Cloud API service quality prediction method based on three-dimensional tensor high-order feature interaction

Abstract

The application discloses a cloud API service quality prediction method based on three-dimensional tensor high-order feature interaction, and belongs to the field of service quality prediction.The prediction method comprises the following steps: step 1, obtaining a cloud API service quality dataset; step 2, designing an abnormal point detection module to detect and screen out outliers of the cloud API service quality dataset; step 3, designing a spatial position information optimization module to fully extract information generated by latitude and longitude coordinates; step 4, constructing a self-adaptive three-dimensional tensor network model, inputting processed user features and cloud API features into the three-dimensional tensor network model to obtain new feature information generated after high-order feature interaction; and step 5, inputting the new feature information into a deep neural network layer for sufficient learning, and finally obtaining a cloud API service quality prediction value through a full connection layer.The method designed in the application has more accurate prediction effect and higher robustness in various data density scenes.

Description

TECHNICAL FIELD

[0001] The application relates to the field of service quality prediction, in particular to a cloud API service quality prediction method based on three-dimensional tensor high-order feature interaction. BACKGROUND

[0002] A cloud API refers to a cloud application programming interface that can be coded by a developer, and the interface has a service capability of a cloud provider. There are a large number of cloud API services with similar functions on the Internet, so that a user is difficult to select a cloud API service suitable for the user, and therefore it is a challenging task to select a suitable service from a large number of cloud API services with similar functions. Service quality is usually used to describe non-functional characteristics (such as response time, throughput, etc.) of a cloud API service, but in practice, a single user is difficult to obtain service quality values of all candidate API services, and therefore service quality prediction of an API service becomes a primary challenge of service recommendation.

[0003] A deep learning-based algorithm can model data types that cannot be modeled by a traditional modeling algorithm, and can extract hidden feature information to improve the effect of a recommendation algorithm. One of the commonly used methods is a multilayer perceptron, which extracts more feature information through a hidden layer, but a traditional deep learning method does not explicitly extract high-order feature cross information. In order to solve the problem of cross feature capture, scholars try to explore network design. There are methods of manually extracting information through cross features after screening first-order features based on GBDT+LR, methods of automatically extracting information through cross features based on FM, and methods of implicitly extracting cross feature information through a hidden layer based on MLP. Although the above methods can extract low-order cross feature information, they do not explicitly extract high-order cross feature information. SUMMARY

[0004] The application needs to solve the technical problem of providing a cloud API service quality prediction method based on three-dimensional tensor high-order feature interaction. The method combines an adaptive three-dimensional tensor and deep learning to solve the problem of the lack of explicit high-order feature cross in the existing cloud API service quality prediction field, can explicitly extract high-order cross feature information, and can improve the prediction accuracy of the model under the condition of data sparsity.

[0005] To solve the above technical problems, the technical solution adopted by the application is:

[0006] A cloud API service quality prediction method based on three-dimensional tensor high-order feature interaction, comprising the following steps:

[0007] Step 1, obtaining a cloud API service quality data set;

[0008] According to the service quality prediction target of the business requirement, a cloud API service quality dataset is collected;

[0009] Step 2, an abnormal point detection module is designed to detect and screen out abnormal outliers in the cloud API service quality dataset;

[0010] When the cloud API service quality data collection is completed, part of the service quality collection values are collected in the network congestion state, which will have a large deviation from the service quality performance in the network smooth state, therefore, the outlier detection technology is used to screen out the cloud API service quality data from the dataset;

[0011] Step 3, a spatial position information optimization module is designed to fully extract the information generated by the latitude and longitude coordinates;

[0012] Usually, the latitude and longitude coordinates representing the geographical position in the collected dataset are in numerical form, and the latitude and longitude coordinates are extracted as specific geographical coordinate information to enhance the interpretability of the features;

[0013] Step 4, a self-adaptive three-dimensional tensor high-order feature interaction network model is constructed, the user features and the cloud API features processed in steps 2-3 are input into the network model, and new feature information generated after high-order feature interaction is obtained;

[0014] The cloud API service quality dataset processed in steps 2-3 is input into the three-dimensional tensor high-order feature interaction model to generate a result vector after high-order feature interaction;

[0015] Step 5, the new feature information is input into a deep neural network layer, and the result vector is obtained through a full connection layer to obtain the cloud API service quality prediction value;

[0016] The result vector of step 4 is input into a deep neural network layer for sufficient learning, and then a full connection layer is used to obtain the final prediction result.

