A service recommendation method based on HGCN and multi-granularity hawkes process

Abstract

The application relates to the field of data mining and a recommendation system, in particular to a service recommendation method based on HGCN and a multi-granularity Hawkes process, which comprises the following steps: inputting historical service interaction data of a user; reading the historical service interaction data of the user, sequentially arranging the historical service interaction data of each user according to the interaction time, and obtaining a service interaction sequence of each user; constructing a user service interaction hypergraph, obtaining a node feature matrix of user interaction through hypergraph convolution, and obtaining a preference of user interaction through user preference aggregation; learning periodic behavior patterns of the user under different time granularities through a multi-granularity Hawkes model and a perception service interval attention mechanism; and model training and service recommendation. The application effectively solves the deficiencies of traditional recommendation methods in processing complex user behavior patterns and periodic changes under multiple time scales, and can more accurately understand the demand and interest changes of the user in different time periods.

Description

Technical Field

[0001] This invention relates to the field of data mining and recommendation systems, specifically a service recommendation method based on HGCN and multi-granularity Hawkes processes. Background Technology

[0002] With the widespread adoption of the internet, user behavior data in big data environments has become massive and diverse. The high dimensionality and complexity of this data have led to increasingly serious information overload problems. Service recommendation has emerged as an important means to address this issue. It is mainly used to accurately identify users' potential preferences by deeply mining user behavior data, thereby providing more personalized and accurate services and improving user experience.

[0003] Currently, existing service recommendation models typically employ sequence models to model user interaction sequences or graph neural network models to capture more complex transformational relationships between services. However, in real-world applications, user behavior patterns are complex and diverse, making it difficult to accurately reflect continuously changing user preferences over time. Recent research has emphasized the importance of temporal information in user interaction sequences, using time intervals from historical access records to capture user preferences for services that change over time. However, existing methods often model user preferences at a single temporal granularity, neglecting periodic preferences across different temporal granularities. Furthermore, most methods only focus on the shallow order of service transitions, while deeper logical dependencies may exist between different services, which are not effectively modeled. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention analyzes user preferences for different services at different times by analyzing user interaction records, and proposes a service recommendation method based on Hypergraph Convolutional Network (HGCN) and multi-granularity Hawkes process, combined with the characteristics of the services themselves.

[0005] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0006] A service recommendation method based on HGCN and multi-granularity Hawkes processes first constructs a hypergraph using user service interaction records and captures users' structural preferences through hypergraph convolution. It then obtains user interaction preferences through user preference aggregation. Next, it uses a multi-granularity Hawkes model combined with a service interval-aware attention mechanism to learn users' periodic behavioral patterns at different time granularities. Finally, it infers the probability of service usage based on the multi-granularity Hawkes intensity function values, providing users with more accurate service recommendations. The method includes the following steps:

[0007] Step 1: Input the user's historical service interaction data;

[0008] Step 2: Read the user's historical service interaction data, sort the historical service interaction data of each user in chronological order according to the interaction time, and obtain the service interaction sequence of each user. Each row corresponds to the service interaction sequence of one user.

[0009] Step 3: Construct a user service interaction hypergraph, obtain the node feature matrix of user interaction through hypergraph convolution, and obtain the user interaction preferences through user preference aggregation;

[0010] Step 4: By introducing an attention mechanism for perceiving service time intervals through a multi-granularity Hawkes process model, and combining it with the user interaction preference vector obtained in Step 3, we learn the periodic behavior patterns of users at different time granularities.

[0011] Step 5: Based on the periodic behavior intensity function modeled by the multi-granularity Hawkes process model constructed in Step 4, design a time-aware service recommendation model. Use the intensity function to estimate the probability of a user triggering various candidate services at the current moment, rank the candidate services and generate recommendation results. During training, optimize the model parameters by minimizing the recommendation ranking loss function.

[0012] Furthermore, in step 1, the historical service interaction data includes the user ID, service ID, and the time of the interaction.

