reliable training signals, thereby enhancing the model's generalization ability in the main task. However, when λ is too large (e.g., λ=1.0), the model training becomes overly biased towards category prediction, leading to suppressed recommendation performance.

[0129] Further analysis of the Accuracy@10 trend reveals that the class prediction accuracy of all three datasets monotonically increases with increasing λ, indicating that higher class loss weights are beneficial for optimizing auxiliary tasks, but do not directly guarantee a sustained improvement in recommendation performance. Specifically, UBF, due to its longer sequences and more complex class structures, achieves the highest accuracy at λ=1.0, but the corresponding HR@10 and MRR@10 show a significant decline, reflecting the trade-off between tasks. In contrast, the Cosmetics dataset shows a smaller improvement in accuracy, indicating that class supervision has limited effectiveness in short-sequence scenarios. Overall, this experiment validates the effectiveness of multi-task collaborative training under a task-aware gating mechanism. Particularly around λ=0.5–0.6, the model achieves an optimal balance between the main and auxiliary tasks.

[0130] To explore the impact of different numbers of convolutional layers on the model's recommendation performance, we tested the performance on three datasets. Experiments were conducted, and the results for HR@10 and MRR@10 were recorded separately, as follows: Figure 4As shown, most datasets achieve the best HR@10 and MRR@10 at L=2, indicating that second-order structural information best enhances the model's ability to model higher-order interest associations. For example, the Cosmetics dataset achieves an HR@10 of 69.90% and an MRR@10 of 32.01 at L=2. Performance begins to decline when L>2, mainly due to overly smooth node embeddings caused by multiple convolutions, leading to reduced discriminativity. Furthermore, deep propagation may introduce redundant noise, weakening training stability. In contrast, the more structurally complex UBF dataset shows similar performance at L=2 and L=3 (HR@10 of 69.70% and 69.81%, respectively), indicating that its longer behavioral chains and richer category structure are robust to multi-layer propagation. However, performance still declines when the number of layers increases to 4. In summary, appropriately increasing the convolution depth can improve the expressive power of the hypergraph structure and enhance the model's ability to capture higher-order semantic relationships in conversations, but excessive layers can lead to over-smoothing and redundant noise. Based on the trade-off between performance and efficiency, this paper ultimately chooses L=2 as the default setting.

[0131] To evaluate the impact of Transformer depth on long-term interest modeling, we set the number of layers. The changes in HR@10 and MRR@10 were measured on three datasets, and the results are as follows: Figure 5 As shown, the Cosmetics dataset peaks at T=2 for both HR@10 and MRR@10, indicating that a two-layer Transformer is sufficient to model long-term preferences in short sequences. For the Multi-category dataset, T=2 also yields the best performance. On the UBF dataset, the performance of T=2 and T=3 is almost equivalent, demonstrating that longer behavioral sequences are robust to deeper structures. When the number of Transformer layers is further increased to T=4, the performance of all three datasets decreases, indicating that excessively deep structures may introduce redundancy and exacerbate the risk of overfitting. The decrease is most pronounced on the Cosmetics dataset, validating that short sequences are more sensitive to modeling depth. Therefore, this paper consistently selects T=2 as the default configuration across different datasets to balance performance and efficiency.

[0132] To evaluate the expressive power of the Transformer in the long interest modeling module, we set different embedding dimensions: {64, 128, 192, 256, 320}, and performed performance evaluations on three datasets. The results are as follows: Figure 6As shown, the embedding dimension of the Transformer has a significant impact on model performance. When the dimension increases from 64 to 192, both HR@10 and MRR@10 steadily improve, indicating that higher-dimensional representations help capture semantic relationships in long-term interests. Further increasing to 256, peak performance is achieved on all three datasets (e.g., HR@10=69.8 and MRR@10=47.7 on UBF), validating the advantage of larger capacity in long sequence modeling scenarios. However, when the dimension continues to increase to 320, performance no longer improves, and some datasets (such as Cosmetics) even show a slight decline. This is mainly because excessively large embedding dimensions introduce parameter redundancy and noise, reducing the discriminative power of the representation while increasing the training burden. Therefore, 256 dimensions achieve a good balance between expressive power and computational efficiency.

[0133] To investigate the impact of GRU embedding dimension on model performance in the short-term interest modeling module, this paper sets the hidden dimensions to {64, 128, 192, 256, 320, 384}. The results are as follows: Figure 7 As shown, the model performance significantly improves in the initial stage as the GRU hidden dimension increases. When the dimension increases from 64 to 192, both HR@10 and MRR@10 continue to rise, indicating that a larger hidden space helps capture complex dependencies in short sequences. Further increasing to 256, the model performance reaches its optimum (HR@10=69.9, MRR@10=32.0 on Cosmetics), validating the effectiveness of medium dimensions in short-term interest modeling. However, when the dimension continues to increase to 320 or 384, the performance improvement is not significant, and some datasets (such as Multi-category) even show a decline. This may be because excessively large dimensions introduce overfitting and training instability, while also incurring additional computational costs and reducing overall efficiency. Therefore, a GRU hidden dimension of 256 is chosen.

