A mobile video freeze prediction model training method and prediction method

Abstract

The present invention provides a mobile video freeze prediction model training method, comprising: S1, obtaining a plurality of historical mobile video data with time series and known freeze conditions to form a data set; S2, using the data set obtained in step S1 to train the initial model in the following manner until convergence: S21, dividing the data set into a training set and a test set; S22, dividing the data in the training set obtained from the historical mobile video data into a plurality of training units according to autocorrelation; S23, respectively using each training unit to train the model until the model prediction accuracy is greater than or equal to a preset accuracy threshold or the training reaches a preset number of iterations to obtain a plurality of unit models, wherein when the training unit is used to train an LSTM network, a dropout function is used to correct weight update; S24, taking the proportion of the prediction accuracy of each unit model on the test set in the sum of all prediction accuracies as its corresponding weight, and merging all the unit models to obtain a final mobile video freeze prediction model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to a mobile video transmission freeze prediction and corresponding resource allocation technology, and more specifically, to a mobile video freeze prediction model training method and a prediction method. Background Art

[0002] With the rise of online video streaming services on the Internet, the influx of many short video content creators has led to the development of the mobile short video market in a refined and vertical direction. The mobile video user base has become increasingly diverse, and mobile video services have gradually become the mainstream service for network users. More and more video content is broadcast to the public on different platforms, and more and more users are involved in the creation and browsing of mobile videos. Ensuring the smoothness of mobile videos is the basic capability of the network, and smooth mobile video playback can greatly improve user experience. Therefore, it is extremely important to understand and predict the QoE (Quality of Experience) of mobile video users. Accurately predicting the QoE problems of mobile video users can help service providers take timely response strategies (for example, allocating bandwidth resources). At the same time, it is also of great significance to improve the QoE of mobile video users and enhance the loyalty of mobile video users.

[0003] Ultra-high-definition mobile video services have large bandwidth requirements and high real-time requirements. Traditional scheduling mechanisms lack perception of mobile video service quality and are difficult to meet the needs of smooth mobile video playback. They are prone to playback jams, and jams are the main factor affecting the user's mobile video service experience. Real-time prediction of mobile video jams can help base stations perceive user service status and adjust user wireless resource scheduling strategies in a timely manner, effectively improving the user experience of mobile video services. However, the user jam state is a service layer indicator and cannot be directly measured and fed back through the network. If the mobile video jams at various time and space are tested manually, a lot of human and material resources will be wasted. Therefore, it is necessary to use AI (Artificial Intelligence) technology to indirectly predict the jam information of mobile video services through network status information, and finally achieve cross-layer intelligent optimization of user experience. As a result, mobile video jam prediction has become a current research hotspot. At present, most researchers are trying to use machine learning and deep learning methods to solve this problem, but accurate prediction is still challenging because mobile video jams are affected by many factors and change over time.

[0004] The prediction results of the mobile video freeze are divided into mobile video freeze and mobile video smooth playback. Theoretically, the applicable algorithms include SVM algorithm (Support Vector Machine), RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), etc. Among them, the SVM algorithm is also a classification algorithm. Through the classification hyperplane, the data is separated on both sides of the plane to achieve the purpose of classification. For the prediction problem of mobile video freeze, SVM can fuzzily classify the relevant feature values ​​of mobile video playback, but because the classification of mobile video freeze is more complicated, the feature values ​​are related, and the results of the previous moment will also affect the current moment, the SVM algorithm is used directly for classification, and the prediction accuracy is low, which is not suitable for the classification of this complex problem. Recurrent neural network RNN ​​is mainly used for processing and prediction of time series data. It is a dynamic neural network suitable for processing and predicting time series data. It has a memory function and can record the influence of previous neurons on current neurons to realize the feedback of network status. However, when processing long-term dependencies, gradient explosion or gradient disappearance may occur. In order to avoid the drawbacks of RNN, LSTM was introduced. In LSTM, the structure of "gate" is a specific form of RNN. Compared with the traditional recurrent neural network, LSTM adds forget gate, input gate and output gate, which can better control the information in the cell, solve the long-term dependence problem on long sequences in RNN, avoid gradient disappearance, and have a good prediction effect on the prediction of data associated with time series. The prediction of mobile video freeze is a prediction problem related to time series. Therefore, LSTM network has good applicability for the prediction of mobile video freeze. However, for the prediction of mobile video freeze, if the traditional LSTM network is directly used, it will also cause overfitting, insufficient prediction accuracy and other problems, affecting the subsequent resource allocation and other processes. Therefore, solving the problems of overfitting and insufficient accuracy of the traditional LSTM network is of great significance for the prediction of mobile video freeze and the corresponding resource allocation. Summary of the invention

[0005] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a mobile video freeze prediction model training method and a prediction method.

