A method for predicting user experience rate of wireless communication network
By building a random forest model, performing feature screening and two trainings, the user experience rate prediction problem in wireless communication networks is solved, and efficient user experience rate prediction and optimization are achieved.
Patent Information
- Application Number
- CN202310161891.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-24
- Publication Date
- 2025-05-16
- Estimated Expiration
- 2043-02-24
AI Technical Summary
The prior art is difficult to effectively predict the user experience rate in wireless communication networks, resulting in the problem of insufficient user experience.
By obtaining and preprocessing the user's scheduling-related information, a random forest model is built for training, important features are screened, and two trainings are performed to improve prediction accuracy.
It realizes intelligent prediction of user experience rate, improves prediction efficiency, and provides a basis for user experience optimization.
Smart Images
Figure CN116170834B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of communication technology and relates to a method for predicting user experience rate of a wireless communication network. The present invention can be applied to intelligent operation, maintenance and optimization of wireless communication networks. Background Art
[0002] A wireless communication network is a communication network composed of wireless communication technology, communication equipment, communication standards and protocols, in which communication terminals can access the network and rely on the network to communicate with each other. In recent years, the fastest-growing and most widely used technology in the field of information and communication is wireless communication technology. Typical scenarios of 5G wireless communication technology involve various areas where people live, work, relax and travel in the future, especially dense residential areas, offices, open-air gatherings, high-speed railways, subways and wide-area coverage. In these scenarios, typical 5G services such as virtual reality, ultra-high-definition video, cloud storage, smart home, Internet of Vehicles, and OTT messaging must be considered. Combined with the possible user distribution, the proportion of various services and the requirements for rate and latency in each scenario, the ITU-R organization has unified and determined the performance indicators of 5G wireless communication technology, including user experience rate, user peak rate, mobility, end-to-end latency, connection density, service density, average spectrum efficiency and energy efficiency. The user experience rate refers to the minimum transmission rate that users can obtain in a real network environment, reflecting the real experience of user services. The prediction of the user experience rate can solve the shortcomings of user experience in a targeted manner. At present, in the existing technology, there is no relevant technology for predicting the user experience rate because the data collection of the user experience rate is time-consuming and labor-intensive, and the sample size is limited. Summary of the invention
[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and propose a method for predicting user experience rate of a wireless communication network, which can realize the prediction of user experience rate.
[0004] To achieve the above object, the technical solution adopted by the present invention includes the following steps:
[0005] (1) Obtain training sample set and test sample set:
[0006] (1a) Obtain the scheduling-related information S of M users after eliminating missing values and outliers v1 ={X (m) |1≤m≤M}, and according to the traffic volume of each user tr m and business duration tt m Calculate the user experience rate tv for each user m , where M>3000,X (m) Indicates that the mth user contains scheduling related information with Q-dimensional features, Q>20;
[0007] (1b) Through each user's experience rate tv m The corresponding scheduling related information X (m) Mark and randomly select S v1 The scheduling related information of U users and their corresponding labels constitute the training sample set X train , S v1 The remaining V users’ scheduling related information constitutes the test sample set X test ,in, M=U+V;
[0008] (2) Constructing a random forest model:
[0009] Construct a random forest model F consisting of H decision trees arranged in parallel and an arithmetic mean module connected sequentially to all decision trees. Each decision tree f h The decision tree f contains a root node, multiple split nodes connected to the root node, and multiple leaf nodes connected to each split node. The connection between the root node and each split node, and the connection between each split node and each leaf node are h directed edges, where H ≥ 50, f h is the hth decision tree;
[0010] (3) Perform the first training of the random forest model:
[0011] (3a) Initialize the random forest model to contain the maximum number of decision trees H ^ , H ^ ≥1000,H=50;
[0012] (3b) The training sample set X train As the input of the random forest model F, each decision tree f h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X u ), the arithmetic average module performs weighted average calculation on the predicted experience rate of each user obtained by the H decision trees to obtain the predicted experience rate of U users
[0013] (3c) adopts the mean square error loss function and passes and User Experience Rate Marker Calculate the loss value L of the random forest model F MSE , and for L MSE and H for preservation;
[0014] (3d) Determine H≥H^ Is it true? If so, for the saved L MSE and H with H as the horizontal axis, L MSE Perform drawing operation for the ordinate and select the ordinate loss value L according to the drawing result MSE The horizontal coordinate H corresponding to the minimum value of the first trained random forest model F containing H decision trees is obtained. * , otherwise, let H = H + 1 and execute step (3b);