[0134] The impact of the number of attention heads on model performance in the Transformer was further investigated. The value range was set to {1, 2, 4, 6, 8, 10}. Experimental results are as follows: Figure 8As shown, appropriately increasing the number of attention heads helps the model learn long-term user preferences from multiple semantic perspectives. When the number of attention heads increases from 1 to 4 or 6, the model performance improves significantly, especially on multi-category datasets (such as Multi-category and UBF), demonstrating better generalization ability. However, when the number of attention heads continues to increase to 8 or 10, the performance on some datasets actually decreases slightly. The reason for this may be that too many attention heads lead to an overly dispersed attention distribution, making it difficult for the model to focus on key interest dimensions, and even introducing redundant path noise, affecting the overall recommendation accuracy. Combining the results of three sets of experiments, setting 4 to 6 attention heads can achieve a better trade-off between expressive power and stability. This paper ultimately selects 6 attention heads as the default setting.

[0135] This invention proposes an interest evolution-based conversation recommendation model that integrates multi-task learning. The MTIESR model starts with multi-scale modeling of user interests, constructing a heterogeneous hypergraph that integrates conversation structure, category semantics, and historical behavior to express high-order semantic relationships between short-term user behavior and long-term preferences. Leveraging the structural modeling advantages of hypergraphs, this invention represents items within the same conversation, cross-conversation items of the same category, and frequently interacted items as multi-type hyperedges, capturing high-order adjacency relationships through a hierarchical hypergraph convolutional neural network. Simultaneously, GRU and Transformer are combined to model short-term interests and long-term evolving preferences respectively, and a task-aware gating mechanism is used to achieve dynamic fusion of multi-scale interests. For the optimization objective, category prediction is introduced as an auxiliary task and jointly trained with the main recommendation task, enhancing the model's generalization ability.

[0136] The present invention has been described by way of example. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A method for recommending projects in online tourism, characterized in that, The method is as follows: (1) Generate session hyperedges, category hyperedges and user history hyperedges based on the interaction behavior sequence of each user on the online travel platform, and construct the connection matrix corresponding to the session hyperedges, category hyperedges and user history hyperedges; (2) The adjacency matrix of the session hyperedge Adjacency matrix of category superedges And the adjacency matrix of user history superedges The inputs to the hypergraph convolutional neural network are respectively, and the hypergraph convolutional neural network outputs the adjacency matrix respectively. Adjacency matrix and adjacency matrix Corresponding project embedding Project Embedding Project Embedding and project embedding The middle representation ; (3) Embed the project Input a short-term interest model, and the short-term interest model outputs a short-term interest representation. ; (4) Embed the project Embedded with project After fusion, the input is the long-term interest model, and the output is the long-term interest representation from the short-term interest model. ; (5) Based on short-term interest representation Long-term interest representation and project embedding The middle representation By integrating these elements, the final representation of user interests can be obtained. ; (6) User interest representation Predict the user's next interest item and the preferred item categories for future time periods.

2. The project recommendation method for online tourism as described in claim 1, characterized in that, Session hyperedge set Construction method: Each visit by a user on an online travel platform will be divided into a session, which contains several items. All the corresponding nodes of each item appearing in a session will be connected by an edge to form a session hyperedge, which will be added to the session hyperedge set. middle.

3. The project recommendation method for online tourism as described in claim 1, characterized in that, Category superedge set Construction method: Based on the mapping table between projects and project categories, each project corresponds to a project category. Nodes belonging to the same category in different time periods are connected by an edge to form a category hyperedge, which is then added to the category hyperedge set. middle.

4. The project recommendation method for online tourism as described in claim 1, characterized in that, User history super-edge set Construction method: Based on all users' historical sessions, the click frequency of items across all historical interactions is counted. Items with click frequencies exceeding a set threshold are added to the long-term interest item set. Items in the long-term interest item set are considered long-term interest items. Nodes corresponding to all items in the long-term interest item set are connected by an edge to form a user history hyperedge, which is then added to the user history hyperedge set. .

5. The project recommendation method for online tourism as described in claim 1, characterized in that, Hypergraph convolutional neural networks consist of three hypergraph convolutional layers, which are adjacency matrices of class hyperedges. After inputting into the hypergraph convolutional neural network, the output of the second hypergraph convolutional layer is the item embedding. The middle representation .

6. The project recommendation method for online tourism as described in claim 1, characterized in that, A gated recurrent unit is used as the short-term interest model, and a Transformer model is used as the long-term interest model.

7. The project recommendation method for online tourism as described in claim 1, characterized in that, Adjacency matrix Each column corresponds to a set of session superedges. A session hyperedge in the context of a session is defined such that the element value of the item appearing in the corresponding column is set to 1, otherwise it is set to 0. Adjacency matrix Each column corresponds to a set of superedges for that category. A category superedge in the table is used to set the element value of the corresponding column of the item appearing in the category superedge to 1, otherwise it is set to 0; Items appearing in the user's history superedge in the adjacency matrix Set the value of the element in the corresponding column to 1, otherwise set it to 0.

8. The project recommendation method for online tourism as described in claim 1, characterized in that, The method for predicting a user's next interest item is as follows: Based on user interest representation Predicting the user's next interest item as a project Click preference score Select items with high click preference scores as candidate items. ; Output candidate items using softmax normalization. Click probability of click preference score The item with the highest click rate is selected as the user's next item of interest; Among them, click preference score The specific calculation formula is as follows: ； in, For the project Learnable embedding vectors.

9. The project recommendation method for online tourism as described in claim 1, characterized in that, User's preferred item categories for the future time period Predicted probability The specific details are as follows: ； in, , These are the weight parameters and bias parameters of the linear projection layer.

Images