[0006] According to a first aspect of the present invention, a method for training a mobile video freeze prediction model is provided, wherein the model comprises an LSTM network, wherein a random deactivation function layer is introduced into the LSTM network, and the method comprises: S1, obtaining a plurality of historical mobile video data with time series of known freeze conditions to form a data set, wherein each historical mobile video data corresponds to a time node, and the known freeze condition refers to a time node at which freeze occurs in the known historical mobile video data; S2, using the data set obtained in step S1 to train the initial model until convergence in the following manner: S21, dividing the data set into a training set and a test set; S22, dividing the data in the training set obtained from the historical mobile video data into a plurality of training units according to autocorrelation; S23, respectively using each training unit to train the model until the model prediction accuracy is greater than or equal to a preset accuracy threshold or the training reaches a preset number of iterations to obtain a plurality of unit models, wherein when the LSTM network is trained using the training unit, a dropout function is used to correct the weight update; S24, taking the proportion of the prediction accuracy of each unit model on the test set in the sum of all prediction accuracies as its corresponding weight, and merging all unit models to obtain a final mobile video freeze prediction model.

[0007] Preferably, in step S21, the acquired historical mobile video data is subjected to standardized preprocessing so that the data satisfies the Gaussian distribution and useless data is removed before the node data in the time series is divided into a training set and a test set. In some embodiments of the present invention, deviation standardization is adopted to perform linear changes on the original historical mobile video data to achieve standardization and normalization of the original data.

[0008] Preferably, step S2 also includes: S25, using a test set to test the prediction accuracy of the final mobile video freeze prediction model, and performing the following operations when its prediction accuracy is less than a preset accuracy threshold: S251, randomly selecting a training unit to train the final mobile video freeze prediction model until the prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations; S252, with a step size of 1, adding a training unit that did not participate in the last training in turn and merging it with all current training units to train the final mobile video freeze prediction model after the last training until the prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations.

[0009] In some embodiments of the present invention, step S2 also includes: S26, correcting the prediction results of the final mobile video freeze prediction model by adjusting the model parameters to correct the prediction results of the inconsistent points with the prediction results of two or more consecutive time nodes before and after to the prediction results of two or more consecutive time points before and after.

[0010] In some embodiments of the present invention, in step S22, the autocorrelation of each mobile video data in the training set is calculated in the following manner and the data with an autocorrelation function value greater than 0.2 are divided into the same unit:

[0011]

[0012] Among them, R(x i , x i+h ) represents x i and x i+h The autocorrelation function, h represents the time interval between two moments of data, x i represents the data of the i-th time node, u 1 represents the mean of the previous set of data, u 2 It represents the mean of the data separated by h in time from the previous set of data, and n represents the amount of autocorrelation data currently calculated.

[0013] Preferably, the preset accuracy threshold is 80%.

[0014] Preferably, the preset number of rounds is 1000.

[0015] According to a second aspect of the present invention, there is provided a method for resource allocation based on prediction of mobile video freezes, the method comprising: P1, using a model trained by the method described in the first aspect of the present invention to predict a time node at which a mobile video freezes; P2, based on the time node at which a mobile video freezes obtained in step P1, allocating resources to the node so that the mobile video does not freeze.

[0016] Compared with the prior art, the present invention has the advantages that: the present invention can effectively predict the time period of mobile video freeze in the network, and provide a basis for allocating and regulating resources before the time node to improve the experience quality of mobile video users. The present invention can predict at which time more resources need to be allocated to users and at which time too many resources are not needed, and more accurately schedule the allocation of resources, effectively improving the quality of user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0018] Figure 1 Schematic diagram of the LSTM network structure according to an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of the structure of a mobile video freeze prediction model according to an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of the principle of a random deactivation function according to an embodiment of the present invention;

[0021] Figure 4 1. A schematic diagram of a flow chart of a method for training a mobile video freeze prediction model according to an embodiment of the present invention;