[0015] (4) Screening characteristics:
[0016] Using the first trained random forest model F * , the importance of the Q-dimensional features of the user's scheduling-related information is evaluated by a feature evaluation method based on impurity reduction, and then the g-dimensional features of the scheduling-related information of each user whose importance is greater than a preset threshold t are selected to form a feature set B;
[0017] (5) Perform a second training on the random forest model:
[0018] (5a) Initialize the random forest model with the maximum number of decision trees H ^ , H ^ ≥1000, and let H1=50, where H1 is the number of decision trees included in the second training of the random forest model;
[0019] (5b) Select training sample set X train The scheduling related information of U users with g-dimensional features of feature set B and their corresponding labels form a new training sample set D train , and D train It is used as the input of the random forest model F and iteratively trained to obtain the second trained random forest model T containing H1 decision trees. * ;
[0020] (6) Obtain the prediction result of user experience rate:
[0021] The test sample set X test As the second trained random forest model T * The input of each decision tree f h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X v ), the arithmetic average module performs weighted average calculation on the predicted experience rate of each user obtained by H1 decision trees to obtain the predicted experience rate of V users
[0022]
[0023] Compared with the prior art, the present invention has the following advantages:
[0024] (1) The present invention evaluates the importance of Q-dimensional features of the user's scheduling-related information by using the random forest model trained for the first time, and can screen a feature set that is highly sensitive to the user experience rate of the wireless communication network and can most affect and reflect the user experience rate.
[0025] (2) The present invention uses the random forest model trained for the second time to intelligently predict the user experience rate based on the feature set while ensuring that the accuracy of the test set is not reduced, thereby improving the efficiency of predicting the user experience rate and providing a basis for user experience issues. It can solve the shortcomings of user experience in a targeted manner and provide ideas for the optimization of wireless communication networks. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is a flow chart for implementing the present invention. DETAILED DESCRIPTION
[0027] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0028] Reference Figure 1 , the present invention comprises the following steps:
[0029] Step 1: Get the training sample set and the test sample set:
[0030] Step 1a) Obtain the scheduling related information S of M users after eliminating missing values and outliers v1 ={X (m) |1≤m≤M}, and according to the traffic volume of each user tr m and business duration tt m Calculate the user experience rate tv for each user m , where M = 9959, X (m)Indicates that the m-th user contains scheduling-related information of Q-dimensional features, Q=89, and the m-th user contains scheduling-related information of 89-dimensional features, including UE power margin, RSRP of uplink SRS, mobility status of different users, RANK value and other scheduling-related information; Missing value elimination mainly includes simple deletion method and weight method. Simple deletion method is the most primitive method for processing missing values. It deletes samples with missing values. If the data missing problem can be achieved by simply deleting a small part of the samples, this method is the most effective; the weight method is that when the type of missing value is non-completely randomly missing, the deviation can be reduced by weighting the complete data. After marking the samples with incomplete data, different weights are given to the complete data samples; the present invention completes the missing value elimination through the simple deletion method; the present invention eliminates outliers. Outlier elimination is to directly delete the outliers, which is equivalent to not having the outliers. If there are not many outliers, it is recommended to use this method. If there are many outliers, the outlier filling method can be used;
[0031] Step 1b) By each user's experience rate tv m The corresponding scheduling related information X (m) Mark and randomly select S v1 The scheduling related information of U users and their corresponding labels constitute the training sample set X train , S v1 The remaining V users’ scheduling related information constitutes the test sample set X test , in this embodiment, U=0.7M, V=0.3M;
[0032] Step 2: Build a random forest model:
[0033] Construct a random forest model F consisting of H decision trees arranged in parallel and an arithmetic mean module connected sequentially to all decision trees. Each decision tree f h The decision tree f contains a root node, multiple split nodes connected to the root node, and multiple leaf nodes connected to each split node. The connection between the root node and each split node, and the connection between each split node and each leaf node are h directed edges, where H ≥ 50, f h is the hth decision tree; each decision tree f of the random forest model h Random sampling is used, and the training samples used are not exactly the same. The variance of the trained model is small, the generalization ability is strong, and the H decision trees are highly parallelized, which has advantages in the training speed of large samples in the big data era.