[0013] Further, in step 2, within each service interaction sequence, the service IDs that user u has interacted with are sorted in chronological order of interaction time, separated by commas, to obtain the user's service interaction sequence: S u ={(Service 1 ID, Service 1 interaction timestamp), (Service 2 ID, Service 2 interaction timestamp), ..., (Service n ID, Service n interaction timestamp)} ={(i u,1 ,t1),(i u,2 ,t2),...,(i u,n ,t n )}.

[0014] Furthermore, in step 3, the hypergraph convolution obtains the service feature matrix of user interaction through hyperedge clipping operation, including the following steps:

[0015] First, construct an interaction hypergraph for the service sequences that the user has interacted with. User u as hyperedge e u ∈V, the services i that user u has interacted with u,k As a node of a hyperedge, through the incidence matrix H∈R |M|×|N| This represents the constructed hypergraph. Where M and N represent the number of users and the number of services, respectively;

[0016] Based on the constructed interaction hypergraph Augmented Graph Convolutional (GCN) selects the node feature matrix relevant to user u through a hyperedge clipping operation; where the node feature matrix... The update method is defined as follows:

[0017]

[0018] Among them, X (l) Let X represent the feature matrix of the nodes in the l-th layer. (l+1) Let A represent the feature matrix of the nodes in the (l+1)th layer, and let A represent the adjacency matrix. The normalized adjacency matrix, W represents the linear transition matrix, D represents the degree matrix, and α1 represents the activation function.

[0019] In a service recommendation method based on HGCN and multi-granularity Hawkes process of the present invention, step 3 obtains user interaction preferences through user preference aggregation, including the following steps:

[0020] First, after performing hypergraph convolution operations with L stacked layers, the output of the Lth layer is... Average pooling is used to obtain the structural preferences of user u;

[0021] Then, structural preferences and user-personalized preferences v u,p The summation yields the final user preference representation v. u Defined as:

[0022]

[0023] Among them, v u Indicates overall user preferences, v u,p X represents the user's personalized preferences. u (l) This represents the node feature matrix obtained by user u after passing through l layers of hypergraph convolution. This indicates the average pooling operation.

[0024] In a service recommendation method based on HGCN and multi-granularity Hawkes process of the present invention, step 4 includes the following sub-steps:

[0025] Step 4-1: Embedding Context Information

[0026] An embedding layer maps discrete user IDs and service IDs to a high-dimensional space to learn potential association features between users and services. Given a user set U = {u1, u2, ..., u...} M} and the service set I = {i1, i2, ..., i N For sets U and I, embedding yields a personalized representation of user preferences:

[0027]

[0028] Among them, v i The feature representation of service i, v u,p This represents the personalized preferences of user u.

[0029] The user interaction timestamp information is encoded using a method similar to location encoding, defined as:

[0030]

[0031] p u (t)={p u (t,1),p u (t,2),...,p u (t,d)}

[0032] Where t represents the timestamp of the user's historical service interactions, j represents the index of the vector dimension, d is the total dimension of the location encoding vector, T represents the time normalization constant, and p u (t) represents the time position code of timestamp t;

[0033] The final service embedding representation is formed by adding the service embeddings that the user interacted with at time t to the time location code at time t, as defined below:

[0034] v i (t)=v i +α2p u (t)

[0035] Among them, v i Let αt represent the service embeddings that have been interacted with at time t, and α2 represent the scaling factor.

