Intelligent space collaborative service recommendation method and system based on graph attention network

Abstract

The application discloses a wisdom space collaborative service recommendation method based on a graph attention network, which comprises the following steps: assigning labels to an attribute set in a service data set; extracting attribute features of the service through service text, and obtaining user feature representation through user attribute set labels; constructing a service attribute graph network and a user and service historical interaction graph network to evaluate service quality level and learn user service preference, training a model based on the user service preference, and generating a personalized service recommendation list.In addition, the application also provides a wisdom space collaborative service recommendation system based on a graph attention network, which learns user features based on personal information of the user, fuses the personal information of the user, user service preference and service quality, calculates a personalized preference score of the user for a candidate service, and can recommend a service that meets the demand of the user and has high quality to the user according to the score.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of service recommendation, and particularly relates to a smart space collaborative service recommendation method and system based on a graph attention network. BACKGROUND

[0002] With the rapid development of society, more and more social public services are emerging. These social public services include but are not limited to services in the fields of education, medical care, employment, social welfare, culture and entertainment, aiming to meet the increasing basic needs of the public, improve the quality of life, and promote social development. However, the diversity of social public services means that there are a large number of service providers and service options, but the public may not be able to obtain sufficient information to make rational choices, thus it is difficult to obtain services that meet their own needs. In addition, the resources of social public services are often limited, and there are problems of insufficient resources and uneven services in some service fields, while there are problems of underutilization of resources and waste of resources in some service fields. Through the study of service recommendation technology, user demand can be better analyzed and predicted, and scientific basis can be provided for the rational allocation of resources, and the social benefits of services can be improved.

[0003] Different users have different needs, and traditional service provision methods often cannot meet individual needs. Service recommendation technology can provide personalized and accurate social public services for users according to their personal characteristics, preferences, needs and other factors, and improve the matching degree of services and user satisfaction. Existing service recommendation technologies mainly include matrix decomposition-based methods and deep learning-based methods. The matrix decomposition-based method uses known user-service interaction data to construct a user-service rating matrix, where each element represents the user's rating or interaction with the service. Then, through matrix decomposition methods such as singular value decomposition, principal component analysis, and latent factor model, the rating matrix is decomposed into a user matrix and a service matrix. Finally, through the low-dimensional representation of these decomposed matrices, unknown user-service pairs are predicted and recommended. However, matrix decomposition can only capture the linear relationship between users and services, and it is difficult to handle more complex nonlinear relationships, so the recommendation effect is limited in some complex scenarios. The deep learning-based method uses neural network models to learn the complex relationship between users and services. This method can learn more complex and abstract representations through multiple levels of nonlinear transformation, thereby improving the accuracy and generalization ability of the recommendation.

[0004] In the interaction between users and services, a user can interact with multiple services, and at the same time a service can be used or evaluated by multiple users, reflecting the complex many-to-many relationship between users and services. Graph networks in deep learning can capture and express complex nonlinear relationships, which are suitable for handling the many-to-many relationship between users and services, and can improve the accuracy and effectiveness of the recommendation.

[0005] One of the current prior arts is a service recommendation method and system based on knowledge graph and graph neural network (CN117688238A), which constructs a user-application collaborative knowledge graph, updates the feature representation of user and application nodes using graph attention network, predicts the probability of user selecting application, and presents the application service recommendation list that meets the user's preferences; the disadvantage of this invention is that the invention only constructs a collaborative knowledge graph based on the interaction data of users and applications, without considering the quality level of each application service itself, which may recommend some inferior services to users and cannot meet the user's needs well.

[0006] The second of the current prior arts is an intelligent community service recommendation method and system (CN116579904A), which analyzes the adaptation degree, convenience degree and execution efficiency between community services and community groups, and selects the community service with the highest matching degree with the community group; the disadvantage of this invention is that the invention regards all users in the community as a whole, evaluates the matching degree of community services and community groups based on simple data analysis method, lacks analysis on the characteristics and preferences of each user, and cannot recommend community services that meet the personal characteristics and needs of specific users. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the existing method, and a wisdom space collaborative service recommendation method and system based on graph attention network are proposed. The main problems solved by the present application are: 1) how to use weak label learning technology to decouple attribute features from text features of services, and obtain accurate attribute feature representation in the case of lacking part of the label; 2) how to evaluate the service quality based on the attribute features of the service itself, and recommend high-quality services to users; 3) how to learn user preferences based on historical interactive service data of users, and effectively fuse user personal characteristics and service quality, and recommend high-quality services that meet the needs and preferences of users.

[0008] In order to solve the above problems, the present application proposes a wisdom space collaborative service recommendation method based on graph attention network, which comprises:

[0009] Collect service data set, and give attribute set label to attribute set, including service attribute set label and user attribute set label;

[0010] Extract the attribute features of the service, first extract the text features from the text description of the service, and then decouple the service attribute features from the text features with the service attribute set label as the supervision signal;

[0011] Extracting user feature representation, extracting attribute feature representation of the user based on the user attribute set label, and aggregating the user attribute features through global average operation to obtain the initial global feature representation of the user;

[0012] Building a service attribute graph network, building an attribute graph network for each service, learning the coordination relationship between different attributes, and updating the feature representation of the service to evaluate the quality level of the service itself;

[0013] Building a user and service historical interaction graph network, taking the services interacted by the user as nodes, building a service interaction graph network for each user, obtaining the service preference feature representation of the user, and thus realizing the learning of the service preference of the user;

[0014] Calculating the preference score of the user for the candidate service, training the model, and generating a personalized service recommendation list.

[0015] Preferably, the service data set is collected, and attribute set labels are assigned to the attribute sets, including service attribute set labels and user attribute set labels, specifically:

[0016] The service data set includes service data, user features, and historical interaction data between users and services; wherein, for service data S, the i-th data contains service number service text description and service attribute set For user features U, the j-th data contains user number and user attribute set For historical interaction data between users and services H, the k-th data represents a service used by a user, containing user number service number and user's evaluation level of the service For the j-th user, the total number of services interacted by the user is recorded as X j ;

[0017] Under normal circumstances, service data can reflect important features of services, user features reflect personal information of each user objectively existing, and historical interaction data between users and services reflect subjective preferences of different users for services;

[0018] A general attribute set is predefined for services:

[0019] S attr ={s n |n=1,2,...,N}

[0020] Wherein, N is the total number of service attributes, including service theme, service content, service object and other factors, s nthe nth attribute of the service;

[0021] For each attribute s n , define its possible values:

[0022]

[0023] where M n is the total number of possible values of the nth attribute;

[0024] At this time, each service has a set of attribute labels:

[0025]

[0026] where, is the true value of the ith service on the nth attribute, if the ith service has this attribute label, then otherwise,

[0027] For users, a general set of attributes is also predefined:

[0028] U attr = {u t |t = 1, 2,..., T}

[0029] where T is the total number of user attributes, including personal information such as user gender, user age, and user occupation;

[0030] Similarly, according to the collected user information, the set of attribute labels that each user has is obtained:

[0031]

[0032] Preferably, the extraction of the attribute features of the service first extracts the text features from the text description of the service, and then decouples the service attribute features from the text features using the set of service attribute labels as a supervision signal, specifically:

[0033] According to the service text description in the collected service data, a pre-trained Transformer model is used for the ith service to extract the text feature representation Text i of the service from the text description of the service.