[0022] Figure 5 A schematic diagram showing changes in the accuracy of the LSTM network prediction according to an embodiment of the present invention and the accuracy of the original LSTM network prediction as a function of the number of iterations;

[0023] Figure 6 This is a schematic diagram of the change of the loss value predicted by the LSTM network according to an embodiment of the present invention and the loss value predicted by the original LSTM network with the number of iterations. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail by specific embodiments below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0025] In order to solve the problems of overfitting and insufficient accuracy of traditional LSTM networks, the present invention provides a method for predicting mobile video freezes using an improved LSTM network to improve prediction accuracy, which includes: establishing a prediction model for mobile video, training the model to improve prediction accuracy, performing prediction correction on the prediction results, and using the trained model to predict mobile video freezes. Among them, to establish a mobile video prediction model and a training model, it is necessary to perform model prediction and training based on monitoring data such as the resource allocation of the network to be predicted. The model is trained by grouping the monitored data according to autocorrelation, and the principle of optimizing the local first and then optimizing the whole is adopted to improve the prediction accuracy and reduce the time cost of prediction. Dropout (random deactivation function) is introduced to prevent model overfitting and improve the generalization ability of the model. The accuracy of the prediction results of the prediction model for sample data is calculated, and the accuracy is used as an evaluation indicator. If the accuracy reaches the expected threshold, the model is output to complete the model training, and the trained model is used to predict the data to be predicted. The next step of prediction correction is required for the prediction result, and the corrected result is output as the final prediction result. This method can provide a basis for subsequent resource regulation, prevent resource waste and insufficient resource allocation, and save a lot of manpower and material resources.

[0026] In order to better understand the present invention, the basic principles of the traditional LSTM network are first introduced.

[0027] like Figure 1As shown in the figure, the traditional LSTM network includes: input gate, forget gate, internal storage unit, output gate. The data of each time node passes through the input gate, forget gate, internal storage unit, and output gate to output the final prediction result.

[0028] Among them, the LSTM network determines the data to be retained and the data to be forgotten at this time node through the input data of this time node and the output data of the previous time node. The calculation formula of the forget gate is formula 1, which needs to be calculated through the input data x of this time node t and input data weight W f , the hidden layer output h of the previous time node t-1 , hidden layer weight U f And the bias value b f Three quantities are used to determine the data f to be forgotten at this time node. t :

[0029] f t =σ(W f x t +U f h t-1 +b f ) (1)

[0030] Formula 1 is the sigmoid function. The calculation result obtained according to formula 1 is between [0, 1]. The output of the forget gate is a real number vector between 0 and 1. When the forget gate output is 0, any vector multiplied with it will get a 0 vector, which is equivalent to nothing can pass through, indicating that the data passing through the forget gate is 0, and all data is retained; when the output is 1, any vector multiplied with it will not change anything, which is equivalent to everything can pass through. The data passing through the forget gate is all the data in the storage unit. According to the data vector of the storage unit and f t The data to be forgotten is determined by multiplying

[0031] The calculation formula of the input gate is formula 2, which requires input data x according to the input gate of this time node t Input data weight W i , the hidden layer output h of the previous time node t-1 , hidden layer weight U i And the bias value b i To determine the data i that will continue to participate in the prediction at this time node t :

[0032] i t =σ(W i x t +U i h t-1 +b i ) (2)

[0033] According to the calculation result of formula 2, when the input gate output is 0, any vector multiplied by it will get a 0 vector, which is equivalent to nothing can pass through, indicating that the input data at the current moment has 0 influence on the model and no input is performed; when the output is 1, any vector multiplied by it will not have any change, which is equivalent to everything can pass through, and the data input at the current moment is the original data that directly enters the model and participates in the prediction. t The product of determines the data to be involved in the prediction, that is, the data stored in the storage unit.

[0034] The calculation formula of the internal storage unit is formula 3 and formula 4, which needs to be based on the input gate input data x at this time node t Input data weight W c , the hidden layer output h of the previous time node t-1 And the hidden layer weight U c , the forget gate f at this time node t The output result and the internal storage unit data c of the previous time node t-1 , the value of the input gate at this time node i t , to determine the data c that needs to be retained in the storage unit at this time node t :

[0035] c′ t =Tanh(W c x t +U c h t-1 ) (3)

[0036] c t =f t c t-1 +i t c′ t (4)

[0037] The output gate calculation formula is Formula 5, which needs to be based on the input data x at this time node t Input data weight W o , the hidden layer output h of the previous time node t-1 And the hidden layer weight U o , bias value b o , three quantities are used to determine the final output result o of this time node t :

[0038] 0 t =σ(W o x t +U 0 h t-1 +b 0 ) (5)

[0039] When the output of the output gate is 0, any vector multiplied by it will result in a 0 vector, which is equivalent to nothing passing through, indicating that the output at the current moment is 0; when the output of the output gate is 1, any vector multiplied by it will not change anything, which is equivalent to everything passing through, and the original data in the storage unit is output. According to the calculation result of formula 5, the output of this time node is determined, and the output of this time node is used as the hidden layer output of the next time node.

[0040] As can be seen from the above introduction, the LSTM network is very suitable for predicting data associated with time series. However, as described in the background technology, if the traditional LSTM network is directly used for predicting the freezing of mobile videos, problems such as overfitting and insufficient prediction accuracy will occur. Therefore, the inventor proposes an improved LSTM network and trains it with historical mobile video data with time series of known freezing conditions to obtain a usable prediction model. The following introduces the present invention from three aspects: the structure of the mobile video freezing model, the training of the model, and the application of the model.

[0041] 1. Model Structure

[0042] The present invention improves the LSTM network to make it more accurate in predicting the freezing of mobile videos. At the same time, it can solve the problem of gradient disappearance, effectively prevent overfitting, and reduce the time and space cost of model training. According to one embodiment of the present invention, Figure 2 As shown in FIG. 1 , the present invention improves the traditional LSTM network to obtain a model more suitable for mobile video freeze prediction, and the model includes a sequence input layer, an LSTM network, a random deactivation function layer (Dropout), a fully connected layer, a softmax layer, and a classification layer connected in sequence. The present invention improves the LSTM network and introduces Dropout in the entire model, such as Figure 3 As shown in the figure, the core idea of ​​Dropout is to change the activation state of neurons during the training process of the model to reduce the dependence of the final prediction results on certain local neurons and time nodes, prevent overfitting, and improve the generalization ability of the model. From a macro perspective, Dropout is outside the LSTM network. From a micro perspective, it acts on the weight parameters in the LSTM during the training process, randomly making some weights not updated at the current time node, so as to reduce the dependence of the final prediction results on certain local neurons and time nodes, prevent overfitting, and improve the generalization ability of the model.

[0043] 2. Model Training

[0044] Simply put, the model training process consists of two stages.

[0045] Phase 1:

[0046] First, feature value selection is performed on mobile video data with known jamming conditions, and the data is normalized and standardized, useless data is removed, and the data is divided into two parts: training data and test data. Then, the autocorrelation of the data is calculated for the training data, and the moment when the autocorrelation function value meets the largest range of conditional data is used as the last data of the group of training data grouping, and the training data is divided into multiple training units according to this classification standard. Finally, for each training unit, the data of this unit is used to import LS-imp (improved LSTM network) in turn to train the model, and the weight values ​​of the input gate, forget gate, storage unit, and output gate are obtained. According to the results obtained from the training of each unit, the test data is input into the model for prediction, and the accuracy of the obtained prediction results is calculated. After the prediction of multiple units is completed, the accuracy of the prediction results of each unit is used as the proportion of each unit model when merging, and the models required for prediction are merged into the model required for the final prediction. If the model has met the preset accuracy requirements, the training can be ended. If the model does not meet the preset accuracy requirements, the next step of model training is required.

[0047] Phase 2:

[0048] According to the principle of optimizing the local first and then optimizing the whole, the training results of the previous stage are first used to train the model obtained in the previous stage with the data of a training unit as training data. After the model reaches the predetermined accuracy, the data of a training unit is added and combined with the previous training unit to train the model together, and so on, until the data of the last unit is added to the training, and the whole is optimized through step-by-step optimization. After the last unit data is added, the model is trained with the overall data until the model meets the predetermined accuracy requirements or reaches the preset number of iterations. Among them, the preset accuracy requirement refers to the prediction accuracy greater than or equal to the preset accuracy threshold. According to one embodiment of the present invention, the preset accuracy threshold is 80%. According to one embodiment of the present invention, the preset number of iterations is 1000.

[0049] For the final output model, after predicting the test data, the predicted results are corrected. The correction is mainly performed on isolated state points, that is, time points that are different from the results predicted by two or more consecutive time points before and after. According to the probability of such a situation occurring in the training data, corresponding operations are performed on the predicted results to improve the accuracy of the predicted results, such as by adjusting the model parameters to change the predicted result of this time point to the same output as the predicted results of the previous and next time points.

[0050] The model training process of the present invention is described in detail below with reference to specific embodiments.

[0051] like Figure 4As shown, the training process includes the following steps:

[0052] Step 1: Data processing

[0053] For historical mobile video data with known freezing conditions, the data is read through the sequence input layer and the original data is linearly transformed using the deviation normalization method to standardize and normalize the data. That is, the maximum and minimum values ​​of each feature dimension are scaled proportionally to between 0 and 1. This can make the new data closer to the standard Gaussian distribution, reduce the pressure on neural network training, and save training time. After standardization, useless data related to the entire system is eliminated, and part of the remaining data is used as the training set data to train the model, and the other part of the data is used as the test set data to test the accuracy of the model.

[0054] For the training data, the autocorrelation of the training data is calculated. When a certain sampling point of the time series and the i-th sampling point with autocorrelation before it are used to train the model together, a model that meets the accuracy requirements can be trained more quickly. The model is divided into a training unit for training the model. The autocorrelation function R(x) is calculated according to Formula 6. A value close to 1 indicates positive correlation, a value close to -1 indicates negative correlation, and a value close to 0 indicates very low correlation. According to the functional relationship of autocorrelation, the data with an autocorrelation function > 0.2 is determined as a training unit:

[0055]

[0056] Among them, R(x i , x i+h ) represents x i and x i+h The autocorrelation function, h represents the time interval between two moments of data, x i represents the data of the i-th time node, u 1 Represents the mean of the previous set of data, the so-called previous set of data refers to x 1 to x n-h The data, u 2 represents the mean of the data separated by h in time from the previous set of data, and n represents the amount of data of the autocorrelation function currently calculated. For example, when n is 100, the calculated value is x 100 With x 100-hThe present invention first takes a value for the data volume. It should be noted that it is meaningless when the data volume is taken too large. For example, the correlation between the first data and the 10,000th data must be very small. Assuming that the value of the data volume is 600, the autocorrelation function with a data interval of 1 to the autocorrelation function value with a data interval of 599 is calculated in sequence. In the present invention, the size of the group is determined by calculating the autocorrelation function for the training data and calculating the autocorrelation of the data. The size of the autocorrelation function fluctuates for different data volumes. The moment when the autocorrelation function value meets the condition (the condition of the present invention is set as the autocorrelation function>0.2) and the data range is the largest is taken as the last data of this group of the training data grouping, and the data is grouped according to this classification standard. For example, the autocorrelation function value with a data interval of 20 is equal to the autocorrelation function value with a data interval of 50, and the subsequent autocorrelation function values ​​are not greater than the autocorrelation function value with a data interval of 50, then the size of the data grouping can be set to 50. Assume that the training set is divided into c training units using the above method for subsequent model training.

[0057] Step 2: Model training

[0058] Step 21: Use each training unit in the divided training set to train the initial model (expressed as LS-imp), such as Figure 4As shown, c training units (respectively represented as unit 1, unit 2, ... unit c) respectively train the initial model (equivalent to each training unit corresponding to an initial model, respectively represented as LS-imp1, LS-imp2, ... LS-impc) until the model prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations, and multiple unit models (respectively represented as model 1, model 2, ... model c) are obtained. Among them, based on the calculation method of each threshold value of the LSTM network, the initial calculation is performed using the initialized weights and the data of the corresponding training unit. The data passes through the input gate, the forgetting gate, the internal storage unit, and the output gate, and then the weight is updated. Since directly using the LSTM network to train the data will produce overfitting, in order to avoid the occurrence of overfitting and improve the generalization ability of the model, the present invention introduces the Dropout function to correct the weight update. Dropout can randomly shield some neurons, and then the gradient descent method is used to update the weights of each threshold to find the threshold weight value that can make the prediction accuracy the highest. Specifically, Dropout can inactivate some neurons, that is, when the weights are updated during the training model, some weights are randomly inactivated at this time point and not updated; for the weights that need to be updated, the gradient descent method is used to update; finally, the prediction data is predicted to check whether the required prediction accuracy is achieved. If the accuracy requirement is met, the iteration is stopped, and if it is not met, the training is continued. When the prediction accuracy of the unit model is greater than or equal to the preset accuracy threshold or the number of training times reaches the upper limit, the update is stopped, and the final model parameters are obtained. The accuracy of the unit model is calculated by inputting test data.

[0059] Step 22: Use the prediction accuracy of the unit model as the weight of the model obtained by training each unit data, merge the models trained by each unit data according to the accuracy, so as to balance the errors generated by the model, prevent overfitting, and at the same time, avoid the influence of the models trained by some special eigenvalues ​​on the whole, and eliminate the dependence of the model on some special eigenvalues. After the model is merged, check the merged model and use the merged model to predict the accuracy. If the prediction accuracy meets the requirements, stop training and proceed to step 3. If the accuracy does not meet the requirements, continue training according to step 23. According to an example of the present invention, it is assumed that the accuracy of model 1, model 2, ... model c obtained after each training unit trains the initial model on the test set is z 1 、z 2 ,…z c , then the weight of model 1 is Weights for Model 2 And so on, the combined model = model1*w 1 +Model2*w2 +…+Model c*w c The merging of models is actually the merging of model array eigenvalues, which is a common operation in this field, and the specific model merging process will not be described here.

[0060] Step 23: According to the result of data grouping in step 1, based on the principle of optimizing the local first and then optimizing the whole, since less data is used to train the model first, the model can be applied to a small part of the data, and then the training data is gradually increased to make the model gradually applicable to more data. The model training according to this method can reduce the complexity of model training and reduce the time cost of training the model. According to an example of the present invention, the merged model is first trained with the data of unit 1 until the model prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations; then the unit 2 data is added, and the unit 1 data and the unit 2 data are combined to train the model trained by unit 1, until the model prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations; then the unit 3 data is added for training, and the unit 1 data, unit 2 data, and unit 3 data are combined to continue training the model, and so on, until all data have been added to the model training process, and the complete data block is used for training until the model prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations, and the model training is terminated.

[0061] Step 3: Correct the prediction results

[0062] The test set data divided in advance is input into the final model obtained by training in step 2 for prediction, and the prediction result is corrected. The correction is mainly performed on isolated state points. By correcting the prediction result, the accuracy of the prediction result can be further improved. Specifically, the correction principle is that the prediction results of two or more consecutive time points are different from the prediction results of the current time point. According to the probability of such a situation in the training data, the prediction result is corrected accordingly. The model parameters are adjusted to change the prediction result of this time point to the same output as the prediction results of the previous and next time points. After the correction is completed, the model prediction result is compared with the true value to evaluate the prediction accuracy. If the expected effect is achieved, for example, the predetermined accuracy threshold is reached, the corrected model is output as the final model, and this model is used to participate in the mobile video freeze prediction work after the network, providing a basis for improving the QoS of mobile video users.

[0063] In order to verify the effect of the present invention, the improved model of the present invention and the traditional LSTM model are used to predict the same network. The changes of the network prediction accuracy of the two with the number of iterations are shown in the following figure. Figure 5As shown in FIG. 1 , the accuracy of the final prediction using the original LSTM network is 63.18%, while the accuracy of the final prediction using the improved LSTM network of the present invention is 83.05%, which is about 20% higher than the accuracy of the prediction. Figure 5 It can be seen that when the number of iterations reaches 1000, the prediction accuracy tends to be stable, and the prediction accuracy of the model obtained by the improved LSTM network training is always higher than the prediction accuracy of the model obtained by the original LSTM network training. Figure 6 It is an image showing the change of the loss function predicted by the improved LSTM network of the present invention and the original LSTM network with the number of iterations. It can be seen from the image that the loss function predicted by the improved LSTM network of the present invention is much smaller than the loss function predicted by the original LSTM network. As the number of iterations increases, the loss functions predicted by the models trained by the two networks gradually decrease, and tend to be stable when the number of iterations is about 1000. It can be seen that the improved LSTM network of the present invention is feasible in predicting the situation of mobile video freeze, and can effectively improve the accuracy of prediction, prevent model overfitting, and improve the generalization ability of the model.

[0064] The prediction model obtained by the present invention can effectively predict the time period of mobile video freeze in the network, and provide a basis for allocating and regulating resources before the time node to improve the experience quality of mobile video users. At different times, users have different needs for the network. At certain times, there are fewer users online, so naturally fewer resources can meet the user needs. However, at certain times, there are more users online, so more resources are needed to meet the needs of users and improve the user experience quality. However, if the allocated resources are constant, when fewer users use the network, there will be a waste of resources. The present invention can predict at which times more resources need to be allocated to users and at which times too many resources do not need to be allocated, and more accurately schedule the allocation of resources to effectively improve the user experience quality.

[0065] Although manual inspection is directly used for evaluation and tuning, the accuracy is higher. However, since it involves the analysis of user experience quality in multiple locations, it will require a lot of manpower and material resources, resulting in unnecessary waste. If this method is used for evaluation and prediction, it will bring more convenience and save manpower and material resources.

[0066] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.

[0067] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0068] Computer readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. Computer readable storage medium can include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.

[0069] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a mobile video freeze prediction model, the model comprising an LSTM network, wherein a random deactivation function layer is introduced into the LSTM network, characterized in that: The method comprises: S1. Acquire a data set composed of a plurality of historical mobile video data with time series and known jamming conditions, wherein each historical mobile video data corresponds to a time node, and the known jamming condition refers to a time node at which jamming occurs in the known historical mobile video data; S2. Use the data set obtained in step S1 to train the initial model until convergence in the following manner: S21, divide the data set into a training set and a test set; S22, dividing the data in the training set obtained from the historical mobile video data into a plurality of training units according to autocorrelation; S23, respectively using each training unit to train the model until the model prediction accuracy is greater than or equal to a preset accuracy threshold or the training reaches a preset number of iterations to obtain multiple unit models, wherein when the LSTM network is trained using the training unit, a dropout function is used to correct the weight update; S24, taking the proportion of the prediction accuracy of each unit model on the test set in the sum of all prediction accuracies as its corresponding weight, merging all unit models to obtain a final mobile video freeze prediction model; S25. Use the test set to test the prediction accuracy of the final mobile video freeze prediction model, and perform the following operations when the prediction accuracy is less than a preset accuracy threshold: S251, randomly selecting a training unit to train the final mobile video freeze prediction model until the prediction accuracy is greater than or equal to a preset accuracy threshold or the training reaches a preset number of iterations; S252. With a step size of 1, add one training unit that did not participate in the last training in turn and merge it with all current training units to train the final mobile video freeze prediction model after the last training until the prediction accuracy is greater than or equal to the preset accuracy threshold or the training reaches a preset number of iterations; S26. Correct the prediction results of the final mobile video freeze prediction model by adjusting the model parameters to correct the prediction results of the points that are inconsistent with the prediction results of two or more consecutive time nodes before and after to the prediction results of two or more consecutive time points before and after.

2. The method according to claim 1, characterized in that In step S21, the acquired historical mobile video data is subjected to standardization preprocessing to make the data satisfy Gaussian distribution and after removing useless data, the node data in the time series is divided into a training set and a test set.

3. The method according to claim 2, characterized in that In step S21, the acquired historical mobile video data is processed as follows: Deviation normalization is adopted to perform linear changes on the original historical mobile video data to achieve standardization and normalization of the original data.

4. The method according to claim 2, characterized in that: In step S22, the autocorrelation of each mobile video data in the training set is calculated in the following manner and the data with an autocorrelation function value greater than 0.2 are divided into the same unit: Among them, R(x i , x i+ h) represents x i and x i+ The autocorrelation function of h, h represents the time interval between two moments of data, x i represents the data at the i-th time node, u1 represents the mean of the previous set of data, u2 represents the mean of the data that is h apart in time from the previous set of data, and n represents the amount of autocorrelation data currently calculated.

5. The method according to claim 2, characterized in that: The preset accuracy threshold is 80%.

6. The method according to claim 2, characterized in that The default number of rounds is 1000.

7. A resource allocation method based on mobile video freeze prediction, characterized in that: The method comprises: P1. Predicting the time point when a mobile video freezes by using a model trained by the method described in any one of claims 1 to 6; P2. Based on the time node at which the mobile video freezes obtained in step P1, allocate resources to the node so that the mobile video does not freeze.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of any method described in claims 1-6 and 7.

9. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, enables the electronic device to implement the steps of the method as described in any one of claims 1-6 and 7.

Images