[0034] Step 3: Perform the first training of the random forest model:
[0035] Step 3a) Initialize the random forest model to contain the largest number of decision trees H^ , this embodiment H ^ =1000,H=50;
[0036] Step 3b) Set the training sample set X train As the input of the random forest model F, each decision tree f h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X u ), the arithmetic average module performs weighted average calculation on the predicted experience rate of each user obtained by the H decision trees to obtain the predicted experience rate of U users Prediction results The expression is:
[0037]
[0038] Step 3c) adopts the mean square error loss function and passes and User Experience Rate Marker Calculate the loss value L of the random forest model F MSE , and for L MSE and H are saved, where the loss value L MSE The expression is:
[0039]
[0040] Step 3d) Determine H≥H ^ Is it true? If so, for the saved L MSE and H with H as the horizontal axis, L MSE Perform drawing operation for the ordinate and select the ordinate loss value L according to the drawing result MSE The horizontal coordinate H corresponding to the minimum value of the first trained random forest model F containing 780 decision trees is obtained. * , otherwise, let H = H + 1 and execute step (3b); The purpose of the first training of the random forest model is to use the random forest model F * Reduce the Q-dimensional features of the user's scheduling-related information, and select the feature set that is highly sensitive to the user's experience rate and can best influence and reflect the user's experience rate;
[0041] Step 4, filter features:
[0042] Using the first trained random forest model F *, the importance of 89-dimensional features of the user's scheduling-related information is evaluated by a feature evaluation method based on impurity reduction, and then 15-dimensional features of each user's scheduling-related information whose importance is greater than a preset threshold of 0.02 are selected to form a feature set B;
[0043] Step 5: Train the random forest model for the second time:
[0044] Step 5a) Initialize the random forest model with the maximum number of decision trees H ^ , H …… =1000, and let H1=50, where H1 is the number of decision trees for the second training of the random forest model;
[0045] Step 5b) Select training sample set X train The scheduling related information of U users and their corresponding labels containing the 15-dimensional features of feature set B constitute a new training sample set D train , and D train It is used as the input of the random forest model F and iteratively trained to obtain the second trained random forest model T containing 220 decision trees. * , where the random forest model T * The training process is the same as the first training process of the random forest model in step (3); the random forest model is trained for the second time to obtain the random forest model T * The purpose is to train a new random forest model T based on the 15-dimensional features of the user's scheduling-related information. * , to predict the user experience rate, and compared with the first trained random forest model F * For example, it improves the efficiency of predicting user experience rate;
[0046] Step 6: Get the prediction result of user experience rate:
[0047] The test sample set X test As the second trained random forest model T * The input of each decision tree f h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X v ), the arithmetic average module performs weighted average calculation on the predicted experience rate of each user obtained by the 220 decision trees to obtain the predicted experience rate of V users
[0048]
[0049] The following is a further description of the technical effects of the present invention in conjunction with the simulation results:
[0050] 1. Experimental conditions and contents:
[0051] The hardware platform of the simulation experiment of the present invention is: the processor is AMD R7 4800H CPU, the main frequency is 2.9G, and the memory is 16GB.
[0052] The software platform for the simulation experiment of the present invention is: Windows 10 operating system and Python 3.6.
[0053] The data set used in the simulation experiment of the present invention is the user's scheduling related information, and the user's scheduling related information includes UE power margin, RSRP of uplink SRS, mobility status of different users, RANK value and other scheduling related information.
[0054] 2. Experimental results analysis:
[0055] The user experience rate effect predicted by the present invention is shown in Table 1.
[0056] The present invention Test set accuracy Prediction time (unit: s) The first trained random forest model (89-dimensional features) 86.63％ 0.54s The second trained random forest model (15-dimensional features) 86.90％ 0.14s
[0057] In conjunction with Table 1, the present invention selects 15-dimensional features of scheduling-related information of each user to form a feature set, thereby improving the efficiency of predicting user experience rate while ensuring that the accuracy of the test set is not reduced.
