Network quality similarity determination method and apparatus, device, medium, product

Abstract

The application relates to a network quality similarity determination method and device, equipment, medium and product, the method comprising: acquiring time domain quality index data corresponding to a plurality of users, including quality index values obtained by discrete sampling along the time domain; transforming each time domain quality index data from the time domain to the frequency domain, converting each quality index value into the amplitude of a data point in frequency spectrum data, and obtaining the frequency spectrum data; extracting comprehensive feature information of the frequency spectrum data corresponding to each user, wherein the comprehensive feature information contains a plurality of point-to-point feature information corresponding to each target data point in the frequency spectrum data, and the point-to-point feature information contains position feature information between a single target data point and the neighborhood data points of the target data point; and determining the feature similarity between two users according to any two comprehensive feature information. The application can accurately measure the similarity between different network conditions according to the frequency spectrum data of the time domain quality index data corresponding to different network conditions.

Description

[0001] Technology Neighborhood

[0002] This application relates to the field of network optimization technology, and in particular to a method for determining network quality similarity and the corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0003] In audio and video co-op and live streaming scenarios, network signals that characterize network transmission quality, such as bandwidth, network latency, and packet loss rate, are often collected from users. Then, the network conditions are simulated offline based on the collected network signals. The audio and video transmission algorithms are then optimized and improved for these network conditions to enhance the audio and video quality under such conditions.

[0004] To simulate these types of network conditions, it is first necessary to classify them. The efficiency of optimization can then be extracted through classification simulation. How to identify similar networks based on network signals is crucial to correctly classifying different network conditions, which in turn affects the effectiveness of network optimization.

[0005] In traditional technologies, network signal similarity determination mainly adopts the following schemes:

[0006] 1) Judgment based on time-domain statistical characteristics, such as mean and variance, is applicable when the variance and mean change little. If the variance and mean change much, and the network signal trends are similar, the statistical characteristics cannot be used to determine the signal. Online networks fall into this category, with severe jitter.

[0007] 2) Based on the cross-correlation coefficient, the cross-correlation coefficient can accurately measure the similarity range [-1,1]. The cross-correlation coefficient is performed in the time domain, and noise has a great impact on the results.

[0008] 2) Based on frequency domain analysis, the method is to first transform the time domain signal to the frequency domain and compare the similarity of the frequency domain amplitudes. For this method, there is a lack of corresponding accurate measurement methods. If only Euclidean distance is used to calculate the similarity of two amplitudes, the results obtained will differ greatly.

[0009] Therefore, it is necessary to explore new approaches to similarity judgment based on network signals in order to effectively classify complex and diverse network conditions and thus help improve network optimization results. Summary of the Invention

[0010] The purpose of this application is to solve the above-mentioned problems by providing a method for determining network quality similarity, and corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product.

[0011] To suit the various purposes of this application, the following technical solution is adopted:

[0012] In one aspect, to suit one of the purposes of this application, a method for determining network quality similarity is provided, comprising:

[0013] Acquire time-domain quality index data representing network transmission quality for multiple users, wherein the time-domain quality index data includes quality index values ​​collected discretely along the time domain;

[0014] The time-domain quality index data of each user is transformed from the time domain to the frequency domain, so that the values ​​of each quality index are converted into the amplitude of the data points in the spectrum data, and the spectrum data is obtained.

[0015] For each user's spectrum data, extract its comprehensive feature information. The comprehensive feature information includes multiple inter-point feature information corresponding to each target data point in the spectrum data. The inter-point feature information includes the positional feature information between a single target data point and its neighboring data points.

[0016] The similarity of network transmission quality between two users is determined based on their combined feature information.

[0017] On the other hand, to meet one of the purposes of this application, a network quality similarity determination device is provided, comprising: a data acquisition module, a time-frequency transformation module, a feature representation module, and a similarity determination module, wherein the data acquisition module is used to acquire time-domain quality index data representing network transmission quality corresponding to multiple users, the time-domain quality index data including quality index values ​​obtained by discrete acquisition along the time domain; the time-frequency transformation module is used to transform the time-domain quality index data of each user from the time domain to the frequency domain, so that each quality index value is converted into the amplitude of data points in the spectrum data, thereby obtaining the spectrum data; the feature representation module is used to extract comprehensive feature information corresponding to the spectrum data of each user, the comprehensive feature information including multiple inter-point feature information corresponding to each target data point in the spectrum data, the inter-point feature information including the positional feature information between a single target data point and its neighboring data points; the similarity determination module is used to determine the feature similarity between two users regarding network transmission quality based on the comprehensive feature information of any two users.