[0034] For the predefined N service attributes, N MLPs are used to decouple the features of the corresponding attributes from the text feature representation of the service, specifically:

[0035]

[0036] where, to decouple the obtained i-th service on the n-th attribute feature representation;

[0037] In order to make the decoupled service attribute feature more accurate, a weak label learning technique is used in the decoupling process to partially supervise it with existing attribute labels, including using N MLPs as attribute classifiers Classfier to predict attribute values according to the decoupled service attribute features, specifically:

[0038]

[0039] wherein, each element in is the probability of the i-th service being the corresponding attribute value on the n-th attribute;

[0040] In order to obtain more accurate attribute features, cross-entropy loss is used for constraint to make the predicted service attribute value closer to the true label, specifically:

[0041]

[0042] wherein, is the true attribute label of the i-th service on the n-th attribute;

[0043] In the above process, N predefined service attribute features are obtained, and an "other" attribute is introduced as the N+1 attribute to include all undefined attributes. This attribute feature is also obtained through a special MLP, specifically:

[0044]

[0045] wherein, the "other" attribute contains information other than the N attributes, and the service feature is supplemented, and the cosine similarity loss is used to constrain the "other" attribute feature and the first N attribute features to be orthogonal, specifically:

[0046]

[0047] wherein, cos(·,·) is the cosine similarity function, and the "other" attribute feature is also regarded as a normal attribute feature of the service, so each service has N+1 attribute features to describe each aspect.

[0048] Preferably, the user feature representation is extracted based on the user attribute set label to extract the attribute feature representation of the user, and the user attribute features are aggregated through global average operation to obtain the initial global feature representation of the user, specifically:

[0049] In the user attribute set U attrIn each user, all the corresponding values of the attributes are filled in, the user attribute label is complete, and therefore, when extracting the attribute features of the user, the Word2Vec word embedding model is used to directly map the attribute label of the user to a low-dimensional continuous vector space to obtain the feature representation of the jth user on the tth attribute

[0050] After obtaining the user attribute features, a global average pooling operation is used to aggregate all the attribute features to obtain the initial global feature representation of the jth user which includes the personal information of the jth user such as gender, age, occupation, etc.

[0051] Preferably, the service attribute graph network is constructed, the coordination relationship between different attributes is learned, and the feature representation of the service is updated to evaluate the quality level of the service itself, specifically:

[0052] For the ith service, a graph network is constructed, specifically:

[0053]

[0054] wherein, is a set of node sets representing the nth attribute of the ith service, and the feature of each node is initialized as the attribute feature representation obtained by decoupling is the edge connection set of the graph, and the edge connection between all pairs of nodes is established, that is, the edge connection between two nodes is defined as 1, and each node is a neighbor node of the remaining nodes;

[0055] The information transmission process on each attribute is combined with the attention mechanism, specifically, for the attribute nodes h p and h q of the ith service, the attention scores of the two are calculated, specifically:

[0056]

[0057] wherein, a pq is the importance score of the node h q to the node h p , || is the vector concatenation operation, W S is the weight matrix of linear transformation, exp(·) is the natural exponential function, which is used to convert the result of linear transformation into a non-negative value, is the neighbor node set of the node h p , and are the nodes h p , h p and h zcharacteristic representation of the node h

[0058] Based on the attention score, define the node h p Pass attribute information to the node h p The process is as follows:

[0059]

[0060] where W p and b p are the weights and biases to be learned, is a LeakyReLU activation function, and is an element-wise multiplication, i.e., the Hadamard product of the node features and , so that the node h p can continuously obtain coordination information with the node h q , which is conducive to discovering the potential relationship between the two;

[0061] By summarizing the information passed by all neighboring nodes, the characteristic representation of the node h p can be updated as:

[0062]

[0063] where is the total number of neighboring nodes of the node h p , and is the updated characteristic representation of the node h p , which has its own characteristics and coordination information with neighboring nodes;

[0064] Concatenate the updated characteristic representations of all attribute nodes and input them into the MLP to obtain the quality score of the service:

[0065]

[0066] where W s and b s are the parameters of the MLP, ω(·) is the ReLU activation function, and score i is the quality score of the i-th service;

[0067] Concatenate the updated node features of the graph network and map them to the service space using global average operation to obtain the overall feature representation of the service:

[0068]

[0069] where mean(·) is the global average pooling operation, f i is the overall feature representation of the i-th service, which contains information of multiple attributes and coordination relationships between attributes.

[0070] Preferably, the user service history interaction graph network is constructed, the services interacted by the user in history are taken as nodes, a service interaction graph network is constructed for each user, a service preference feature representation of the user is obtained, so as to realize learning of the service preference of the user, specifically:

[0071] According to the collected historical interaction data of the user on the service, a service history interaction graph network is constructed for each user, for the jth user, the service history interaction graph network is constructed as follows:

[0072]

[0073] wherein, is a node set composed of services interacted by the user in history and a candidate service to be evaluated;

[0074] For the ith service interacted by the jth user in history, the feature is initialized as the updated service overall feature representation f i of the service attribute graph network, is a set of edge connections of the graph, except the candidate service node, all the remaining nodes in the same graph network represent services interacted by the same user in history, representing the historical preference of the user; an edge connection is established between the candidate service node and all the remaining nodes, for the ith service interacted by the jth user in history, the connection weight between the corresponding node and the candidate service node is the historical score of the user on the service The higher the score is, the higher the satisfaction of the user on the service is, which can better reflect the historical service preference of the user, at this time all the nodes are neighbor nodes of the candidate service node;

[0075] Based on different edge connection weights, the information transmission process on each service is combined with the attention mechanism to update the feature of the candidate service node, for the service node h r and the candidate service node h i , the attention scores of the two are calculated as follows:

[0076]

[0077] wherein, a ir is the importance score of the node h r to the node h i , W U is a weight matrix of linear transformation, is a neighbor node set of the node h i , f i , f r and f o are the features of the service nodes h i , hr and h o 's overall feature representation;

[0078] based on the attention score α ir , the process of passing service information from node h r to node h i is as follows:

[0079]

[0080] where W r and b r are the weights and biases to be learned, and is the element-wise multiplication, i.e., the Hadamard product of the node features f i and f r , so that node h i can continuously obtain coordination information with node h r , which is beneficial to modeling the potential association between the two, and by aggregating the information passed by all neighboring nodes, the feature representation of node h i can be updated as:

[0081]

[0082] is the total number of neighboring nodes of node h i , is the updated feature representation of node h i , which has its own features and coordination information with neighboring nodes, i.e., the user's historical service preference information is fused into the candidate service node features to be evaluated.

[0083] Preferably, the user's preference score for the candidate service is calculated, the model is trained, and a personalized service recommendation list is generated, specifically:

[0084] based on the user's initial feature representation and the candidate service features containing user preference information , the two are concatenated and input into the MLP, which outputs the candidate service score fused with the user's personal features and user preferences, and on this basis, the service quality score is fused to calculate the preference score fused with the user's features and the quality of the candidate service:

[0085]

[0086] where W c and b c are the parameters of the MLP, is the preference score of the jth user for the ith service, and λ is in the range [0, 1], representing the proportion of user features and service quality in calculating the personalized service preference score.

[0087] The historical interaction data H of the user with the service is taken as a positive training sample ε + A service is randomly extracted from the service data S to form a user-service interaction negative training sample ε with the user - The number of positive samples and negative samples of each user is the same, and a Bayesian personalized ranking algorithm BPR is used to construct a loss function according to a personalized service preference score:

[0088]

[0089] Wherein, The personalized preference scores of the user to the service in the positive sample and the negative sample are respectively σ, and σ is a sigmoid activation function;

[0090] The total loss function of the model is:

[0091] L = L S + L cos + L bpr

[0092] The neural network model parameters are trained by using a stochastic gradient descent algorithm until the model converges.

[0093] After the model is trained, the candidate services to be evaluated are input into the network in turn to obtain the personalized preference scores of the user to the candidate services, the preference scores of all the candidate services are ranked from high to low, and a corresponding high-quality service personalized recommendation list is generated.

[0094] Correspondingly, the application also provides a wisdom space collaborative service recommendation system based on a graph attention network, comprising:

[0095] A data collection unit collects a service data set, and attribute set labels are given to an attribute set, including a service attribute set label and a user attribute set label.

[0096] A feature extraction unit extracts attribute features of the service, first extracts text features from the text description of the service, and then decouples the service attribute features from the text features by taking the service attribute set label as a supervision signal.

[0097] A feature aggregation unit extracts user feature representation, extracts attribute feature representation of the user based on the user attribute set label, and aggregates the user attribute features by global average operation to obtain the initial global feature representation of the user.

[0098] A network construction unit constructs a service attribute graph network, constructs an attribute graph network for each service, learns the coordination relationship between different attributes, and updates the feature representation of the service to evaluate the quality level of the service itself.

[0099] The preference learning unit constructs a service historical interaction graph network of a user, takes the services interacted by the user as nodes, constructs a service interaction graph network for each user, obtains a service preference feature representation of the user, and thus learns the service preference of the user;

[0100] The service recommendation unit calculates a preference score of the user for a candidate service, trains a model, and generates a personalized service recommendation list.

[0101] The present application has the following beneficial effects:

[0102] The present application extracts text features based on the text description of a service, decouples attribute features from the text features in a partially supervised manner, can make the decoupled attribute features more accurate, and overcomes the hindrance of the partial absence of attribute labels of the service. The present application introduces an "other" attribute to include the attribute features that are not defined, so that the attribute features can express more complete service information and reduce the information loss in the decoupling process. The present application constructs an attribute graph network for each service based on the decoupled service attribute features, learns the attention weights between different attributes through an attention mechanism, and updates the attribute node features with the attention weights. Through the information transmission of the graph attention network, the coordination relationship between different attributes can be learned, which is helpful for evaluating the quality of the service and thus can recommend high-quality services for the user. Since different users like or need different services, the service preference of the user needs to be learned. The present application constructs a service historical interaction graph network for each user based on the historical interaction data of the user and the service, learns the service preference of the user, and further learns the user features based on the personal information of the user. The present application fuses the personal information of the user, the service preference of the user, and the service quality, calculates the personalized preference score of the user for a candidate service, and recommends services that meet the needs of the user and have high quality to the user according to the score. BRIEF DESCRIPTION OF DRAWINGS

[0103] Figure 1 is a flow chart of a smart space collaborative service recommendation method based on a graph attention network according to an embodiment of the present application;

[0104] Figure 2 is a structural diagram of a smart space collaborative service recommendation system based on a graph attention network according to an embodiment of the present application. DETAILED DESCRIPTION

[0105] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0106] Figure 1 is a flow chart of a smart space collaborative service recommendation method based on a graph attention network according to an embodiment of the present application, as shown in the figure, the method comprises: Figure 1

[0107] S1, collect a service data set, and assign attribute set labels to an attribute set, including service attribute set labels and user attribute set labels;

[0108] S2, extract the attribute features of the service, first extract the text features from the text description of the service, and then decouple the service attribute features from the text features using the service attribute set labels as a supervision signal;

[0109] S3, extract the user feature representation, extract the attribute feature representation of the user based on the user attribute set labels, and aggregate the user attribute features through global average operation to obtain the initial global feature representation of the user;

[0110] S4, construct a service attribute graph network, construct an attribute graph network for each service, learn the coordination relationship between different attributes, and update the feature representation of the service to evaluate the quality level of the service itself;

[0111] S5, construct a user and service historical interaction graph network, construct a service interaction graph network for each user using the services that the user has interacted with as nodes, obtain the service preference feature representation of the user, and thus realize the learning of the service preference of the user;

[0112] S6, calculate the preference score of the user for the candidate service, train the model, and generate a personalized service recommendation list.

[0113] Step S1 is as follows:

[0114] S1-1: The collected service data set includes service data, user features, and historical interaction data between users and services; wherein the i-th data of the service data S contains the service number service text description and the service attribute set The j-th data of the user feature U contains the user number and the user attribute set The k-th data of the historical interaction data between users and services H represents a service used by a user, containing the user number service number and the evaluation level of the user for the service For the j-th user, the total number of services that the user has interacted with is denoted as X j ;

[0115] ​In general, service data can reflect the important features of the service, user features reflect the personal information of each user objectively, and historical interaction data of users and services reflect the subjective preferences of different users for services.

[0116] S1-2: Assign a label to the attribute set, and define a general attribute set for the service in advance:

[0117] S attr ={s n |n=1,2,...,N}

[0118] Where N is the total number of service attributes, including service theme, service content, service object and other factors, and s n is the nth attribute of the service.

[0119] For each attribute s n , define its possible values:

[0120]

[0121] Where M n is the total number of possible values of the nth attribute.

[0122] At this time, each service has an attribute set label:

[0123]

[0124] Where, is the true value of the ith service on the nth attribute, and if the ith service has the attribute label, Otherwise,

[0125] For users, a general attribute set is also defined in advance:

[0126] U attr ={u t |t=1,2,...,T}

[0127] Where T is the total number of user attributes, including user gender, user age, user occupation and other personal information.

[0128] Similarly, according to the collected user information, the attribute set label that each user has is obtained as:

[0129]

[0130] Step S2 is as follows:

[0131] S2-1: According to the service text description in the collected service data, a pre-trained Transformer model is used to extract the text feature representation Text of the i-th service from its text description. i ;

[0132] S2-2: Decouple the attribute features from the text features of the service. For the pre-defined N service attributes, N MLPs are used to decouple the features of the corresponding attributes from the text feature representation of the service, specifically:

[0133]

[0134] wherein, is the feature representation of the i-th service on the n-th attribute obtained by decoupling;

[0135] S2-3: Service attribute decoupling supervision. In order to make the decoupled service attribute features more accurate, weak label learning technology is used in the decoupling process to partially supervise them with existing attribute labels. N MLPs are used as attribute classifiers Classfier to predict attribute values according to the decoupled service attribute features, specifically:

[0136]

[0137] wherein, each element in is the probability of the i-th service having the corresponding attribute value on the n-th attribute;

[0138] In order to obtain more accurate attribute features, cross-entropy loss is used for constraint to make the predicted service attribute values closer to the true labels, specifically:

[0139]

[0140] wherein, is the true attribute label of the i-th service on the n-th attribute;

[0141] S2-4: Complete the service attribute features. In the above process, N pre-defined service attribute feature representations are obtained, and an "other" attribute is introduced as the N+1-th attribute to include all undefined attributes. This attribute feature is also mapped through a special MLP, specifically:

[0142]

[0143] wherein, the "other" attribute contains information other than the N attributes, and the service features are supplemented. Cosine similarity loss is used to constrain the "other" attribute features and the first N attribute features to be orthogonal, specifically:

[0144]

[0145] where cos(·,·) is the cosine similarity function, and the "other" attribute feature is also regarded as a normal attribute feature of the service, so each service has N+1 attribute features to describe its aspects.

[0146] Step S3 is specifically as follows:

[0147] S3-1: Extract the attribute features of the user, in the user attribute set U attr , each user fills in the corresponding value of all attributes, the user attribute label is complete, so when extracting the attribute features of the user, the Word2Vec word embedding model is used to directly map the attribute label of the user to a low-dimensional continuous vector space, and the feature representation of the jth user on the tth attribute is obtained.

[0148] S3-2: Obtain the global feature representation of the user, after obtaining the attribute features of the user, the global average pooling operation is used to aggregate all the attribute features, and the initial global feature representation of the jth user is obtained. which contains the gender, age, occupation and other personal information of the jth user.

[0149] Step S4 is specifically as follows:

[0150] S4-1: Construct an attribute graph network for each service, for the ith service, construct a graph network, specifically:

[0151]

[0152] where is a set of nodes representing the nth attribute of the ith service, and the feature of each node is initialized to the attribute feature representation obtained by decoupling is the edge connection set of the graph, and the edge connection between all pairs of nodes is established, that is, the edge connection between two nodes is defined as 1, and each node is a neighbor node of the remaining nodes.

[0153] S4-2: Attribute node information transmission, since different attributes have different contributions to service quality, the information transmission process on each attribute is combined with the attention mechanism, specifically, for the attribute nodes h p and h q of the ith service, the attention scores of the two are calculated, specifically:

[0154]

[0155] where αpq For node h q For node h p Importance score, || for vector concatenation operation, W S Let be the weight matrix for the linear transformation, and exp(·) be the natural exponential function used to convert the result of the linear transformation into a non-negative value. For node h p The set of neighboring nodes, and They are nodes h respectively p h p and h z Feature representation;

[0156] S4-3: Attribute node feature update, based on the attention score, define the feature update from node h. q Pass attribute information to node h p The process is as follows:

[0157]

[0158] Among them, W p and b p These are the weights and biases to be learned, respectively. The function is LeakyReLU activation function, and ⊙ represents element-wise multiplication, i.e., calculating node features. and The Hadamard product makes node h p Able to continuously acquire information about node h q Coordination information is helpful in discovering potential relationships between the two;

[0159] By aggregating the information transmitted by all neighboring nodes, node h can be... p The feature representation is updated as follows:

[0160]

[0161] in, For node h p The total number of neighboring nodes, For node h p The updated feature representation has its own characteristics as well as coordination information with neighboring nodes;

[0162] S4-4: Service quality assessment. The updated feature representations of all attribute nodes are concatenated and input into the MLP to obtain the service quality score.

[0163]

[0164] Among them, W s and bs is the parameter of the MLP, ω(·) is the ReLU activation function, score i is the quality score of the i-th service;

[0165] S4-5: Service feature representation update, concatenating the updated node features of the graph network, and using global average operation to map to the service space to obtain the overall feature representation of the service:

[0166]

[0167] wherein mean(·) is the global average pooling operation, f i is the overall feature representation of the i-th service, which contains information of multiple attributes and the coordination relationship between the attributes.

[0168] Step S5 is as follows:

[0169] S5-1: Construct a service historical interaction graph network for each user, and construct a service historical interaction graph network for each user according to the collected historical interaction data of the user to the service, and construct a service historical interaction graph network for the j-th user:

[0170]

[0171] wherein, is a node set composed of services that have been interacted by the user and candidate services to be evaluated;

[0172] For the i-th service that has been interacted by the j-th user, the feature is initialized as the updated overall feature representation f i of the service attribute graph network, is the edge connection set of the graph, except for the candidate service node, all other nodes in the same graph network represent services that have been interacted by the same user, representing the historical preference of the user; an edge connection is established between the candidate service node and all other nodes, and for the i-th service that has been interacted by the j-th user, the connection weight between the corresponding node and the candidate service node is the historical score of the user to the service The higher the score, the higher the satisfaction of the user to the service, which can better reflect the historical service preference of the user, and at this time all nodes are neighbor nodes of the candidate service node;

[0173] S5-2: Service node information transmission, combining the information transmission process on each service with the attention mechanism based on different edge connection weights to update the feature of the candidate service node, and for the service node h r and the candidate service node h i , the attention score of the two is calculated as follows:

[0174]

[0175] where α ir is the importance score of node h r , W i is the weight matrix of linear transformation, U is the set of neighbor nodes of node h i , f i , f r and f o are the overall feature representations of service nodes h i , h r and h o , respectively.

[0176] S5-3: Candidate service node feature update, based on the attention score α ir , the process of passing service information from node h r to node h i is defined as follows:

[0177]

[0178] where W r and b r are the weights and bias to be learned, respectively, and is the element-wise multiplication, i.e., the Hadamard product of node features f i and f r , so that node h i can continuously obtain coordination information with node h r , which is conducive to modeling the potential association between the two, and by aggregating the information passed by all neighbor nodes, the feature representation of node h i can be updated as:

[0179]

[0180] is the total number of neighbor nodes of node h i , is the updated feature representation of node h i , which has its own features and coordination information with neighbor nodes, i.e., the user's historical service preference information is fused into the candidate service node feature to be evaluated.

[0181] Step S6, specifically as follows:

[0182] S6-1: Calculate the user's preference score for the candidate service, based on the user's initial feature representation and the candidate service feature containing user preference information ​The two are spliced and input into the MLP, and the candidate service score fused with the user personal characteristics and user preferences is output, and on this basis, the quality score of the service is fused, and the preference score of the user characteristics and the quality of the candidate service is calculated:

[0183]

[0184] Wherein, W c and b c are parameters of the MLP, is the preference score of the jth user for the ith service, and the value range of lambda is [0, 1], indicating the proportion of user characteristics and service quality in calculating the personalized service preference score;

[0185] S6-2: Construct positive and negative training sample data, train the model, and use the historical interaction data H of the user and the service as the positive training sample epsilon + , and randomly extract a service from the service data S to form a user service interaction negative training sample epsilon - The number of positive samples and negative samples of each user is the same, and the Bayesian personalized ranking algorithm BPR is used to construct a loss function according to the personalized service preference score:

[0186]

[0187] Wherein, respectively, the personalized preference score of the user for the service in the positive and negative samples, and sigma is a sigmoid activation function;

[0188] The total loss function of the model is:

[0189] L=L S +L cos +L bpr

[0190] The random gradient descent algorithm is used to train the neural network model parameters until the model converges.

[0191] S6-3: Generate a personalized service recommendation list, after training the model, input the candidate service to be evaluated into the network in turn to obtain the personalized preference score of the user, sort the preference scores of all candidate services from high to low, and generate a corresponding high-quality service personalized recommendation list.

[0192] Correspondingly, the application also provides a wisdom space collaborative service recommendation system based on a graph attention network, as shown in Figure 2 , comprising:

[0193] A data set unit 1 is used to collect a service data set, and attribute set labels are assigned to the attribute set, including service attribute set labels and user attribute set labels.

[0194] The feature extraction unit 2 is used for extracting attribute features of the service, and first extracts text features from the text description of the service, and then decouples the attribute features of the service from the text features by taking the attribute set label of the service as a supervision signal;

[0195] The feature aggregation unit 3 is used for extracting user feature representation, and extracts attribute features of the user based on the attribute set label of the user, and aggregates the attribute features of the user by a global average operation to obtain an initial global feature representation of the user;

[0196] The network construction unit 4 is used for constructing a service attribute graph network, constructing an attribute graph network for each service, learning the coordination relationship between different attributes, and updating the feature representation of the service to evaluate the quality level of the service itself;

[0197] The preference learning unit 5 is used for constructing a user-service historical interaction graph network, taking the services interacted by the user in the past as nodes, constructing a service interaction graph network for each user, obtaining the service preference feature representation of the user, and thus learning the service preference of the user;

[0198] The service recommendation unit 6 is used for calculating the preference score of the user to the candidate service, training the model, and generating a personalized service recommendation list.

[0199] Therefore, based on the text description of the service, the text features are extracted, and the attribute features are decoupled from the text features in a partially supervised manner, so that the decoupled attribute features are more accurate, and the obstacle of the partial missing of the attribute label of the service is overcome; the application introduces the “other” attribute to include the undefined attribute features, so that the attribute features can express more complete service information, and the information loss in the decoupling process is reduced; based on the decoupled service attribute features, the application constructs an attribute graph network for each service, learns the attention weight between different attributes through an attention mechanism, and updates the attribute node features with the attention weight. Through the information transmission of the graph attention network, the coordination relationship between different attributes can be learned, which is helpful for evaluating the quality of the service, and thus the high-quality service can be recommended for the user; since different users like or need different services, the service preference of the user needs to be learned, and the application constructs a service historical interaction graph network for each user based on the historical interaction data of the user and the service, and learns the service preference of the user; in addition, the user features are learned based on the personal information of the user, the application fuses the personal information of the user, the service preference of the user and the service quality, calculates the personalized preference score of the user to the candidate service, and recommends the service that meets the needs of the user and has high quality to the user according to the score.