Claims
1. A method for predicting user experience rate of a wireless communication network, characterized in that: The steps include: (1) Obtain training sample set and test sample set: (1a) Obtain the scheduling-related information S of M users after eliminating missing values and outliers v1 ={X (m) |1≤m≤M}, and according to the traffic volume of each user tr m and business duration tt m Calculate the user experience rate tv for each user m , where M>3000,X (m) Indicates that the mth user contains scheduling related information with Q-dimensional features, Q>20; (1b) Through each user's experience rate tv m The corresponding scheduling related information X (m) Mark and randomly select S v1 The scheduling related information of U users and their corresponding labels constitute the training sample set X train , S v1 The remaining V users’ scheduling related information constitutes the test sample set X test ,in, M=U+V; (2) Constructing a random forest model: Construct a random forest model F consisting of H decision trees arranged in parallel and an arithmetic mean module connected sequentially to all decision trees. Each decision tree f h The decision tree f contains a root node, multiple split nodes connected to the root node, and multiple leaf nodes connected to each split node. The connection between the root node and each split node, and the connection between each split node and each leaf node are h directed edges, where H ≥ 50, f h is the hth decision tree; (3) Perform the first training of the random forest model: (3a) Initialize the random forest model to contain the maximum number of decision trees H ^ , H ^ ≥1000,H=50; (3b) The training sample set X train As the input of the random forest model F, each decision tree f h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X u ), the arithmetic average module performs weighted average calculation on the predicted experience rate of each user obtained by the H decision trees to obtain the predicted experience rate of U users (3c) adopts the mean square error loss function and passes and User Experience Rate Marker Calculate the loss value L of the random forest model F MSE , and for L MSE and H for preservation; (3d) Determine H≥H ^ Is it true? If so, for the saved L MSE and H with H as the horizontal axis, L MSE Perform drawing operation for the ordinate and select the ordinate loss value L according to the drawing result MSE The horizontal coordinate H corresponding to the minimum value of the first trained random forest model F containing H decision trees is obtained. * , otherwise, let H = H + 1 and execute step (3b); (4) Screening characteristics: Using the first trained random forest model F * , the importance of the Q-dimensional features of the user's scheduling-related information is evaluated by a feature evaluation method based on impurity reduction, and then the g-dimensional features of the scheduling-related information of each user whose importance is greater than a preset threshold t are selected to form a feature set B; (5) Perform a second training on the random forest model: (5a) Initialize the random forest model with the maximum number of decision trees H ^ , H ^ ≥1000, and let H1=50, where H1 is the number of decision trees included in the second training of the random forest model; (5b) Select training sample set X train The scheduling related information of U users with g-dimensional features of feature set B and their corresponding labels form a new training sample set D train , and D train It is used as the input of the random forest model F and iteratively trained to obtain the second trained random forest model T containing H1 decision trees. * ; (6) Obtain the prediction result of user experience rate: The test sample set X test As the second trained random forest model T * The input of each decision tree f h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X v ), the arithmetic average module performs weighted average calculation on the predicted experience rate of each user obtained by H1 decision trees to obtain the predicted experience rate of V users 2. The method for predicting user experience rate of a wireless communication network according to claim 1, characterized in that: The traffic volume tr of each user described in step (1a) m and business duration tt m Calculate the user experience rate tv for each user m , the calculation formula is:
3. The method for predicting user experience rate of a wireless communication network according to claim 1, characterized in that: Each decision tree f described in step (3b) h Make decisions on the scheduling related information of each user and get f h The corresponding experience rate prediction value f for each user h (X u ), the implementation steps are: (3b1) From the training sample set X train u training samples are randomly selected with replacement, and q-dimensional features are randomly selected from the Q-dimensional features of the user's scheduling-related information, where q<Q; (3b2) Traverse the q-dimensional features and solve the (j, s) corresponding to the minimum value of the loss function L(j, s) as the optimal segmentation feature j and segmentation point s, and let q = q-1. The expression of the loss function L(j, s) is: Among them, x i is the u values corresponding to feature j, y i is the experienced rate of u users, 1≤i≤u, R1(j,s) is x i <s training sample space, R2(j,s) is x i >s training sample space, c1 is x i <s corresponding to y i The average value of the value, c2 is x i >s corresponding to y i The average of the values; (3b3) Determine whether q=0 is true. If so, we get the decision tree f h , otherwise, execute step (3b2); (3b4) The scheduling related information of each user is calculated according to the decision tree f h The segmentation features and segmentation points are judged one by one from top to bottom until the leaf node is reached, and the x contained in the training sample space of the leaf node is calculated. i The corresponding y i The average of the experienced rate prediction value f of each user is obtained h (X u ).
4. The method for predicting user experience rate of a wireless communication network according to claim 1, characterized in that: The arithmetic average module described in step (3b) performs weighted average calculation on the predicted experience rate of each user obtained by the H decision trees to obtain the predicted experience rate of U users. The calculation formula is:
5. The method for predicting user experience rate of a wireless communication network according to claim 1, characterized in that: Step (3c) described by and User Experience Rate Marker Calculate the loss value L of the random forest model F MSE , the calculation formula is: