Multi-attribute joint prediction method for predicting service quality of Internet of Things service

Abstract

The invention discloses a multi-attribute joint prediction method for predicting the service quality of an Internet of Things service, and belongs to the technical field of service quality prediction, and the prediction method comprises the steps: obtaining a service quality data set of the Internet of Things service; preprocessing the context features of the user and the service; inputting the processed context features into a first multi-view residual interactive network MRINrt and a second multi-view residual interactive network MRINtp, and respectively executing response time prediction and throughput prediction; respectively calculating loss functions between the response time predicted value and the corresponding true value and between the throughput predicted value and the corresponding true value, obtaining a joint loss function through weighting operation, and training a prediction model by using the joint loss function; and predicting the service quality of the Internet of Things service by using the trained prediction model. According to the method, the potential beneficial incidence relation between different service qualities is considered, and the method has high prediction precision and robustness in data sparse and cold start scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of service quality prediction, and in particular to a multi-attribute joint prediction method for predicting the service quality of an Internet of Things service. Background Art

[0002] IoT applications usually consist of a large number of heterogeneous services provided by various sensor devices, edge servers or cloud servers. For example, the IoT services used by intelligent transportation systems include but are not limited to target detection, traffic flow prediction and travel route planning. 5G technology provides IoT services with higher bandwidth (throughput), lower latency (response time) and more reliable connection (stability), which has promoted a significant increase in the number of IoT services. However, with the increase in their number, the functional homogeneity of IoT services has become more and more serious. For IoT application developers (i.e., users), it is difficult to quickly select high-quality services from a large number of candidate services with similar functions. Therefore, researchers have proposed the concept of service quality to distinguish functionally similar services from a non-functional perspective. Service quality represents the quality information of IoT services in a certain aspect, including response time, throughput and stability. Service quality can help users distinguish functionally similar services and help service providers improve service performance. However, it is not feasible to determine the service quality by calling all candidate services because it is time-consuming and resource-consuming. IoT service quality prediction methods have been widely studied in academia and industry due to their advantages such as low resource overhead and efficient service recommendation.

[0003] Currently popular deep learning-based service quality prediction methods use neural networks to learn the potential characteristics of users and services or the interaction characteristics between users and services, and use the above two characteristics to make predictions. According to the different ways in which the features are input into the model, deep learning-based methods can be roughly divided into two categories: (1) interactive tower method: the characteristics of users and services are jointly input into the model; (2) dual-tower method: the characteristics of users and services are input into the model separately. Although the above two methods can achieve high-precision service quality prediction, both the interactive tower method and the dual-tower method ignore the potentially beneficial correlation between different service qualities. In addition, the delayed interaction strategy of the dual-tower method causes the loss of interaction characteristics between users and services, affecting its prediction accuracy. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a multi-attribute joint prediction method for Internet of Things service quality prediction, which realizes high-precision Internet of Things service quality prediction and effectively meets the objective needs of developers for high-quality Internet of Things service recommendations during the Internet of Things application development process.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A multi-attribute joint prediction method for predicting the quality of service of an Internet of Things service comprises the following steps:

[0007] S1. Obtain IoT service quality dataset;

[0008] S2, preprocessing the context features of users and services;

[0009] S3: Input the processed context features into the first multi-view residual interaction network MRIN rt and the second multi-view residual interaction network MRIN tp In, response time prediction and throughput prediction are performed separately;

[0010] S4, respectively calculating the loss functions between the response time prediction value and the throughput prediction value and the corresponding true value, and obtaining a joint loss function through weighted operation, and using the joint loss function to train the prediction model;

[0011] S5. Use the trained prediction model to predict the quality of IoT services.

[0012] A further improvement of the technical solution of the present invention is that in S1, obtaining a service quality data set of an Internet of Things service specifically includes:

[0013] Use service quality monitoring tools to collect IoT service quality data, except for the required user ID u , IoT service ID s In addition to the service quality value, the context features of useful users and IoT services are selected and added to the dataset; finally, the user context feature F u and IoT service context features F s It is expressed as:

[0014] F u =[ID u ,AS u ,Longitude u ,Latitude u ,Country u ,Province u ,City u ]

[0015] F s =[ID s ,Provider s ,AS s ,Longitude s ,Latitudes ,Country s ,Province s ,City s ]

[0016] Among them, AS u Indicates the network autonomous domain to which the user's IP address belongs. Latitude u and Longitude u Respectively represent the user's longitude and latitude information, Country u , Province u and City u Respectively represent the country, province and city where the user is located; Provider s Indicates the service provider, AS s Indicates the network autonomous domain to which the service IP address belongs. Latitude s and Longitude s Respectively represent the longitude and latitude information of the service, Country s , Province s and City s Respectively represent the country, province, and city where the service is located.

[0017] A further improvement of the technical solution of the present invention is that: in S2, the context features of the user and the service are preprocessed, specifically including the following steps:

[0018] S2.1. OneHot encoding technology is used to convert the context features of users and IoT services into sparse feature vectors.

[0019]

[0020] Among them, F u,i represents the i-th feature of the user, F s,j represents the jth feature of the IoT service; x u,i represents the sparse feature vector of the i-th feature of user, x s,j The sparse feature vector representing the jth feature of the IoT service; O i represents the dimension of the sparse feature vector of the user i, and also represents the number of valid values ​​of the user i feature. j It represents the dimension of the j-th sparse feature vector of the IoT service, and also represents the number of valid values ​​of the j-th feature of the IoT service;

[0021] S2.2, using embedding technology to transform the sparse feature vectors of users and IoT services into embedded vectors in a shared feature space;

[0022] (c u,i ) E =Embedding(x u,i )=W i ·x u,i

[0023] (c s,j ) E =Embedding(x s,j )=W j ·x s,j

[0024] Among them, c u,i represents the embedding vector of the i-th feature of the user, c s,j represents the embedding vector of the jth feature of the IoT service; E represents the dimension of the embedding vector; W represents the weight matrix of the feature embedding layer;

[0025] S2.3. Connect the embedding vectors of users and IoT services Φ to generate user features C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s :

[0026]

[0027] Among them, m represents the number of user context features, and n represents the number of IoT service context features.

[0028] A further improvement of the technical solution of the present invention is that in S3, the following steps are specifically included:

[0029] S3.1. User feature C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s Input to the first multi-view residual interaction network MRIN rt , perform response time prediction;

[0030]

[0031] in, represents the final predicted value of response time;

[0032] The first multi-view residual interaction network MRIN rt By the first user tower First Service Tower and the first interactive tower composition;

[0033] Use the First User Tower First Service Tower To learn the first user's latent characteristics and first service potential characteristics Use the First Interactive Tower To supplement the study of the first user tower First Service Tower Lost first interaction feature between user and service

[0034]

[0035] Among them, the first user tower First Service Tower and the first interactive tower They are all stacked by residual blocks ResBlock;

[0036] The first interaction feature between users and services Input into a residual block ResBlock to get the first response time prediction value

[0037]

[0038] The first user potential characteristics and first service potential characteristics Perform inner product operation to obtain the second response time prediction value

[0039]

[0040] Wherein, T represents the vector transpose operation;

[0041] Prediction of first response time and the second response time prediction value Perform weighted operations to obtain the final predicted value of response time

[0042]

[0043] Among them, α represents the view weight, which determines the first response time prediction value The importance of , and 1-α determines the second response time prediction value The importance of

[0044] S3.2. User feature C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s Input to the second multi-view residual interaction network MRIN tp , perform throughput prediction:

[0045]

[0046] in, represents the final predicted value of throughput;

[0047] The Second Multi-view Residual Interaction Network MRIN tp By the second user tower Second Service Tower and the second interactive tower composition;

[0048] Using the Second User Tower Second Service Tower To learn the second user's latent features and second service potential characteristics Use the second interactive tower To supplement the learning of the second user tower and the second service tower The missing secondary interaction feature between the user and the service