[0036] Step 4-2: Time Hawkes Process

[0037] Given a user u and its service interaction sequence S u , in t h The Hawkes process intensity function of the target service i at time moment i is given by the basic intensity μ. u,i And the self-motivation part The result of addition is calculated as follows:

[0038]

[0039] Where Δt represents the interval between the historical service time and the current time, Used to model the impact of past services on the present, the baseline strength μ of service occurrence. u,i The higher the similarity between user u and target service i, the greater the underlying strength of the service occurrence. Cosine similarity is used to calculate the embedding v of user u. u Embedded with target service i i The similarity between them, used to represent the basic strength of the occurrence of target service i, is defined as:

[0040]

[0041] Self-motivation part Indicates the impact of historical services on current services; α s′,i To represent the degree of influence of historical service s' on target service i, cosine similarity is used to calculate the embedding v of historical service. s′ Embedded with the target service v i The higher the similarity, the greater the influence, defined as:

[0042]

[0043] Attenuation factor β u The exponential function represents the rate at which the influence of historical services on current services decays over time. This indicates that the impact of user u's historical services on the target service gradually decreases as time Δt increases;

[0044] Step 4-3: Multi-particle Hawkes intensity function polymerization

[0045] Considering the different behavioral patterns of user interaction at different time granularities, time is divided into four granularities: month (m), week (w), day (d), and hour (h). For each time granularity t∈T={m,w,d,h}, its corresponding Hawkes intensity function λ is calculated. m , λ w , λ d , λ h Based on hyperparameter {θ m θ w θ d θ h The contribution of intensity functions at different time granularities is adaptively adjusted. The definition of multi-granularity intensity functions is as follows:

[0046]

[0047] Where, θ j λ represents the contribution value of the intensity function at different time granularities. j This represents the intensity function at different time granularities.

[0048] In the service recommendation method based on HGCN and multi-granularity Hawkes process of the present invention, step 5 optimizes the service recommendation performance by maximizing the discriminative power between positive and negative samples. The formula for the loss function is defined as follows:

[0049]

[0050] Among them, i + This represents a positive sample, i.e., the true label. - Indicates other negative samples; and These represent the multi-granularity Hawkes intensity function values ​​for positive and negative samples, respectively.

[0051] The final multi-granularity intensity function λ is obtained by aggregating the intensity function values ​​at different time granularities. u (t) represents the recommendation score, and the top-K with the highest probability are selected as the final recommendation candidate list.

[0052] This invention has the following characteristics and beneficial effects:

[0053] This invention establishes a hypergraph structure between users and multiple services, and utilizes hypergraph convolution to capture the complex structured relationships between users and services, thereby effectively modeling users' structured preferences. Furthermore, users exhibit different behavioral patterns at different time granularities; therefore, this invention models the periodicity of user behavior at different time scales through multi-granularity Hawkes processes, enabling a more accurate understanding of changes in users' needs and interests over different time periods.

[0054] This invention effectively addresses the shortcomings of traditional recommendation methods in handling complex user behavior patterns and periodic changes across multiple time scales. Experiments show that, on multiple real-world datasets, the service recommendation method based on hypergraph convolutional networks and multi-granularity Hawkes processes in this invention exhibits superior recommendation capabilities compared to traditional methods, outperforming traditional methods in multiple metrics (Recall, NDCG). Attached Figure Description

[0055] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0056] Figure 1 This is a flowchart illustrating a service recommendation method based on HGCN and multi-granularity Hawkes processes according to an embodiment of the present invention.

[0057] Figure 2 This is a comparison of the experimental results of the method of the present invention and the control method. Detailed Implementation

[0058] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0059] To address the shortcomings of traditional recommendation methods in handling complex user behavior patterns and periodic changes across multiple time scales, this invention proposes a service recommendation method based on Hypergraph Convolutional Networks (HGCN) and multi-granularity Hawkes processes. This method analyzes user preferences for different services at different time points by leveraging user interaction records. Specifically, by combining hypergraph convolution and multi-granularity Hawkes processes, this invention captures the temporal changes in user interests across multiple dimensions of interaction. Specifically, to effectively capture the multi-layered interest structure of users, we utilize hypergraph convolution to model the interaction relationships between users and multiple services. Within the hypergraph structure, we not only model the relationship between users and individual services but also identify logical relationships between services at a higher-order level. To further enhance the model's ability to model the periodicity of user interest changes, we introduce multi-granularity Hawkes processes. This process can capture the periodicity and self-motivating effect of user behavior at different time scales, i.e., the changes in user preferences for certain services in the short and long term. By considering the temporal dependence of user behavior, multi-granularity Hawkes processes can accurately simulate the evolution of user interests in different time scenarios. Finally, the user embedding obtained by hypergraph convolution is combined with the output of a multi-granularity Hawkes process to generate more personalized and timely recommendation results, thereby effectively improving the prediction accuracy and user satisfaction of the service recommendation system.

[0060] The sequence service recommendation method based on hypergraph convolutional networks and multi-granularity Hawkes processes of the present invention will be described in detail below with reference to the accompanying drawings. Figure 1 As shown, the sequence service recommendation method based on hypergraph convolutional networks and multi-granularity Hawkes processes of the present invention specifically includes the following steps:

[0061] Step 1: Input the user's historical service interaction data, including user ID, service ID, and interaction time.

[0062] Step 2: Read the user's historical service interaction data. Sort the historical service interaction data for each user according to the interaction time to obtain the service interaction sequence for each user. Each line corresponds to a user's service interaction sequence. Specifically, in each service interaction sequence, sort the service IDs that user u has interacted with in chronological order of interaction time, separated by commas, to obtain the user's service interaction sequence: S u={(Service 1 ID, Service 1 interaction timestamp), (Service 2 ID, Service 2 interaction timestamp), ..., (Service n ID, Service n interaction timestamp)} ={(i u,1 ,t1),(i u,2 ,t2),...,(i u,n ,t n )}.

[0063] Step 3: Construct a user service interaction hypergraph, and use hypergraph convolution to capture the complex structured relationships between users and services, generating user interaction preferences. Specifically:

[0064] First, construct an interaction hypergraph for the service sequences that the user has interacted with. User u as hyperedge e u ∈V, the services i that user u has interacted with u,k As a node of a hyperedge, through the incidence matrix H∈R |M|×|N| This represents the constructed hypergraph. Where M and N represent the number of users and the number of services, respectively.

[0065] Based on the constructed interaction hypergraph Augmented Graph Convolutional (GCN) selects the node feature matrix X related to user u through a hyperedge clipping operation. The update method for the node feature matrix X is defined as follows:

[0066]

[0067] In the formula, X (l) Let X represent the feature matrix of the nodes in the l-th layer. (l+1) Let A represent the feature matrix of the nodes in the (l+1)th layer, and let A represent the adjacency matrix. The normalized adjacency matrix, W represents the linear transition matrix, D represents the degree matrix, and α1 represents the activation function.

[0068] Then, after stacking L layers of hypergraph convolution operations, the output of the Lth layer is... Average pooling is used to obtain the structural preferences of user u; these structural preferences are then combined with the user's personalized preferences v. u,p The summation yields the final comprehensive user preference representation v. u Defined as:

[0069]

[0070] Among them, v u Indicates overall user preferences, v u,p X represents the user's personalized preferences. u (l) This represents the node feature matrix obtained by user u after passing through l layers of hypergraph convolution. This indicates the average pooling operation.

[0071] Step 4: Using a multi-granularity Hawkes model, learn the periodic behavioral patterns of users at different time granularities using the attention mechanism of perceived service intervals; specifically:

[0072] Step 4-1: Embedding Context Information

[0073] An embedding layer maps discrete user IDs and service IDs to a high-dimensional space to learn potential association features between users and services. Given a user set U = {u1, u2, ..., u...} M} and the service set I = {i1, i2, ..., i N For U and I, embedding yields a personalized representation of user preferences:

[0074]

[0075] Among them, v i The feature representation of service i, v u,p This represents the personalized preferences of user u.

[0076] To make the model sensitive to timestamp information, a method similar to location encoding is used to encode user interaction timestamp information, defined as: p u (t)={p u (t,1),p u (t,2),...,p u (t,d)}, where t represents the timestamp of the user's historical service interactions, p u (t) represents the time location code of timestamp t. The final service embedding representation is formed by adding the service embeddings that the user interacted with at time t to the time location code of time t. The definition is as follows: v i (t)=v i +α2p u (t), where v i Let α represent the service embeddings that have been interacted with at time t, and α2 represent the scaling factor.

[0077] Step 4-2, Time Hawkes Process

[0078] Given a user u and its service interaction sequence S u , in t h The Hawkes process intensity function of the target service i at time moment i is given by the basic intensity μ. u,i And the self-motivation part The addition is calculated as follows:

[0079]

[0080] Where Δt represents the interval between the historical service time and the current time, Used to model the impact of past services on the present, the baseline strength μ of service occurrence. u,i The higher the similarity between user u and target service i, the greater the underlying strength of the service occurrence. Cosine similarity is used to calculate the embedding v of user u. u Embedded with target service i i The similarity between them, used to represent the basic strength of the occurrence of target service i, is defined as:

[0081]

[0082] Self-motivation part This indicates the impact of historical services on current services. α s′,i To represent the degree of influence of historical service s' on target service i, cosine similarity is used to calculate the embedding v of historical service. s′ Embedded with the target service v i The higher the similarity, the greater the influence, defined as:

[0083]

[0084] Attenuation factor β u This represents the rate at which the impact of historical services on current services decays over time. (Exponential function) This indicates that the impact of user u's historical services on the target service gradually decreases as time Δt increases.

[0085] Step 4-3: Multi-particle size Hawkes intensity function polymerization

[0086] Considering the different behavioral patterns of user interaction at different time granularities, time is divided into four granularities: month (m), week (w), day (d), and hour (h). For each time granularity t∈T={m,w,d,h}, its corresponding Hawkes intensity function λ is calculated. m , λ w , λ d , λ h To reflect the impact of different time granularities, we set the hyperparameter {θ}. m θ w θ d θ h The contribution of intensity functions at different time granularities is adaptively adjusted. The definition of multi-granularity intensity functions is as follows:

[0087] λ i|u (t)=∑ j∈T θ j λj ;

[0088] Where, θ j λ represents the contribution value of the intensity function at different time granularities. j This represents the intensity function at different time granularities.

[0089] Step 5: Model Training and Service Recommendation

[0090] The service recommendation performance is optimized by maximizing the discriminative power between positive and negative samples. The loss function is defined as follows: Where i + This represents a positive sample, i.e., the true label. - Indicates other negative samples. and These represent the multi-granularity Hawkes intensity function values ​​for positive and negative samples, respectively. The final multi-granularity intensity function λ is obtained by aggregating the intensity function values ​​at different time granularities. u (t) represents the recommendation score, and the top-K with the highest probability are selected as the final recommendation candidate list.

[0091] Example 1

[0092] The steps in this implementation are the same as those in steps 1 to 5 above, which describe the sequence service recommendation method based on hypergraph convolutional networks and multi-granularity Hawkes processes. Therefore, they will not be repeated here. The main focus is on demonstrating the specific dataset, specific parameter settings, and implementation results of this embodiment. For ease of description, the methods shown in steps 1 to 5 are referred to as the method of this invention, and the service recommendation model used, namely the model trained in step 5, is MTHP-HC. MTHP-HC mainly includes two modules: a structured hypergraph convolutional module, used to capture users' structured preferences, and the generated representation is added to the user's personalized preference vector to jointly constitute the user's comprehensive interaction preferences; and a multi-granularity Hawkes process module, used to model users' behavioral patterns at different time scales (such as hours, days, weeks, etc.), aggregating multi-granularity time patterns into a unified intensity function through learnable weight parameters, and deriving the recommended service based on the intensity function.

[0093] The raw data used in this embodiment are four widely used real-world datasets: LastFM, Foursquare, MovieLens-100k, and MovieLens-10m. In the LastFM dataset, 921 users engaged in 533,604 interactions across 10,000 services; in the Foursquare dataset, 1,083 users engaged in 227,428 interactions across 38,333 services; in the MovieLens-100k dataset, 943 users engaged in 100,000 interactions across 1,625 services; and in the MovieLens-10m dataset, 26,885 users engaged in 4,548,704 interactions across 5,726 services.

[0094] In this embodiment, some parameters of MTHP-HC are selected as follows: the sequence length n of historical behavior is 5, the embedding dimension dim is 256; the number of stacked layers l of the hypergraph convolution operation is 1, and the α2 scaling factor is 0.001. The iteration termination condition set during the training of the service recommendation model is that the number of iterations reaches 100.

[0095] Furthermore, the experiments in this embodiment also compared the method of the present invention with several traditional service recommendation methods. The traditional service recommendation methods for comparison are: (1) GRU4Rec: a recommendation model based on gated recurrent units (2) SASRec: a recommendation model based on self-attention (3) SRGNN: a recommendation model based on conversational graph neural networks (4) TransRec: a transferable recommendation model based on hybrid modal feedback (5) BERT4Rec: a recommendation model based on BERT (6) FEARec: a recommendation model using frequency-enhanced hybrid attention networks (7) TTT4Rec: a sequence recommendation model based on test-time training. In this embodiment, recall (Recal l@K) and normalized depreciation cumulative gain (NDCG@K) are used as evaluation metrics for the service recommendation model. Recal l@K represents the proportion of the number of actually relevant services to the total number of all relevant services in the first K service recommendation results. In the experiment, K = {10, 15, 20} is set. NDCG can reflect the ranking quality of the recommendation list.

[0096] The final experimental results are as follows Figure 2As shown, MTHP-HC performs best in most test cases across the four datasets, demonstrating the effectiveness of the proposed method. On the Recall@20 and NDCG@20 metrics, the model exhibits significant advantages over traditional service recommendation models. Furthermore, across different metrics and datasets, the best results are achieved by several different service models besides MTHP-HC, while the performance of MTHP-HC remains relatively stable. This comparison clearly demonstrates the effectiveness of the proposed method.

[0097] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A service recommendation method based on HGCN and multi-granularity Hawkes processes, characterized in that, Includes the following steps: Step 1: Input the user's historical service interaction data; Step 2: Read the user's historical service interaction data, sort the historical service interaction data of each user in chronological order according to the interaction time, and obtain the service interaction sequence of each user. Each row corresponds to the service interaction sequence of one user. Step 3: Construct a user service interaction hypergraph, obtain the node feature matrix of user interaction through hypergraph convolution, and obtain the user interaction preferences through user preference aggregation; Hypergraph convolution obtains the service feature matrix of user interaction through hyperedge clipping operations, including the following steps: First, construct an interaction hypergraph for the service sequences that the user has interacted with. User u as a superedge Services that user u has interacted with As a node of a hyperedge, through the incidence matrix This represents the constructed hypergraph. , where M and N represent the number of users and the number of services, respectively; Based on the constructed interaction hypergraph Augmented Graph Convolutional Network (GCN) selects the node feature matrix related to user u through a hyperedge clipping operation; where the node feature matrix... The update method is defined as follows: in, This represents the feature matrix of the nodes in the l-th layer. Let A represent the feature matrix of the nodes in the (l+1)th layer, and let A represent the adjacency matrix. The normalized adjacency matrix, where W represents the linear transition matrix and D represents the degree matrix. Indicates the activation function; Obtaining user interaction preferences through user preference aggregation includes the following steps: First, after performing hypergraph convolution operations with L stacked layers, the output of the Lth layer is... Average pooling is used to obtain the structural preferences of user u; Then, structural preferences and user-personalized preferences... The summation yields the final comprehensive user preference representation. Defined as: in, Indicates overall user preferences. It indicates the user's personalized preferences. This represents the node feature matrix obtained by user u after passing through l layers of hypergraph convolution. This indicates an average pooling operation; Step 4: By introducing an attention mechanism for perceiving service time intervals through a multi-granularity Hawkes process model, and combining it with the user interaction preference vector obtained in Step 3, we learn the periodic behavior patterns of users at different time granularities. Step 5: Based on the periodic behavior intensity function modeled by the multi-granularity Hawkes process model constructed in Step 4, design a time-aware service recommendation model. Use the intensity function to estimate the probability of a user triggering various candidate services at the current moment, rank the candidate services and generate recommendation results. During training, optimize the model parameters by minimizing the recommendation ranking loss function.