[0200] The method and system for recommending intelligent space collaborative services based on a graph attention network provided by the embodiments of the present application are described in detail above, and the principles and implementation manners of the present application are described by using specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for smart space collaborative service recommendation based on a graph attention network, characterized in that, The method comprises: A service dataset is collected, including service data, user features, and historical interaction data of users with services; wherein for service data S, the i-th data contains service number service text description and a set of service attributes For user features U, the j-th data contains user number and a set of user attributes For historical interaction data H of users with services, the k-th data represents a service used by a user, containing user number service number and user rating of the service For the j-th user, the total number of services that the user has interacted with is denoted as X j ; and attribute set labels are assigned to the attribute sets, including service attribute set labels and user attribute set labels; Extracting the attribute features of the service, first extracting the text features from the text description of the service, and then decoupling the service attribute features from the text features with the service attribute set label as the supervision signal; Extracting the user feature representation, extracting the attribute feature representation of the user based on the user attribute set label, and aggregating the user attribute features through global average operation to obtain the initial global feature representation of the user; Building a service attribute graph network, building an attribute graph network for each service, learning the coordination relationship between different attributes, and updating the feature representation of the service to evaluate the quality level of the service itself, including the following six steps: Step one, for the ith service, build a graph network; Step two, combine the information passing process on each attribute with the attention mechanism: for the i-th service's attribute node h p and h q , compute their attention scores, specifically: Where, α pq For node h q For node h p Importance score, || for vector concatenation operation, W S Let be the weight matrix for the linear transformation, and exp(·) be the natural exponential function used to convert the result of the linear transformation into a non-negative value. For node h p The set of neighboring nodes, and They are nodes h respectively p h q and h z Feature representation; Step three, based on the attention score, define a process to propagate attribute information to node h q Step three, based on the attention score, define a process to propagate attribute information to node h p Step three, based on the attention score, define a process to propagate attribute information to node h where W p and b p are the weights and biases to be learned, is a LeakyReLU activation function, and is an element-wise multiplication, which calculates the Hadamard product of the node features and , enabling the node h p to continuously obtain coordination information with the node h q , which is conducive to discovering the potential correlation between the two nodes. Step four, update the feature representation of node h p by aggregating all the information passed by its neighbors; Step five, splice the updated feature representation of all attribute nodes, input the MLP to obtain the quality score of the service; Step six, splice the node features of the updated graph network, and map to the service space through global average operation to obtain the overall feature representation of the service; Building a user and service historical interaction graph network, taking the services interacted by the user as nodes, building a service interaction graph network for each user, obtaining the service preference feature representation of the user, and thus realizing the learning of the service preference of the user; Calculate the preference score of the user for the candidate service, train the model, and generate a personalized service recommendation list. 2.The wisdom space collaboration service recommendation method based on the graph attention network according to claim 1, wherein, The attribute set label is assigned to the attribute set, including the service attribute set label and the user attribute set label, specifically: Service data can reflect the important features of the service, user features reflect the personal information of each user objectively, and user and service historical interaction data reflect the subjective preferences of different users for services; A general attribute set is predefined for the service: S attr = {s n |n = 1, 2,..., N} Wherein, N is the total number of service attributes, including service theme, service content, service object, s n is the nth attribute of the service. For each attribute s n , define its possible values: where M n is the total number of possible values for the nth attribute; At this time, each service has an attribute set label: wherein, is the true value of the ith service on the nth attribute, if the ith service has the attribute set label, otherwise, For users, a general attribute set is also predefined: U attr = {u t |t = 1, 2,..., T} Wherein, T is the total number of user attributes, including user gender, user age, and user occupation; Similarly, according to the collected user information, the attribute set label possessed by each user is obtained as: 3.The wisdom space collaboration service recommendation method based on the graph attention network according to claim 2, wherein, The attribute features of the service are extracted, first the text features are extracted from the text description of the service, and then the service attribute features are decoupled from the text features with the service attribute set label as the supervision signal, specifically: According to the service text description in the collected service data, a pre-trained Transformer model is used for the i-th service to extract the text feature representation Text of the service for its text description i ; For the predefined N service attributes, N MLPs are used to decouple the features of the corresponding attributes from the text feature representation of the service, specifically: wherein, is the decoupled representation of the ith service on the nth attribute. In order to make the decoupled service attribute features more accurate, weak label learning technology is used in the decoupling process to partially supervise it with existing attribute labels, including using N MLPs as attribute classifiers Classfier to predict attribute values according to the decoupled service attribute features, specifically: wherein, each element in the vector is the probability that the i-th predicted service has the corresponding attribute value on the n-th attribute. In order to obtain more accurate attribute features, cross-entropy loss is used for constraint to make the predicted service attribute values closer to the true labels, specifically: wherein, is the true attribute label that the ith service has on the nth attribute. In the above process, N predefined service attribute feature representations are obtained, and a "other" attribute is introduced as the N+1 attribute to include all undefined attributes, which is also obtained through a special MLP mapping, specifically: Wherein, the "other" attribute contains information other than the N attributes, thereby supplementing the service features, and the cosine similarity loss is used to constrain the orthogonality between the "other" attribute feature and the first N attribute features, specifically: Wherein, cos(·,·) is the cosine similarity function, and the "other" attribute feature is also regarded as a normal attribute feature of the service, so each service has N+1 attribute features to describe its aspects. 4.The wisdom space collaborative service recommendation method based on the graph attention network according to claim 2, wherein, The user attribute feature representation is extracted based on the user attribute set label, and the user attribute features are aggregated through global average operation to obtain the initial global feature representation of the user, specifically: In the user attribute set U attr In the user attribute set U After obtaining the user attribute features, a global average pooling operation is used to aggregate all the attribute features to obtain an initial global feature representation of the jth user which includes the gender, age, and occupation of the jth user. 5.The wisdom space collaboration service recommendation method based on the graph attention network according to claim 3, wherein, The graph network is constructed for the i-th service, specifically: wherein, is a set of nodes representing the i-th service n-th attribute, and the feature of each node is initialized as the attribute feature representation obtained by the decoupling is an edge connection set of the graph, and an edge connection is established between all pairs of nodes, and the edge connection between each pair of nodes is defined as 1, and each node is a neighbor node of the remaining nodes; The information delivered by all the neighbor nodes is summarized to update the feature representation of the node h p , in particular: wherein, the total number of neighbor nodes of node h p , the updated feature representation of node h p , with its own features and coordination information with neighbor nodes; The updated feature representations of all attribute nodes are spliced and input into the MLP to obtain the quality score of the service, specifically: where W s and b s are parameters of the MLP, ω(·) is the ReLU activation function, score i is the quality score of the i-th service; The node features of the updated graph network are spliced and mapped to the service space through global average operation to obtain the overall feature representation of the service, specifically: where mean(·) is a global average pooling operation, f i is the overall feature representation of the ith service, which contains information of multiple attributes and the coordination relationship among attributes. 6.The wisdom space collaboration service recommendation method based on the graph attention network according to claim 4 or 5, characterized in that, The user and service historical interaction graph network is constructed, and the service interaction graph network is constructed for each user based on the historical interaction data of the user, and the service preference feature representation of the user is obtained, so as to realize the learning of the service preference of the user, specifically: According to the collected historical interaction data of the user on the service, the service historical interaction graph network is constructed for each user, and for the j-th user, the service historical interaction graph network is constructed: wherein, is a set of nodes consisting of services interacted by the user history and candidate services to be evaluated; For the ith service that the jth user has historically interacted with, the feature is initialized as the updated overall service feature representation f of the service attribute graph network i , is the edge connection set of the graph, and all nodes in the same graph network represent services that the same user has historically interacted with, representing the historical preferences of the user; an edge connection is established between the candidate service node and all the remaining nodes, and the connection weight between the corresponding node and the candidate service node is the historical score of the user for the ith service that the jth user has historically interacted with The higher the score, the higher the degree of satisfaction of the user for the service, and the more it reflects the historical service preferences of the user, and at this time all nodes are neighbor nodes of the candidate service node; Based on different edge connection weights, the information transmission process on each service is combined with the attention mechanism, the features of the candidate service nodes are updated, and the service nodes h r and the candidate service nodes h i interacted by the jth user are calculated as follows: Where, α ir For node h r For node h i Importance score, W U The weight matrix is ​​the linear transformation matrix. For node h i The set of neighboring nodes, f i f r and f o They are service nodes h respectively i h r and h o Overall characteristic representation; Based on the attention score a ir , define from node h r Pass the service information to node h i The process is as follows: where W r and b r are the weights and biases to be learned, is the element-wise multiplication, and the Hadamard product of the node features f i and f r enables node h i to continuously obtain coordination information with node h r , which is conducive to modeling the potential association between the two nodes, and the feature representation of node h i is updated by aggregating the information transmitted by all neighboring nodes as follows: the total number of neighbor nodes of node h i the total number of neighbor nodes of node h the updated feature representation of node h i The updated feature representation of node h has its own features and coordination information with neighbor nodes, and the historical service preference information of the user is fused into the candidate service node features to be evaluated. 7.The wisdom space collaboration service recommendation method based on the graph attention network according to claim 6, wherein, The preference score of the user on the candidate service is calculated, the model is trained, and a personalized service recommendation list is generated, specifically: Based on the initial feature representation of the user and candidate service features containing user preference information After splicing, input the MLP, output the candidate service score fused with the user's personal features and user preferences, and on this basis, fuse the service quality score to calculate the preference score of the user features and the candidate service quality: where W c and b c are parameters of the MLP, is the preference score of the jth user for the ith service, and λ is in the range [0, 1] and represents the proportion of user characteristics and service quality in the calculation of the personalized service preference score. The historical interaction data H of the user with the service is taken as a positive training sample ε + A service is randomly extracted from the service data S to form a user-service interaction negative training sample ε with the user - The number of positive samples and negative samples of each user is the same, and a loss function is constructed according to the personalized service preference score by using a Bayesian personalized ranking algorithm BPR. wherein, respectively the personalized preference score of the user to the service in the positive and negative samples, and σ is the sigmoid activation function. The total loss function of the model is: L = L S + L cos + L bpr The random gradient descent algorithm is used to train the neural network model parameters until the model converges; After training the model, the candidate services to be evaluated are input into the network in turn to obtain the personalized preference score of the user, and the preference scores of all candidate services are sorted from high to low to generate a corresponding high-quality personalized service recommendation list.