[0049]

[0050] Among them, the second user tower Second Service Tower and the second interactive tower They are also stacked by residual blocks ResBlock;

[0051] The second interaction feature between users and services Input into a residual block ResBlock to get the first throughput prediction value

[0052]

[0053] The second user potential characteristics And the second service potential characteristics Perform inner product operation to obtain the second throughput prediction value

[0054]

[0055] The first throughput prediction value and the second throughput prediction value Perform weighted operations to obtain the final predicted value of throughput

[0056]

[0057] Among them, α represents the view weight, which determines the first throughput prediction value The importance of , and 1-α determines the second throughput prediction value The importance of.

[0058] A further improvement of the technical solution of the present invention is that: in S3.1 and S3.2, the residual block ResBlock structure is based on the first user tower The stacking is described as follows:

[0059]

[0060] Where K represents the number of residual blocks;

[0061] Each residual block is composed of two types of first residual units ResUnit1 and second residual units ResUnit2 stacked together:

[0062]

[0063] Where D represents the total number of residual units in each residual block; the number of the first residual unit ResUnit1 is 1, and the number of the second residual unit ResUnit2 is D-1; the first residual unit ResUnit1 can reduce the dimension of the input vector while learning the interaction between features:

[0064]

[0065] The second residual unit ResUnit2 only learns the interaction between features without changing the dimension of the input vector:

[0066]

[0067] Where d represents the dth residual unit in the current residual block, and its maximum value does not exceed D; X and Y represent the output vectors of the two fully connected layers in the residual block; W and b represent the weight matrix and bias term of the fully connected layer respectively; GELU represents the activation function, and BN represents the batch normalization operation; Represents the output vector of the d-th residual unit and also serves as the input vector of the d+1-th residual unit.

[0068] A further improvement of the technical solution of the present invention is that in S4, the following steps are specifically included:

[0069] S4.1. Calculate the first response time prediction value according to the Huber loss function Second response time prediction value and the corresponding true value y rt The first loss function between The second loss function Among them, the threshold of the Huber loss function is set to 1;

[0070] According to the first loss function Calculate first response time prediction and the true value y rt The loss between:

[0071]

[0072] According to the second loss function Calculate the second response time prediction value and the true value y rt The loss between:

[0073]

[0074] S4.2. Calculate the first throughput prediction value according to the Huber loss function Second throughput prediction value and the corresponding true value y tp The third loss function between The fourth loss function Among them, the threshold of the Huber loss function is set to 1;

[0075] According to the third loss function Calculate the first throughput prediction value and the true value y tp The loss between:

[0076]

[0077] According to the fourth loss function Calculate the second throughput prediction value and the true value y tp The loss between:

[0078]

[0079] S4.3. The first loss function The second loss function The third loss function The fourth loss function Perform weighted operations to obtain the joint loss function L, and use the joint loss function L to train the prediction model:

[0080]

[0081] Among them, β represents the attribute weight, which determines the importance of the response time loss function, and 1-β determines the importance of the throughput loss function; α represents the view weight, which determines the importance of the interactive tower loss function, and 1-α determines the importance of the dual-tower loss function.

[0082] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0083] 1. The multi-attribute joint prediction method for predicting the service quality of Internet of Things services provided by the present invention first pre-processes the context features of users and services; then the processed features are input into two multi-view residual interaction networks to perform response time prediction and throughput prediction respectively; finally, the joint loss function is calculated according to the predicted values, and the prediction model is trained using the joint loss function; the potentially beneficial correlation between different service qualities is taken into account, and this correlation improves the prediction accuracy of the model.

[0084] 2. In the present invention, multi-attribute joint training is performed on the embedded vectors in the shared feature space, thereby improving the generalization of the embedded vectors, thereby improving the generalization ability of the prediction model in the cold start situation, and having higher efficiency than the single attribute prediction method.

[0085] 3. In the present invention, an interactive tower is added on the basis of the dual towers, and the interactive tower is used to supplement and learn the interactive features between users and services lost by the dual towers.

[0086] 4. The residual block structure designed in the present invention can significantly enhance the feature interaction learning ability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0088] Figure 1 It is a flowchart of a multi-attribute joint prediction method for predicting the service quality of an Internet of Things service provided in an embodiment of the present invention;

[0089] Figure 2 It is a schematic diagram of the overall structure of a multi-attribute joint prediction method for predicting the service quality of an Internet of Things service provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0090] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatus.

[0091] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "several" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0092] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0093] like Figure 1 and Figure 2 As shown, a multi-attribute joint prediction method for predicting the quality of service of IoT services includes the following contents:

[0094] S1. Obtain the IoT service quality dataset.

[0095] Use service quality monitoring tools to collect service quality data of IoT services in different regions, except for the required user ID u , IoT service ID s In addition to the service quality value, the context features of useful users and IoT services are selected and added to the dataset. Finally, the user context feature F u and IoT service context features F s It is expressed as:

[0096] F u =[ID u ,AS u ,Longitude u ,Latitude u ,Country u ,Province u ,City u ]

[0097] F s =[ID s ,Provider s ,AS s ,Longitude s ,Latitude s,Country s ,Province s ,City s ]

[0098] Among them, AS u Indicates the network autonomous domain to which the user's IP address belongs. Latitude u and Longitude u Respectively represent the user's longitude and latitude information, Country u , Province u and City u Respectively represent the country, province and city where the user is located. s Indicates the service provider, AS s Indicates the network autonomous domain to which the service IP address belongs. Latitude s and Longitude s Respectively represent the longitude and latitude information of the service, Country s , Province s and City s Respectively represent the country, province, and city where the service is located.

[0099] S2. Preprocess the context features of users and services.

[0100] The specific steps include:

[0101] S2.1. Use the one-hot encoding technique OneHot to convert the context features of users and IoT services into sparse feature vectors:

[0102]

[0103] Among them, F u,i represents the i-th feature of the user, F s,j represents the jth feature of IoT service; x u,i represents the sparse feature vector of the i-th feature of user, x s,j The sparse feature vector representing the jth feature of the IoT service; O i represents the dimension of the sparse feature vector of the user i, and also represents the number of valid values ​​of the user i feature. j It represents the dimension of the j-th sparse feature vector of the IoT service, and also represents the number of valid values ​​of the j-th feature of the IoT service.

[0104] S2.2, using embedding technology to transform the sparse feature vectors of users and IoT services into embedded vectors in the shared feature space:

[0105] (c u,i ) E =Embedding(x u,i )=W i ·x u,i

[0106] (c s,j ) E =Embedding(x s,j )=W j ·x s,j

[0107] Among them, c u,i represents the embedding vector of the i-th feature of the user, c s,j represents the embedding vector of the jth feature of the IoT service; E represents the dimension of the embedding vector; W represents the weight matrix of the feature embedding layer.

[0108] S2.3. Connect the embedding vectors of users and IoT services Φ to generate user features C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s :

[0109]

[0110] Among them, m represents the number of user context features, and n represents the number of IoT service context features.

[0111] S3. Input the processed context features into two multi-view residual interaction networks to perform response time prediction and throughput prediction respectively.

[0112] The specific steps include:

[0113] S3.1. User feature C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s Input to the first multi-view residual interaction network MRIN rt , perform response time prediction;

[0114]

[0115] in, Represents the final predicted value of the response time.

[0116] The first multi-view residual interaction network MRIN rt The first user tower First Service Tower (the two are also collectively referred to as the Twin Towers) and the First Interactive Tower composition.

[0117] Use dual towers (first user tower First Service Tower ) to learn the first user's latent features and first service potential characteristics Use the First Interactive Tower To supplement the learning of the dual towers (the first user tower First Service Tower ) Lost first interaction feature between user and service

[0118]

[0119] Among them, the first user tower First Service Tower and the first interactive tower They are all stacked by residual blocks ResBlock.

[0120] First User Tower Take this as an example, the description is as follows:

[0121]

[0122] Where K represents the number of residual blocks.

[0123] Each residual block is composed of two types of first residual units ResUnit1 and second residual units ResUnit2 stacked together:

[0124]

[0125] Where D represents the total number of residual units in each residual block. The number of the first residual unit ResUnit1 is 1, and the number of the second residual unit ResUnit2 is D-1. The first residual unit ResUnit1 can reduce the dimension of the input vector while learning the interaction between features:

[0126]

[0127]

[0128] The second residual unit ResUnit2 only learns the interaction between features without changing the dimension of the input vector:

[0129]

[0130] Where d indicates that the current residual block is the dth residual unit, and its maximum value does not exceed D; X and Y represent the output vectors of the two fully connected layers in the residual block; W and b represent the weight matrix and bias term of the fully connected layer respectively; GELU represents the activation function, and BN represents the batch normalization operation; Represents the output vector of the d-th residual unit and also serves as the input vector of the d+1-th residual unit.

[0131] The first interaction feature between users and services Input into a residual block ResBlock to get the first response time prediction value

[0132]

[0133] The first user potential characteristics and first service potential characteristics Perform inner product operation to obtain the second response time prediction value

[0134]

[0135] Wherein, T represents the vector transpose operation;

[0136] Prediction of first response time and the second response time prediction value Perform weighted operations to obtain the final predicted value of response time

[0137]

[0138] Among them, α represents the view weight, which determines the first response time prediction value The importance of , and 1-α determines the second response time prediction value The importance of.

[0139] S3.2. User feature C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s Input to the second multi-view residual interaction network MRIN tp , perform throughput prediction:

[0140]

[0141] in, Represents the final predicted value of throughput.

[0142] The Second Multi-view Residual Interaction Network MRIN tp By the second user tower Second Service Tower (the two are also collectively referred to as the Twin Towers) and the Second Interactive Tower composition.

[0143] Use dual towers (second user tower Second Service Tower ) to learn the second user's latent features and second service potential characteristics Use the second interactive tower To supplement the learning of the double tower (the second user tower Second Service Tower ) The second interaction feature between the user and the service is lost

[0144]

[0145] Among them, the second user tower Second Service Tower and the second interactive tower They are also stacked by residual blocks ResBlock.

[0146] The second interaction feature between users and services Input into a residual block ResBlock to get the first throughput prediction value

[0147]

[0148] The second user potential characteristics And the second service potential characteristics Perform inner product operation to obtain the second throughput prediction value

[0149]

[0150] The first throughput prediction value and the second throughput prediction value Perform weighted operations to obtain the final predicted value of throughput

[0151]

[0152] Among them, α represents the view weight, which determines the first throughput prediction value The importance of , and 1-α determines the second throughput prediction value The importance of.

[0153] S4. Calculate the losses between the response time prediction value and the throughput prediction value and the corresponding true value respectively, obtain the joint loss function through weighted operation, and use the joint loss function to train the prediction model.

[0154] The specific steps include:

[0155] S4.1. Calculate the first response time prediction value according to the Huber loss function Second response time prediction value and the corresponding true value y rt The first loss function between The second loss function Among them, the threshold of the Huber loss function is set to 1;

[0156] According to the first loss function Calculate first response time prediction and the true value y rt The loss between:

[0157]

[0158] According to the second loss function Calculate the second response time prediction value and the true value y rt The loss between:

[0159]

[0160] S4.2. Calculate the first throughput prediction value according to the Huber loss function Second throughput prediction value and the corresponding true value y tp The third loss function between The fourth loss function Among them, the threshold of the Huber loss function is set to 1;

[0161] According to the third loss function Calculate the first throughput prediction value and the true value y tp The loss between:

[0162]

[0163] According to the fourth loss function Calculate the second throughput prediction value and the true value y tp The loss between:

[0164]

[0165] S4.3. The first loss function The second loss function The third loss function The fourth loss function Perform weighted operations to obtain the joint loss function L, and use the joint loss function L to train the prediction model:

[0166]

[0167] Among them, β represents the attribute weight, which determines the importance of the response time loss function, and 1-β determines the importance of the throughput loss function; α represents the view weight, which determines the importance of the interactive tower loss function, and 1-α determines the importance of the dual-tower loss function.

[0168] S5. Use the trained prediction model to predict the quality of IoT services.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-attribute joint prediction method for Internet of Things service quality prediction, characterized in that: The following steps are involved: S1. Obtain IoT service quality dataset; S2, preprocessing the context features of users and services; S3: Input the processed context features into the first multi-view residual interaction network MRIN rt and the second multi-view residual interaction network MRIN tp In, response time prediction and throughput prediction are performed separately; S4, respectively calculating the loss functions between the response time prediction value and the throughput prediction value and the corresponding true value, and obtaining a joint loss function through weighted operation, and using the joint loss function to train the prediction model; S5. Use the trained prediction model to predict the quality of IoT services.

2. A multi-attribute joint prediction method for Internet of Things service quality prediction according to claim 1, characterized in that: In S1, the IoT service quality data set is obtained, including: Use service quality monitoring tools to collect IoT service quality data, except for the required user ID u , IoT service ID s In addition to the service quality value, the context features of useful users and IoT services are selected and added to the dataset; finally, the user context feature F u and IoT service context features F s It is expressed as: F u =[ID u ,AS u ,Longitude u ,Latitude u ,Country u ,Province u ,City u ] F s =[ID s ,Provider s ,AS s ,Longitude s ,Latitude s ,Country s ,Province s ,City s ] Among them, AS u Indicates the network autonomous domain to which the user's IP address belongs. Latitude u and Longitude u Respectively represent the user's longitude and latitude information, Country u , Province u and City u Respectively represent the country, province and city where the user is located; Provider s Indicates the service provider, AS s Indicates the network autonomous domain to which the service IP address belongs. Latitude s and Longitude s Respectively represent the longitude and latitude information of the service, Country s , Province s and City s Respectively represent the country, province, and city where the service is located.

3. A multi-attribute joint prediction method for Internet of Things service quality prediction according to claim 1, characterized in that: In S2, the context features of users and services are preprocessed, which specifically includes the following steps: S2.

1. OneHot encoding technology is used to convert the context features of users and IoT services into sparse feature vectors. Among them, F u,i represents the i-th feature of the user, F s,j represents the jth feature of the IoT service; x u,i represents the sparse feature vector of the i-th feature of user, x s,j The sparse feature vector representing the jth feature of the IoT service; O i represents the dimension of the sparse feature vector of the user i, and also represents the number of valid values ​​of the user i feature. j It represents the dimension of the j-th sparse feature vector of the IoT service, and also represents the number of valid values ​​of the j-th feature of the IoT service; S2.2, using embedding technology to transform the sparse feature vectors of users and IoT services into embedded vectors in a shared feature space; (c u,i ) E =Embedding(x u,i )=W i ·x u,i (c s,j ) E =Embedding(x s,j )=W j ·x s,j Among them, c u,i represents the embedding vector of the user’s i-th feature, c s,j represents the embedding vector of the jth feature of the IoT service; E represents the dimension of the embedding vector; W represents the weight matrix of the feature embedding layer; S2.

3. Connect the embedding vectors of users and IoT services Φ to generate user features C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s : Among them, m represents the number of user context features, and n represents the number of IoT service context features.

4. The multi-attribute joint prediction method for Internet of Things service quality prediction according to claim 1 is characterized by: In S3, the following steps are specifically included: S3.

1. User feature C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s Input to the first multi-view residual interaction network MRIN rt , perform response time prediction; in, represents the final predicted value of response time; The first multi-view residual interaction network MRIN rt By the first user tower First Service Tower and the first interactive tower composition; Use the First User Tower First Service Tower To learn the first user's latent characteristics and first service potential characteristics Use the First Interactive Tower To supplement the study of the first user tower First Service Tower Lost first interaction feature between user and service Among them, the first user tower First Service Tower and the first interactive tower They are all stacked by residual blocks ResBlock; The first interaction feature between users and services Input into a residual block ResBlock to get the first response time prediction value The first user potential characteristics and first service potential characteristics Perform inner product operation to obtain the second response time prediction value Wherein, T represents the vector transpose operation; Prediction of first response time and the second response time prediction value Perform weighted operations to obtain the final predicted value of response time Among them, α represents the view weight, which determines the first response time prediction value The importance of , and 1-α determines the second response time prediction value The importance of S3.

2. User feature C u 、IoT Service Characteristics C s , and the joint characteristics of users and IoT services C u,s Input to the second multi-view residual interaction network MRIN tp , perform throughput prediction: in, represents the final predicted value of throughput; The Second Multi-view Residual Interaction Network MRIN tp By the second user tower Second Service Tower and the second interactive tower composition; Using the Second User Tower Second Service Tower To learn the second user's latent features and second service potential characteristics Use the second interactive tower To supplement the learning of the second user tower and the second service tower The missing secondary interaction feature between the user and the service Among them, the second user tower Second Service Tower and the second interactive tower They are also stacked by residual blocks ResBlock; The second interaction feature between users and services Input into a residual block ResBlock to get the first throughput prediction value The second user potential characteristics And the second service potential characteristics Perform inner product operation to obtain the second throughput prediction value The first throughput prediction value and the second throughput prediction value Perform weighted operations to obtain the final predicted value of throughput Among them, α represents the view weight, which determines the first throughput prediction value The importance of , and 1-α determines the second throughput prediction value The importance of.

5. The multi-attribute joint prediction method for Internet of Things service quality prediction according to claim 1 is characterized by: In S3.1 and S3.2, the residual block ResBlock structure is based on the first user tower The stacking is described as follows: Where K represents the number of residual blocks; Each residual block is composed of two types of first residual units ResUnit1 and second residual units ResUnit2 stacked together: Where D represents the total number of residual units in each residual block; the number of the first residual unit ResUnit1 is 1, and the number of the second residual unit ResUnit2 is D-1; the first residual unit ResUnit1 can reduce the dimension of the input vector while learning the interaction between features: The second residual unit ResUnit2 only learns the interaction between features without changing the dimension of the input vector: Where d represents the dth residual unit in the current residual block, and its maximum value does not exceed D; X and Y represent the output vectors of the two fully connected layers in the residual block; W and b represent the weight matrix and bias term of the fully connected layer respectively; GELU represents the activation function, and BN represents the batch normalization operation; Represents the output vector of the d-th residual unit and also serves as the input vector of the d+1-th residual unit.

6. A multi-attribute joint prediction method for Internet of Things service quality prediction according to claim 1, characterized in that: In S4, the following steps are specifically included: S4.

1. Calculate the first response time prediction value according to the Huber loss function Second response time prediction value and the corresponding true value y rt The first loss function between The second loss function Among them, the threshold of the Huber loss function is set to 1; According to the first loss function Calculate first response time prediction and the true value y rt The loss between: According to the second loss function Calculate the second response time prediction value and the true value y rt The loss between: S4.

2. Calculate the first throughput prediction value according to the Huber loss function Second throughput prediction value and the corresponding true value y tp The third loss function between The fourth loss function Among them, the threshold of the Huber loss function is set to 1; According to the third loss function Calculate the first throughput prediction value and the true value y tp The loss between: According to the fourth loss function Calculate the second throughput prediction value and the true value y tp The loss between: S4.

3. The first loss function The second loss function The third loss function The fourth loss function Perform weighted operations to obtain the joint loss function L, and use the joint loss function L to train the prediction model: Among them, β represents the attribute weight, which determines the importance of the response time loss function, and 1-β determines the importance of the throughput loss function; α represents the view weight, which determines the importance of the interactive tower loss function, and 1-α determines the importance of the dual-tower loss function.

Images