2. The service recommendation method based on HGCN and multi-granularity Hawkes process according to claim 1, characterized in that, In step 1, the historical service interaction data includes user data. ,Serve And the time of interaction.

3. The service recommendation method based on HGCN and multi-granularity Hawkes process according to claim 1, characterized in that, In step 2, in each service interaction sequence, the user... Interacted services Sort the user's service interaction sequence chronologically by interaction time, separated by commas: {(Service 1 ID, timestamp of service 1 interaction), (Service 2 ID, timestamp of service 2 interaction), ..., (Service n ID, timestamp of service n interaction)} = .

4. The service recommendation method based on HGCN and multi-granularity Hawkes process according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4-1: Embedding Context Information An embedding layer maps discrete user IDs and service IDs to a high-dimensional space to learn potential association features between users and services, given a user set. and service collection For sets and set Embedding yields a personalized representation of user preferences: in, The feature representation of service i. This represents the personalized preferences of user u. The user interaction timestamp information is encoded using location encoding, defined as follows: in, This represents the timestamps of the user's historical service interactions, j represents the index of the vector dimension, d is the total dimension of the location encoding vector, and T represents the time normalization constant. Represents timestamp Time location encoding; User Embedded services that have been interacted with at any time and time The time and location codes are added together to form the final service embedding representation, as defined below: in, Indicates the first Embedding services that have been interacted with at all times Indicates the scaling factor; Step 4-2: Time Hawkes Process Given user u and its service interaction sequence ,exist The Hawkes process strength function of the target service i at time step is derived from the basic strength. And the self-motivation part The result of addition is calculated as follows: in, This indicates the interval between historical service time and the current time. Used to model the impact of past services on the present, the underlying strength of service occurrence. The higher the similarity between user u and target service i, the greater the underlying strength of the service occurrence. Cosine similarity is used to calculate the embedding of user u. Embedding with target service i The similarity between them, used to represent the basic strength of the occurrence of target service i, is defined as: Self-motivation part This indicates the impact of historical services on current services; To represent the degree of influence of historical service s' on target service i, cosine similarity is used to calculate the embedding of historical service s'. Embedding with target service The higher the similarity, the greater the influence, defined as: Attenuation factor The exponential function represents the rate at which the influence of historical services on current services decays over time. This indicates the impact of user u's historical services on the target service over time. It gradually decreases as it grows; Step 4-3: Multi-particle Hawkes intensity function polymerization Considering the different behavioral patterns of user interaction at different time granularities, time is divided into four granularities: month (m), week (w), day (d), and hour (h); for each time granularity... Calculate its corresponding Hawkes intensity function. Based on hyperparameters The contribution of intensity functions at different time granularities is adaptively adjusted. The definition of multi-granularity intensity functions is as follows: in, This represents the contribution value of the intensity function at different time granularities. This represents the intensity function at different time granularities.

5. The service recommendation method based on HGCN and multi-granularity Hawkes process according to claim 1, characterized in that, In step 5, the service recommendation performance is optimized by maximizing the discriminative power between positive and negative samples. The loss function is defined as follows: in, This represents a positive sample, i.e., a true label. Indicates other negative samples; and These represent the multi-granularity Hawkes intensity function values ​​for positive and negative samples, respectively. The final multi-granularity intensity function is obtained by aggregating the intensity function values ​​at different time granularities. This is the recommendation score, and the top-K with the highest probability are selected as the final recommendation candidate list.

Images