8. The intelligent space collaborative service recommendation system based on a graph attention network, characterized in that, The system comprises: A data collection unit collects a service dataset, including service data, user features, and historical interaction data of users with services; wherein for service data S, the i-th data contains a service number a service text description and a service attribute set For user features U, the j-th data contains a user number and a user attribute set For historical interaction data H of users with services, the k-th data represents a service used by a user, containing a user number a service number and a user rating of the service For the j-th user, the total number of services that he / she has interacted with is denoted as X j ; and attribute set labels are assigned to the attribute sets, including service attribute set labels and user attribute set labels; A feature extraction unit extracts the attribute features of the service, first extracts the text features from the text description of the service, and then decouples the service attribute features from the text features using the service attribute set label as a supervision signal; A feature aggregation unit extracts the user feature representation, extracts the attribute feature representation of the user based on the user attribute set label, and aggregates the user attribute features through global average operation to obtain the initial global feature representation of the user; A network construction unit constructs a service attribute graph network, constructs an attribute graph network for each service, learns the coordination relationship between different attributes, and updates the feature representation of the service to evaluate the quality level of the service itself, including the following six steps: Step one, for the i-th service, construct a graph network; Step two, combine the information passing process on each attribute with the attention mechanism: for the i-th service's attribute node h p and h q , compute their attention scores, specifically: Where, α pq For node h q For node h p Importance score, || for vector concatenation operation, W S Let be the weight matrix for the linear transformation, and exp(·) be the natural exponential function used to convert the result of the linear transformation into a non-negative value. For node h p The set of neighboring nodes, and They are nodes h respectively p h q and h z Feature representation; Step three, based on the attention scores, define a process to propagate attribute information to node h p Step three, based on the attention scores, define a process to propagate attribute information to node h p Step three, based on the attention scores, define a process to propagate attribute information to node h where W p and b p are the weights and biases to be learned, is a LeakyReLU activation function, and is an element-wise multiplication, which calculates the Hadamard product of the node features and , enabling the node h p to continuously obtain coordination information with the node h q and facilitating the discovery of potential relationships between the two nodes. Step four, update the feature representation of node h p by aggregating all the information passed by its neighbors; Step five, splice the updated feature representations of all attribute nodes and input them into the MLP to obtain the quality score of the service; Step six, splice the updated node features of the graph network and map them to the service space through global average operation to obtain the overall feature representation of the service. Step six, the updated node features of the graph network are spliced together and mapped to the service space by global average operation to obtain the overall feature representation of the service; The preference learning unit constructs a user and service historical interaction graph network, constructs a service interaction graph network for each user using the services that the user has historically interacted with, obtains a service preference feature representation of the user, and thus learns the service preferences of the user; The service recommendation unit calculates the preference score of the user for the candidate service, trains the model, and generates a personalized service recommendation list. 9.The wisdom space collaboration service recommendation system based on the graph attention network according to claim 8, wherein, The attribute set is assigned an attribute set label, including a service attribute set label and a user attribute set label, specifically: Service data can reflect important features of services, user features reflect personal information of each user, and historical interaction data of users and services reflect subjective preferences of different users for services; A general attribute set is predefined for services: S attr = {s n |n = 1, 2,..., N} Wherein, N is the total number of service attributes, including service theme, service content, service object, s n is the nth attribute of the service. For each attribute s n , define its possible values: where M n is the total number of possible values for the nth attribute; At this time, each service has an attribute set label: wherein, is the true value of the ith service on the nth attribute, if the ith service has the attribute set label, otherwise, For users, a general attribute set is also predefined: U attr = {u t |t = 1, 2,..., T} T is the total number of user attributes, including user gender, user age, and user occupation. Similarly, based on the collected user information, each user has an attribute set label: 10.The wisdom space collaboration service recommendation system based on the graph attention network according to claim 9, wherein, The attribute features of the service are extracted, first the text features are extracted from the text description of the service, and then the service attribute features are decoupled from the text features using the service attribute set label as a supervision signal, specifically: According to the service text description in the collected service data, a pre-trained Transformer model is used for the i-th service to extract the text feature representation Text of the service for its text description i ; For the predefined N service attributes, N MLPs are used to decouple the corresponding attribute features from the text feature representation of the service, specifically: wherein, is a decoupled representation of the i-th service's characteristic on the n-th attribute. In order to make the decoupled service attribute features more accurate, a weak label learning technique is used in the decoupling process to partially supervise them with existing attribute labels, including using N MLPs as attribute classifiers Classfier to predict attribute values based on the decoupled service attribute features, specifically: wherein, each element in the vector is the probability that the i-th predicted service has the corresponding attribute value on the n-th attribute. In order to obtain more accurate attribute features, a cross-entropy loss is used to constrain the predicted service attribute values to be closer to the true labels, specifically: wherein, is the true attribute label that the ith service has on the nth attribute. In the above process, N predefined service attribute feature representations are obtained, and a "other" attribute is introduced as the N+1 attribute to include all undefined attributes. This attribute feature is also mapped by a special MLP, specifically: The "other" attribute contains information other than the N attributes, and the service features are supplemented accordingly. A cosine similarity loss is used to constrain the "other" attribute feature to be orthogonal to the first N attribute features, specifically: Where cos(·,·) is the cosine similarity function. The "other" attribute feature is also considered as a normal attribute feature of the service, so each service has N+1 attribute features to describe various aspects. 11.The wisdom space collaboration service recommendation system based on the graph attention network according to claim 9, wherein, The user feature representation is extracted based on the user attribute set label, and the user attribute features are aggregated by global average operation to obtain the initial global feature representation of the user, specifically: In the user attribute set U attr In the user attribute set U After obtaining the user attribute features, a global average pooling operation is used to aggregate all the attribute features to obtain an initial global feature representation of the jth user which includes the gender, age, and occupation of the jth user. 12.The intelligent space collaborative service recommendation system based on graph attention network according to claim 10, wherein, For the i-th service, a graph network is constructed, specifically: wherein, is a set of nodes representing the i-th service n-th attribute, and the feature of each node is initialized as the attribute feature representation obtained by the decoupling is a set of edge connections of the graph, and an edge connection is established between all pairs of nodes, and the edge connection between each pair of nodes is defined as 1, and each node is a neighbor node of the remaining nodes; The information delivered by all the neighbor nodes is summarized to update the feature representation of the node h p , in particular: wherein, the total number of neighbor nodes of node h p the updated feature representation of node h p with its own features and coordination information with neighbor nodes;​ The updated attribute node feature representations are spliced, input into an MLP, and a quality score of the service is obtained, specifically as follows: where W s and b s are parameters of the MLP, ω(·) is the ReLU activation function, score i is the quality score of the i-th service; The updated node features of the graph network are spliced, and global average operation is adopted to map to a service space, and overall feature representation of the service is obtained, specifically as follows: where mean(·) is a global average pooling operation, f i is the overall feature representation of the ith service, which contains information of multiple attributes and the coordination relationship among attributes. 13.The wisdom space collaboration service recommendation system based on the graph attention network according to claim 11 or 12, wherein, The user-service historical interaction graph network is constructed, a service interaction graph network is constructed for each user, service preference feature representation of the user is obtained, learning of service preference of the user is realized, and specifically as follows: According to the collected historical interaction data of the user on the service, a service historical interaction graph network is constructed for each user, and for the jth user, the service historical interaction graph network is constructed as follows: wherein, is a set of nodes consisting of services interacted by the user history and candidate services to be evaluated; For the ith service that the jth user has historically interacted with, the feature is initialized as the updated overall service feature representation f of the service attribute graph network i , is the edge connection set of the graph, and all nodes in the same graph network represent services that the same user has historically interacted with, representing the historical preferences of the user; an edge connection is established between the candidate service node and all the remaining nodes, and the connection weight between the corresponding node and the candidate service node is the historical score of the user for the ith service that the jth user has historically interacted with The higher the score, the higher the degree of satisfaction of the user for the service, and the more it reflects the historical service preferences of the user, and at this time all nodes are neighbor nodes of the candidate service node; Based on different edge connection weights, the information transmission process on each service is combined with the attention mechanism, the features of the candidate service nodes are updated, and the service nodes h r and the candidate service nodes h i interacted by the jth user are calculated as follows: Where, α ir For node h r For node h i Importance score, W U The weight matrix is ​​the linear transformation matrix. For node h i The set of neighboring nodes, f i f r and f o They are service nodes h respectively i h r and h o Overall characteristic representation; Based on the attention score a ir , define from node h r Pass the service information to node h i The process is as follows: where W r and b r are the weights and biases to be learned, is the element-wise multiplication, and the Hadamard product of the node features f i and f r enables node h i to continuously obtain coordination information with node h r , which is conducive to modeling the potential association between the two nodes, and the feature representation of node h i is updated by aggregating the information transmitted by all neighboring nodes as follows: the total number of neighbor nodes of node h i the updated feature representation of node h i with its own features and coordination information with neighbor nodes, the historical service preference information of the user is fused into the feature of the candidate service node to be evaluated.​ 14.The wisdom space collaboration service recommendation system based on the graph attention network according to claim 13, wherein, The preference score of the user on the candidate service is calculated, the model is trained, and a personalized service recommendation list is generated, specifically as follows: Based on the initial feature representation of the user and candidate service features containing user preference information After splicing, input the MLP, output the candidate service score fused with the user's personal features and user preferences, and on this basis, fuse the service quality score to calculate the preference score of the user features and the candidate service quality: where W c and b c are parameters of the MLP, is the preference score of the jth user for the ith service, and λ is in the range [0, 1] and represents the proportion of user characteristics and service quality in the calculation of the personalized service preference score. The historical interaction data H of the user with the service is taken as a positive training sample ε + A service is randomly extracted from the service data S to form a user-service interaction negative training sample ε with the user - The number of positive samples and negative samples of each user is the same, and a loss function is constructed according to the personalized service preference score by using a Bayesian personalized ranking algorithm BPR. wherein, respectively the personalized preference score of the user to the service in the positive and negative samples, and σ is the sigmoid activation function. The total loss function of the model is as follows: L = L S + L cos + L bpr A random gradient descent algorithm is adopted to train the neural network model parameters until the model converges. After the model is trained, the candidate services to be evaluated are sequentially input into the network to obtain the personalized preference score of the user, the preference scores of all candidate services are sorted from high to low, and a corresponding high-quality personalized service recommendation list is generated.

Images