A location-aware network service recommendation method

Abstract

The present invention relates to a location-aware network service recommendation method, and belongs to the technical fields of recommendation algorithms, service calculation, and service recommendation. The present invention considers the geographical location relationship between users and services, calculates the geographical location similarity between users and services, and then comprehensively considers the geographical location similarity, user network environment, service network environment, and service function class to predict the service quality value (QoS, Quality of Service) of related users when calling related services. And generate a service recommendation list based on the QoS value sorting and recommend it to users, so that users can accurately and quickly select suitable services to build their own service platforms, further improving the prediction accuracy and saving the economic costs and time generated by actual calls and records.

Description

Technical Field

[0001] The invention relates to a location-aware network service recommendation method, belonging to the technical fields of recommendation algorithms, service calculation and service recommendation. Background Art

[0002] In today's information explosion, recommendation algorithms, as one of the effective methods to solve information overload, have emerged many classic algorithms and have been widely used in various fields. Among them, the Internet is one of the fastest-growing fields. Various functional requirements have derived a variety of functional services. From June 2011 to July 2022, the number of network services counted on ProgrammableWeb has increased from 3261 to 24471, an increase of 750%. How to accurately find high-quality services among many services and recommend them to users has become a major problem.

[0003] Since Quality of Service (QoS) is an indicator that describes the quality of Web services, many scholars focus on the prediction and calculation of service QoS, including response time and throughput, etc. However, in real scenarios, the QoS values ​​recorded by actual service calls only account for a very small part of all network service QoS values, and the extremely sparse data makes it difficult to predict QoS.

[0004] Secondly, when faced with data features, most traditional algorithms (such as user collaborative filtering, item collaborative filtering, matrix decomposition, etc.) do not consider the impact of feature interactions or only consider the impact of second-order feature interactions. This idea will bring relatively low prediction accuracy to QoS prediction. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a location-aware network service recommendation method, which can quickly and accurately predict the quality of network services and make relevant recommendations for high-quality network services based on the network service quality.

[0006] The technical solution of the present invention is: a location-aware network service recommendation method, which comprises the following specific steps:

[0007] Step 1: Convert the geographic location text features into numerical features and use them to calculate the geographic location similarity;

[0008] Step 2: Through the LDIF method, the learning and fusion of feature interaction information will be completed in each SIS layer, so that the fusion of low-order feature interaction information will be completed in the shallow layer, and the fusion of high-order feature interaction information will be completed in the deep layer;

[0009] Step 3: Integrate the information of low-order to high-order feature interactions, combine QoS features for sorting, generate a recommendation list and make predictions.

[0010] The Step 1 is specifically as follows: the GloVe word embedding algorithm is used to convert the geographic location text features into dense numerical features of 50, 200, and 300 dimensions respectively, and then the Cosine similarity is used to calculate the similarity between the user and service geographic locations. The similarity calculation formula is as follows:

[0011]

[0012] in, is the user's geographic category vector, For the geographic category vector of the service, the value of Cosθ will be added as a new numerical feature into the original feature.

[0013] The specific step 2 is as follows: first, distinguish the relevant feature information of users and services, scan the distinguished user and service features with a fixed window size, use the deep forest multi-granularity scanning mechanism, treat local features as overall samples, thereby increasing the sample size of effective samples to alleviate the impact of data sparsity. On this basis, the feature interaction information calculation step is used to express the feature interaction information, and finally the feature interaction information generated by the local features is fused into the original features to achieve the effect of fusion feature interaction information.

[0014] The feature interaction information calculation step is specifically used to express the feature interaction information:

[0015] The scanned local features are recorded as matrix A, where A is expressed as: The size of the matrix is ​​n*m, where U 11 Represents the feature information of the first row and first column of the user, S nm Represents the feature information of the service at row n and column m.

[0016] To calculate the second-order feature interaction information based on this matrix, use A*A T To represent various combinations of second-order feature interactions;

[0017] To calculate the third-order feature interaction information based on this matrix, use A*A T *A is used to represent various combinations of third-order feature interactions; by analogy, it can be extended to feature interactions of various orders to achieve the effect of calculating feature interaction information.

[0018] The Step 3 is specifically as follows:

[0019] Assume that the input vector is 18-dimensional, of which 9 dimensions are user features and the other 9 dimensions are service features. This paper believes that feature interaction is meaningful only when users and features work together, so it is expanded into two columns of 9-dimensional features, representing user and service features respectively. When the local feature size is 2, four numerical features, U1, U2, S1, and S2, will be extracted and recorded as matrix A:

[0020]

[0021] When it is necessary to calculate the second-order and third-order feature interaction information, the calculation formulas are as follows:

[0022]

[0023]

[0024] From the above formula, we can see that in the second-order and third-order feature interaction information matrices, the four values ​​can represent the combination of different orders and their influences. Therefore, the four values ​​in the second-order and third-order feature interaction matrices are concatenated with the original four numerical features to achieve the effect of feature enhancement and fusion of feature interaction information, thereby realizing predictive recommendation.

[0025] The beneficial effects of the present invention are as follows: the present invention solves the problems of data sparsity and insufficient fusion of feature interaction information, further improves prediction accuracy and saves economic costs and time generated by actual calls and records. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of the steps of the present invention;

[0027] Figure 2 It is a diagram of the SIS layer structure of the present invention;

[0028] Figure 3 It is the structural diagram of the LDIF method of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0030] Example 1: Figure 1 As shown, a location-aware network service recommendation method includes the following specific steps:

[0031] Step 1: Convert the geographic location text features into numerical features, and use them to calculate the geographic location similarity. Most traditional word embedding algorithms are word2vec and one-hot encoding. The Glove algorithm has the following advantages over the above algorithms: a) it considers the global impact of the corpus more, b) it has high training efficiency, c) the calculation scale is proportional to the size of the corpus, and d) when the corpus is small or the vector dimension is small, it can still achieve good results and can consider the global impact of the corpus more. Therefore, the Glove word embedding algorithm is used to convert the geographic location text features into dense numerical features of 50, 200, and 300 dimensions respectively. If these numerical features are directly used for feature interaction operations, the calculation complexity will be greatly increased. Therefore, Cosine similarity is used to calculate the similarity between the user and service geographic locations. The similarity calculation formula is as follows:

[0032]

[0033] in, is the user's geographic category vector, For the geographic category vector of the service, the value of Cosθ will be added as a new numerical feature into the original feature.

[0034] Step 2: Through the LDIF method, the feature interaction information is learned and integrated in each SIS layer, so that the low-order feature interaction information is integrated in the shallow layer, and the high-order feature interaction information is integrated in the deep layer; and whether the next layer should be generated in a loop is determined based on whether the average prediction accuracy of the final output of this round is improved compared with the prediction accuracy of the previous layer. This mechanism effectively controls the complexity of the model and the depth of the hierarchy, avoiding unnecessary high-order feature interaction calculations.

[0035] The LDIF method is composed of SIS (Scaninteractstructure) layers. Since the learning and fusion of the corresponding hierarchical feature interaction information are completed in each SIS layer, the fusion of low-order feature interaction information is completed in the shallow layer, and the fusion of high-order feature interaction information is completed in the deep layer. Overall, the effect of considering low-order feature interaction information to high-order feature interaction information is achieved.

[0036] First, the relevant feature information of users and services is distinguished, and the distinguished user and service features are scanned with a fixed window size. The deep forest multi-granularity scanning mechanism is used to treat local features as overall samples, thereby increasing the sample size of effective samples to alleviate the impact of data sparsity. On this basis, the feature interaction information calculation step is used to express the feature interaction information, and finally the feature interaction information generated by the local features is fused into the original features to achieve the effect of fusion feature interaction information.

[0037] Since the characteristic information of users and services has been distinguished before, and the user and service characteristics are scanned together by a scanning window of a specific size, the scanned local features are recorded as matrix A, where A is expressed as: The size of the matrix is ​​n*m, where U 11 Represents the feature information of the first row and first column of the user, S nm Represents the feature information of the service at row n and column m.

[0038] To calculate the second-order feature interaction information based on this matrix, use A*A T To represent various combinations of second-order feature interactions;

[0039] To calculate the third-order feature interaction information based on this matrix, use A*A T *A is used to represent various combinations of third-order feature interactions; by analogy, it can be extended to feature interactions of various orders to achieve the effect of calculating feature interaction information.

[0040] Step 3: Integrate the information of low-order to high-order feature interactions, combine QoS features for sorting, generate a recommendation list and make predictions. The sorting is based on the predicted QoS value of the network service. The network service is recommended based on multiple QoS value indicators. For example, the response time (ResponseTime) is an evaluation standard of the shorter the better, and the throughput (Throughout) is an evaluation standard of the larger the better. Here, it is necessary to combine QoS features to determine whether the sorting order is descending or ascending. Similarly, since each network service has different functional characteristics, it is necessary to complete the classification according to the functional characteristics, and sort according to the QoS value under each category to generate a recommendation list. Ultimately, the effect of quickly recommending high-quality network services to users is achieved.

[0041] For the problem of fusing feature interaction information, the expression and fusion of feature interaction information rely on the processing of dimension scanning. Through the above iterative structure, the final SIS layer will fuse the information of low-order to high-order feature interactions and make predictions. Figure 2As shown in the figure, it is assumed that the input vector is 18-dimensional, of which 9 dimensions are user features and the other 9 dimensions are service features. Since feature interaction is meaningful only when users and features work together, it is expanded into two columns of 9-dimensional features, representing the features of users and services respectively. When the local feature size is 2, four numerical features U1, U2, S1, and S2 will be extracted, which are recorded as matrix A:

[0042]

[0043] When it is necessary to calculate the second-order and third-order feature interaction information, the calculation formulas are as follows:

[0044]

[0045]

[0046] From the above formula, we can see that in the second-order and third-order feature interaction information matrices, the four values ​​can represent the combination of different orders and their influences. Therefore, the four values ​​in the second-order and third-order feature interaction matrices are concatenated with the original four numerical features to achieve the effect of feature enhancement and fusion of feature interaction information, thereby realizing predictive recommendation.

[0047] Embodiment 2: A location-aware network service recommendation method, comprising the following specific steps:

[0048] When a network service call is needed, the user's location information, network-related information, and the network service's functional information, network-related information, and location information can be obtained relatively easily. Based on the above information, the geographic location similarity needs to be calculated first. The text vector is replaced with a dense numerical vector according to the Glove algorithm proposed in the invention content. The geographic location vectors of the user and the network service can be obtained here, and the Cosine function is used to calculate the geographic location similarity between the user and the network service, and enter the LDIF method together with other information.

[0049] In the LDIF method, feature interaction information is learned and integrated in each SIS layer, so that low-order feature interaction information is integrated in the shallow layer, and high-order feature interaction information is integrated in the deep layer. And whether to generate the next layer in a loop is determined based on whether the average prediction accuracy of the final output of this round is improved relative to the prediction accuracy of the previous layer. This mechanism effectively controls the complexity of the model and the depth of the layers, avoiding unnecessary high-order feature interaction calculations.

[0050] Furthermore, for the problem of fusing feature interaction information, the expression and fusion of feature interaction information rely on the processing of the dimension scanning block. Through the above iterative structure, the final SIS layer will fuse the information of low-order to high-order feature interactions and make predictions. Figure 3 As shown in the formulas (2), (3), and (4) in the invention content, when the input vector is 18-dimensional, it will be scanned 8 times in the first 2-dimensional SIS layer. If all the values ​​in the matrices (3) and (4) are directly concatenated to the original features, it will cause a dimensionality explosion problem. Therefore, it is chosen here to perform weighted averaging on the values ​​in the matrices (3) and (4), and the input vector is expanded to 34 dimensions. However, in Figure 2 In the SIS structure, the input vector will be predicted once and twice, and 6 prediction values ​​will be obtained respectively. The 6 prediction values ​​also need to be spliced ​​with the 34-dimensional input vector. Then, after passing through the first 2-dimensional SIS layer, 40-dimensional information will be output. Before entering the next layer, it is necessary to obtain the scan dimension size of the next layer. Here, the scan dimension size is set to Where k represents the dimension of the vector output by the previous layer. The reason for choosing a specific scan dimension size here is that if the scan size is too large, the dimension will grow slowly, while if the scan size is too small, the dimension will explode.

[0051] When the accuracy of the prediction results of the current SIS layer does not improve, or when the maximum depth is reached, the next SIS layer will be stopped from being generated, and the prediction results of the last layer will be used as the final prediction results, that is, the QoS prediction value. According to the predicted QoS values ​​of all network services, different categories can be sorted to output recommendation lists under different network service categories.

[0052] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A location-aware network service recommendation method, characterized in that: Step 1: Convert the geographic location text features into numerical features and use them to calculate the geographic location similarity; Step 2: Through the LDIF method, the learning and fusion of feature interaction information will be completed in each SIS layer, so that the fusion of low-order feature interaction information will be completed in the shallow layer, and the fusion of high-order feature interaction information will be completed in the deep layer; Step 3: Integrate the information of low-order to high-order feature interactions, combine QoS features to perform sorting, generate a recommendation list and make predictions; The Step 2 is specifically as follows: firstly, the relevant feature information of the user and the service is distinguished, and the distinguished user and service features are scanned with a fixed window size, and the deep forest multi-granularity scanning mechanism is used to treat the local features as the overall sample, and the feature interaction information calculation step is adopted to express the feature interaction information, and finally the feature interaction information generated by the local features is fused into the original features to achieve the effect of fusion feature interaction information; The Step 3 is specifically as follows: Assume that the input vector is 18-dimensional, of which 9 dimensions are user features and the other 9 dimensions are service features. When the local feature size is 2, four numerical features, U1, U2, S1, and S2, will be extracted and recorded as matrix A: When it is necessary to calculate the second-order and third-order feature interaction information, the calculation formulas are as follows: The four numerical values ​​in the second-order and third-order feature interaction matrices are concatenated with the original four numerical features to achieve the effect of feature enhancement and fusion of feature interaction information, thereby realizing predictive recommendation.

2. The location-aware network service recommendation method according to claim 1, characterized in that: The Step 1 is specifically as follows: the GloVe word embedding algorithm is used to convert the geographic location text features into dense numerical features of 50, 200, and 300 dimensions respectively, and then the Cosine similarity is used to calculate the similarity between the user and service geographic locations. The similarity calculation formula is as follows: in, is the user's geographic category vector, For the geographic category vector of the service, the value of Cosθ will be added as a new numerical feature into the original feature.

Images