[0017] The further improvement of the technical scheme of the application is that the cloud API service quality dataset in step 1 is specifically:

[0018] The cloud API service quality monitoring tool is used to collect cloud API service quality data in different regions in a random distribution manner, in addition to the necessary user label, cloud API label and service quality value, beneficial user and cloud API feature data are selected to join the collected service quality dataset; finally, the user features UF and the cloud API features AF are represented as:

[0019] UF=UID+∑U c

[0020] AF=AID+∑A c

[0021] Wherein, UID is a user identifier, AID is a cloud API identifier, UID and AID are both mandatory items; U c Indicates the selected user beneficial features, A c Indicates the selected cloud API beneficial features.

[0022] Further improvement of the technical scheme of the application is that the abnormal point detection module in step 2 is specifically:

[0023] Based on the cloud API service quality data set collected in step 1, a two-dimensional matrix ua is constructed, with user identifier UID as row and cloud API identifier AID as column, and the value in the matrix is the service quality value; two two-dimensional sets are constructed, which are user outlier set o user and cloud API outlier set oapi respectively;

[0024] Traverse each row of the matrix ua, sort the service quality values of all cloud API identifiers AID corresponding to the user identifier UID of the row from small to large, calculate the average value AR of all values within the 50th percentile and the 75th percentile of the data, and set the outlier determination threshold k, then all data points greater than AR*k are regarded as outliers, and the formula for detecting user-side outliers is as follows:

[0025]

[0026] Wherein, o user is a user outlier set; ua represents a two-dimensional matrix with user identifier UID as row and cloud API identifier AID as column, and the value is the service quality value; i represents the row of the matrix ua; j represents the column of the matrix ua; m represents the total number of rows of the matrix ua; n represents the total number of columns of the matrix ua; AR is the average value of all values within the 50th percentile and the 75th percentile of the data; k is the determination threshold of outliers;

[0027] Similarly, traverse each column of the matrix ua again, sort the service quality values of all user identifiers UID corresponding to the cloud API identifier AID of the column from small to large, calculate the average value AR of all values within the 50th percentile and the 75th percentile of the data, and set the outlier determination threshold k, then all data points greater than AR*k are regarded as outliers, and the formula for detecting cloud API-side outliers is as follows:

[0028]

[0029] Wherein, oapi is a cloud API outlier set; ua represents a two-dimensional matrix with a user identifier UID as a row and a cloud API identifier AID as a column, and the value is a service quality value; i represents a row of the matrix ua; j represents a column of the matrix ua; m represents the total number of rows of the matrix ua; n represents the total number of columns of the matrix ua; AR is the average value of all values in the range of the 50th percentile and the 75th percentile of data; k is a determination threshold of outliers;

[0030] After the construction of the two two-dimensional sets of ouser and oapi, if all combinations of rows and columns of the matrix ua, assuming that the row is tu and the column is ta, if tu is in oapi[ta] and ta is in ouser[tu], it is indicated that the value of ua[tu][ta] is a specific point in the network congestion state, and the detection function check is defined as follows:

[0031] check(tu,fa)=(fa∈ouser[tu])∧(tu∈oapi[ta])

[0032] Wherein, tu is a user identifier; ta is a cloud API identifier; ouser is a user outlier set; oapi is a cloud API outlier set; check is an outlier specific point, 1 is an outlier specific point, and 0 is not; after the check function is true, the corresponding outlier is deleted from the data set.

[0033] The further improvement of the technical scheme of the application is that the spatial position information optimization module in step 3 is specifically:

[0034] The original longitude and latitude coordinate features are replaced after analyzing the longitude and latitude coordinates; the longitude and latitude coordinate analysis formula T LL The definition is as follows:

[0035] T LL (Lon, Lat)=(Country, Prov, City, Street)

[0036] Wherein, Lon is longitude, Lat is latitude, Country is country, Prov is province, City is city, and Street is street.

[0037] The further improvement of the technical scheme of the application is that the three-dimensional tensor high-order feature interaction network model in step 4 is specifically:

[0038] The sparse features in the embedding vector high-order feature interaction stage are converted into embedding vectors for modeling, all features are converted into integer encoding by using LabelEncoder, and then all features are converted by using Embedding function, and the definition is as follows:

[0039] UID Emb = Embedding(LabelEncoder(UID))

[0040] AID Emb = Embedding(LabelEncoder(AID))

[0041] U cEmb = Embedding(LabelEncoder(U c ))

[0042] A cEmb = Embedding(LabelEncoder(A c ))

[0043] Wherein, the LabelEncoder method is to encode each feature into an integer, and the Embedding method is to convert the integer- encoded feature into an embedding vector;

[0044] The high-order feature interaction method of the self- adaptive three-dimensional tensor to the embedding vector adopts a TDN module to generate; after being converted into an embedding vector for processing, a self- adaptive three- dimensional tensor W is constructed, the first dimension is the number of features of the data set, the second dimension is the length of the mapping vector set in the Embedding process, and the third dimension is the depth of feature cross; the cross method of the TDN is to repeatedly perform Hadamard product on each embedding vector in the data to obtain a new feature interaction vector, and the interaction of the embedding vector in the self- adaptive three- dimensional tensor is defined as follows:

[0045] x l+1 = x0⊙(W l x l +b l )+x l

[0046] Wherein, x l is a feature vector of the lth intermediate layer of the TDN, x0 is an initial embedding vector, W l is a matrix obtained by taking l as the first dimension of the three- dimensional tensor set in the front, and b l is a bias term of the lth intermediate layer.

[0047] The further improvement of the technical scheme of the application is that the deep neural network layer in step 5 is specifically:

[0048] A result vector set is obtained from a multilayer perceptron, wherein a plurality of hidden layers are set, and the result vector after cross in the TDN module is input into the perceptron as an initial h0, and the perceptron is defined as follows:

[0049] h l+1 = f(Wl h l +b l )

[0050] wherein, h l represents the value in the process of neuron transmission of the feature vector of the lth intermediate layer of the multi-layer perception, W l is a weight matrix, and b l is a bias term of the lth intermediate layer.

[0051] Further improvement of the technical scheme of the application is that the full connection layer in step 5 is specifically:

[0052] The full connection layer outputs a prediction result of a cloud API service quality value, and the formula definition is as follows:

[0053] QoS pred =f(W 1×n h MLP )

[0054] wherein, W 1xn is a weight matrix with 1 row and n columns, h is a result vector of a deep neural network layer, and the QoS pred output by the last layer is a cloud API service quality prediction result.

[0055] Due to the adoption of the above technical scheme, the application has achieved the following technical progress:

[0056] 1. The outlier detection module designed by the application can reduce the influence of abnormal service quality prediction values collected under network congestion state on the model, and the scheme is simple and effective, thereby improving the prediction accuracy of the model.

[0057] 2. The spatial position information optimization module designed by the application can enhance the feature relationship between adjacent users, better utilize the interpretability of latitude and longitude coordinate information, and better classify users and services in terms of regions.

[0058] 3. The three-dimensional tensor high-order feature interaction module designed by the application can fully mine new feature information after high-order feature interaction between original features, fully utilize existing feature data, improve the prediction effect of the model, and has lower calculation complexity, and has high applicability in multiple scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical scheme in the embodiments of the application or the prior art, the drawings needed to be used in the embodiment or the prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor;

[0060] Figure 1 is a flowchart of the method for adaptively implementing three-dimensional tensor high-order feature interaction in the embodiment of the application;

[0061] Figure 2 is a schematic diagram of the overall structure of the model for adaptively implementing three-dimensional tensor high-order feature interaction in the embodiment of the application. DETAILED DESCRIPTION

[0062] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to the process, method, product or device.

[0063] The application will be further described in detail below in combination with the drawings and embodiments:

[0064] As shown in Figure 1 , a cloud API service quality prediction method based on three-dimensional tensor high-order feature interaction includes the following contents:

[0065] Content 1, the RT data set and the TP data set in the WSDREAM data set are used for training and test verification, first, the specific outliers in the data set are cleaned based on the outlier detection method, and then the beneficial features are selected to obtain the cleaned data set.

[0066] Content 1-1, the data set is collected in multiple regions by using a network API request tool, and contains multiple user features and cloud API service features.

[0067] The user in the data set contains user label UID, user IP address User IP Address, user IP country User Country, user IP number User IP No., user IP application server User AS, user longitude information UserLat, user latitude information User Lon, etc.

[0068] The API in the data set contains cloud API label AID, cloud API network service address WSDL Address, cloud API service domain name Service Provider, cloud API IP address API IP Address, cloud API IP country API Country, cloud API IP number API IP No., cloud API IP application server API AS, cloud API longitude information API Lat, cloud API latitude information API Lon, etc.

[0069] Content 1-2, Outlier detection method, a two-dimensional matrix ua (matrix_user_api) is constructed with user identifier UID as row and cloud API identifier AID as column, and the value in the matrix is the service quality value. Two two-dimensional sets are constructed, which are user outlier set ouser (outliers_user) and cloud API outlier set oapi (outliers_api), respectively.

[0070] Each row of the matrix ua is traversed, and the service quality values of all cloud API identifiers AID corresponding to the user identifier UID of the row are sorted from small to large, and the average value AR (average range) of all values within the 50th percentile and the 75th percentile of the data is calculated. The determination threshold of the outlier is k, and then all data points greater than AR*k are regarded as outliers, and the formula for detecting user-side outliers is as follows:

[0071]

[0072] Where, ouser is the user outlier set; ua represents a two-dimensional matrix with user identifier UID as row and cloud API identifier AID as column, and the value is the service quality value; i represents the row of the matrix ua; j represents the column of the matrix ua; m represents the total number of rows of the matrix ua; n represents the total number of columns of the matrix ua; AR is the average value of all values within the 50th percentile and the 75th percentile of the data; k is the determination threshold of the outlier.

[0073] Similarly, each column of the matrix ua is traversed again, and the service quality values of all user identifiers UID corresponding to the cloud API identifier AID of the column are sorted from small to large, and the average value AR (average range) of all values within the 50th percentile and the 75th percentile of the data is calculated. The determination threshold of the outlier is k, and then all data points greater than AR*k are regarded as outliers, and the formula for detecting cloud API-side outliers is as follows:

[0074]

[0075] Where, oapi is the cloud API outlier set; ua represents a two-dimensional matrix with user identifier UID as row and cloud API identifier AID as column, and the value is the service quality value; i represents the row of the matrix ua; j represents the column of the matrix ua; m represents the total number of rows of the matrix ua; n represents the total number of columns of the matrix ua; AR is the average value of all values within the 50th percentile and the 75th percentile of the data; k is the determination threshold of the outlier.

[0076] After the two-dimensional sets of ouser and oapi are constructed, if in all combinations of the matrix ua, assuming that the behavior tu, the column is ta, if tu appears in oapi[ta] and ta appears in ouser[tu], it is explained that the value of ua[tu][ta] is a specific point in the network congestion state, and the detection function check is defined as follows:

[0077] check(tu, ta) = (ta e ouser[tu]) A (tu e oapi[ta])

[0078] Wherein, tu is a certain user label assumed; ta is a certain cloud API label assumed; ouser is a user outlier point set; oapi is a cloud API outlier point set. Check is an inspection whether it is an outlier specific point, 1 is an outlier specific point, 0 is not. After the check function is true, the corresponding outlier point is deleted from the data set.

[0079] Content 1-3, the feature engineering of longitude and latitude coordinates, the longitude and latitude coordinate conversion code is written to realize the following mapping, and the longitude and latitude coordinate analysis formula T LL The definition is as follows:

[0080] T LL (Lon, Lat) = (Country, Prov, City, Street)

[0081] Wherein, Lon is the longitude, Lat is the latitude, Country is the country, Prov is the province, City is the city, and Street is the street. Country, Prov, City, Street are the strings spliced by the country, province, city and street corresponding to the coordinates, T LL The four strings obtained are used as sparse feature fields to replace the original dense feature longitude and latitude coordinate fields.

[0082] Content 2, construct a self-adapting three-dimensional tensor high-order feature interaction model, the three-dimensional tensor high-order feature interaction model proposed by the application is as shown in Figure 2 The self-adapting three-dimensional tensor high-order feature interaction model includes an embedding vector high-order feature cross stage and a cross multi-layer perception machine training stage, the embedding vector high-order feature cross stage includes sparse feature conversion embedding vector modeling, self-adapting three-dimensional tensor initialization, and vector cross scheme design and application.

[0083] Content 2-1: The sparse feature conversion of the embedding vector high-order feature interaction stage is modeled by converting all features into integer codes using LabelEncoder first, and then converting all features using the Embedding function, which is defined as follows:

[0084] UID Emb = Embedding(LabelEncoder(UID))

[0085] AID Emb = Embedding(LabelEncoder(AID))

[0086] U cEmb = Embedding(LabelEncoder(U c ))

[0087] A cEmb = Embedding(LabelEncoder(A c ))

[0088] wherein the LabelEncoder method is to encode each feature into an integer, and the Embedding method is to convert the integer-coded feature into an embedding vector.

[0089] Content 2-2: The high-order feature interaction method of the embedding vector using a self-adaptive three-dimensional tensor generates a TDN module. After converting to an embedding vector, a self-adaptive three-dimensional tensor W is constructed, the first dimension is the number of features of the data set, the second dimension is the mapping vector length set in the Embedding process, and the third dimension is the depth of feature cross. The cross method of TDN is to repeatedly do Hadamard product on each embedding vector in the data to get a new feature interaction vector. The interaction of the embedding vector in the self-adaptive three-dimensional tensor is defined as follows:

[0090] x l+1 = x0⊙(W l x l +b l )+x l

[0091] wherein x l is the feature vector of the lth intermediate layer of TDN, x0 is the initial embedding vector, W l is the matrix obtained by setting the three-dimensional tensor to l, and b lThe bias term of the lth intermediate layer is calculated layer by layer, and after all the intermediate output result vectors are accumulated, they are transmitted to the next layer. During the calculation of the Hadamard product in each layer, the components of each dimension of x are crossed with the components of other dimensions, so the maximum value of l is also a hyperparameter that can control the maximum feature cross order that the adaptive three-dimensional tensor layer can achieve.

[0092] Content 2-3: Multi-layer perceptron, which processes the high-order feature interaction vector after TDN processing to obtain a result vector set, wherein a plurality of hidden layers are set, and the result vector after crossing in the TDN module is transmitted to the perceptron as the initial h0, and the perceptron is defined as follows:

[0093] h l+1 =f(W l h l +b l )

[0094] Where h l represents the feature vector of the lth intermediate layer of the multi-layer perceptron, x0 is the initial embedding vector, W l is the weight matrix, and b l is the bias term of the lth intermediate layer.

[0095] Content 2-4: Finally, a fully connected layer is added to output a predicted result of API response time or throughput, and its formula is defined as follows:

[0096] QoS pred =f(W 1×n h MLP )

[0097] Where W 1xn is a weight matrix with 1 row and n columns, h is the result vector of the deep neural network layer, and the QoS pred output by the last layer is the prediction result.

[0098] Content 3: Training of the adaptive three-dimensional tensor high-order feature interaction model: the multi-layer perceptron training module includes a fully connected layer module and a mean square error loss function, the fully connected layer module converts the result vector into a predicted value, and the mean square error loss function is used to guide the training of the adaptive three-dimensional tensor high-order feature cross model; the mean square error loss is represented as:

[0099]

[0100] Where L is the mean square error loss, N is the total number of samples, y i is the true value of the ith sample, and y is the predicted value of the ith sample under the adaptive three-dimensional tensor model.

[0101] Content 3-1: Selecting appropriate hyperparameters of the neural network, in addition to the learning rate, step size and other parameters required by traditional deep learning, the embedding vector length and the self-adaptive three-dimensional tensor layer feature cross-depth need to be determined in advance in the TDN model.

[0102] Content 3-2: Training the self-adaptive three-dimensional tensor model TDN and using MSE as the loss function for training, when the number of iterations reaches the set number or the loss error is less than the set threshold, the training stops, thereby obtaining the optimal self-adaptive three-dimensional tensor model TDN; the weight decay method is introduced in the calculation process, and L2 regularization is used for constraint, and the implementation scheme is as follows:

[0103]

[0104] wherein, new loss is a new loss function, L is a mean square error loss, is an L2 regularization constraint term.

[0105] Content 4: After the training is completed, the cloud API features and user features to be predicted are input into the trained model, and the inference result of the model is used as the predicted value to predict the service quality of the cloud API.

[0106] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A cloud API quality of service prediction method based on three-dimensional tensor high-order feature interaction, characterized by: The method comprises the following steps: Step 1, obtaining a cloud API service quality dataset; According to the service quality prediction target of the business requirement, the cloud API service quality dataset is collected; Specifically, the cloud API service quality data is collected in a random distribution manner in different regions by using a cloud API service quality monitoring tool, and in addition to the necessary user identifier, cloud API identifier and service quality value, beneficial user and cloud API feature data are selected to be added to the collected service quality dataset; the final user feature UF and cloud API feature AF are represented as: Wherein, UID is a user identifier, AID is a cloud API identifier, UID and AID are both mandatory; U c represents a selected user beneficial feature, A c represents a selected cloud API beneficial feature; Step 2, designing an outlier detection module to detect and exclude outliers in the cloud API service quality dataset; Step 3, designing a spatial position information optimization module to fully extract information generated by latitude and longitude coordinates; The latitude and longitude coordinates are extracted into specific geographic coordinate information to enhance the interpretability of the user features and the cloud API features; Step 4, constructing a self-adaptive three-dimensional tensor high-order feature interaction network model, inputting the enhanced user features and cloud API features in step 3 into the TDN network model to obtain new feature information generated after high-order feature interaction; The new feature information is as follows: wherein x l is the feature vector of the lth intermediate layer of TDN, x0is the initial embedding vector, W l is the matrix obtained by taking the first dimension of the three-dimensional tensor set in the front, b l is the bias term of the lth intermediate layer; The cloud API service quality dataset processed in steps 2-3 is input into the three-dimensional tensor high-order feature interaction model to generate a result vector after high-order feature interaction; Step 5, inputting the new feature information into a deep neural network layer, and obtaining a cloud API service quality prediction value through a fully connected layer from the result vector; The result vector in step 4 is input into a deep neural network layer for sufficient learning, and then a fully connected layer is used to obtain the final prediction result.

2. The cloud API quality of service prediction method based on three-dimensional tensor high-order feature interaction of claim 1, characterized in that: The outlier detection module in step 2 is specifically: Based on the cloud API service quality dataset collected in step 1, a two-dimensional matrix ua is constructed with the user identifier UID as the row and the cloud API identifier AID as the column, and the values in the matrix are service quality values; two two-dimensional sets are further constructed, namely a user outlier set o user and a cloud API outlier set o api; Traverse each row of the matrix ua, sort the quality of service values of all cloud API identifiers AID corresponding to the user identifier UID of the row from small to large, calculate the average AR of all values in the range of the 50th percentile and the 75th percentile of data, and set the threshold value k for the outlier as follows: All data points greater than the threshold value k are regarded as outliers, and the formula for detecting user-side outliers is as follows: Wherein, o user is a user outlier set; ua represents a two-dimensional matrix with the user identifier UID as the row and the cloud API identifier AID as the column, and the values are service quality values; i represents the row of the matrix ua; j represents the column of the matrix ua; m represents the total number of rows of the matrix ua; n represents the total number of columns of the matrix ua; AR is the average value of all values within the 50th and 75th percentiles of the data; k is the determination threshold of outliers; Similarly, traverse each column of the matrix ua again, sort the quality of service values of all user identifiers UID corresponding to the cloud API identifier AID in the column from small to large, calculate the average AR of all values in the range of the 50th percentile and the 75th percentile of data, and set the threshold value k for the outlier. All data points greater than are regarded as outliers, and the formula for detecting cloud API side outliers is as follows: Wherein, o api is a cloud API outlier set; ua represents a two-dimensional matrix with the user identifier UID as the row and the cloud API identifier AID as the column, and the values are service quality values; i represents the row of the matrix ua; j represents the column of the matrix ua; m represents the total number of rows of the matrix ua; n represents the total number of columns of the matrix ua; AR is the average value of all values within the 50th and 75th percentiles of the data; k is the determination threshold of outliers; After constructing the two-dimensional set of ouser and oapi, if the matrix ua has all combinations of rows and columns, assuming that the behavior tu, column is ta, if tu appears in oapi[ta] and ta appears in ouser[tu], it indicates that the value of ua[tu][ta] is a specific point in the network congestion state, and the detection function check is defined as follows: Where, tu is a certain user label assumed; ta is a certain cloud API label assumed; ouser is a user outlier point set; oapi is a cloud API outlier point set; check is an outlier specific point check, 1 is an outlier specific point, 0 is not; After the check function is true, the corresponding outlier point is deleted from the data set.

3. The cloud API quality of service prediction method based on three-dimensional tensor high-order feature interaction of claim 1, characterized in that: The spatial position information optimization module in step 3 is specifically: The original longitude and latitude coordinates are replaced by analyzing the longitude and latitude coordinates; the longitude and latitude coordinate analysis formula T LL are defined as follows: Where, Lon is the longitude, Lat is the latitude, Country is the country, Prov is the province, City is the city, and Street is the street.

4. The cloud API quality of service prediction method based on three-dimensional tensor high-order feature interaction of claim 1, characterized in that: The three-dimensional tensor high-order feature interaction network model in step 4 is specifically: The sparse features in the embedding vector high-order feature interaction stage are converted into embedding vectors for modeling. First, all features are converted into integer encoding using LabelEncoder, and then all features are converted using the Embedding function. The definition is as follows: Where, LabelEncoder method is to encode each feature into an integer, and Embedding method is to convert the integer encoded feature into an embedding vector; The adaptive three-dimensional tensor high-order feature interaction method for embedding vectors uses the TDN module to generate; after converting into embedding vectors, an adaptive three-dimensional tensor W is constructed, the first dimension is the number of features of the data set, the second dimension is the length of the mapping vector set in the Embedding process, and the third dimension is the depth of feature cross. The cross method of TDN is to repeatedly do Hadamard product on each embedding vector in the data to get a new feature interaction vector.

5. The cloud API quality of service prediction method based on three-dimensional tensor high-order feature interaction according to claim 1, characterized in that: The deep neural network layer in step 5 is specifically: The result vector set is obtained by multilayer perceptron, where multiple hidden layers are set, and the result vector after cross in the TDN module is input into the perceptron as the initial h0, and the perceptron is defined as follows: where h l represents the feature vector of the lth intermediate layer of the multi-layer perceptron, W l is the weight matrix, and b l is the bias term of the lth intermediate layer.

6. The cloud API quality of service prediction method based on three-dimensional tensor high-order feature interaction of claim 1, characterized in that: The full connection layer in step 5 is specifically: The full connection layer outputs a prediction result of the cloud API service quality value, and its formula definition is as follows: wherein W 1xn is a weight matrix of 1 row and n columns, h is a result vector of a deep neural network layer, and QoS pred is a cloud API service quality prediction result.

Images