[0018] In another aspect, a computer device provided to suit one of the purposes of this application includes a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the network quality similarity determination method described in this application.

[0019] In another aspect, a computer-readable storage medium is provided to suit another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the network quality similarity determination method, which, when invoked by a computer, executes the steps included in the method.

[0020] In another aspect, a computer program product provided for another purpose of this application includes a computer program / instructions that, when executed by a processor, implement the steps of the network quality similarity determination method described in any embodiment of this application.

[0021] Compared with existing technologies, this application has several technical advantages, including but not limited to:

[0022] First, this application performs time-frequency transformation on the user's time-domain quality index data to obtain the corresponding frequency-domain spectrum data. Then, it extracts features from each data point in the spectrum data, using the positional feature information between two data points within a neighborhood to represent the features of each data point. Each data point is represented by its inter-point feature information through the positional feature information between it and its neighboring data points. Based on this, the inter-point feature information of each data point in the spectrum data is determined as the comprehensive feature information corresponding to that spectrum data, thus achieving an effective feature representation of each user's time-domain quality index data in the frequency domain mapping. This allows each spectrum data to be distinguished and compared through its comprehensive feature information. Therefore, the feature similarity can be quantitatively determined based on the pairwise comprehensive feature information, achieving an effective measurement of the similarity of the time-domain quality index data.

[0023] Secondly, when representing the spectral data, this application uses the positional feature information between data points. According to the characteristics of spectral data, the positional feature information is mainly composed of time-domain and frequency-domain information. Therefore, it does not need to rely on the amplitude of each data point for feature statistics. It is not sensitive to amplitude or noise, but only sensitive to frequency in the frequency domain. It can effectively measure the same network change trend. Compared with other methods, the positional feature information between two points on the time spectrum can provide a more accurate and interference-resistant feature representation method, which is superior to other methods.

[0024] Furthermore, based on the effective measurement of time-domain quality index data for each user and the determination of feature similarity between pairs of users, this application can be used to classify the network conditions of a large number of users, making the classification more accurate. Therefore, it is not difficult to understand how the technical solution of this application can be extended to application scenarios such as live streaming. By conducting corresponding simulation tests based on the similar network conditions obtained from the correct classification, network optimization can be subdivided into user groups corresponding to different categories of network conditions, so as to further match different optimal algorithms for different user groups, thereby achieving overall network optimization of the network transmission quality of live streaming services. Attached Figure Description

[0025] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0026] Figure 1 This is a flowchart illustrating one embodiment of the network quality similarity determination method of this application.

[0027] Figure 2 Here is an example of a spectrum diagram corresponding to the spectrum data of this application.

[0028] Figure 3 , Figure 4 , Figure 5 These are all examples of time-frequency diagrams obtained by abstracting the spectrum diagrams of this application, wherein Figure 4 Corresponding to Figure 3 A window showing the neighborhood range corresponding to the target data point. Figure 5 Location information is displayed.

[0029] Figure 6 This is a schematic diagram of the process for determining feature similarity in an embodiment of this application.

[0030] Figure 7 This is a schematic diagram of a dual-tower structure for an exemplary neural network model of this application.

[0031] Figure 8 This is a schematic diagram of the process of performing time-frequency transformation on time-domain quality index data in an embodiment of this application.

[0032] Figure 9 This illustration shows the windowing and framing process during time-frequency conversion in the embodiments of this application.

[0033] Figure 10 This is a flowchart illustrating the process of constructing comprehensive feature information of spectrum data in the implementation of this application.

[0034] Figure 11 This is a schematic diagram of the simulation test based on similar user clusters in an embodiment of this application.

[0035] Figure 12 This is a flowchart illustrating an exemplary clustering process in an embodiment of this application.

[0036] Figure 13 This is a schematic diagram of the network quality similarity determination device of this application.

[0037] Figure 14 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0038] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0039] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0040] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in various ways to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0041] This application discloses a method for determining network quality similarity, which can be programmed into a computer program and deployed on a client or server for implementation. Please refer to [link to relevant documentation]. Figure 1 The network quality similarity determination method of this application, in its typical embodiment, includes the following steps:

[0042] Step S1100: Obtain time-domain quality index data representing network transmission quality corresponding to multiple users, wherein the time-domain quality index data includes quality index values ​​obtained by discrete collection along the time domain.

[0043] An exemplary application scenario involves live streaming. Live streaming requires the transmission of large amounts of audio and video data, placing higher demands on network quality, specifically the network transmission quality between terminal devices and the server. Therefore, it necessitates timely optimization to adapt to the network conditions shaped by the varying network transmission quality of a massive number of users. To monitor the network transmission quality of each user, corresponding tracking code is typically embedded in the webpage or application providing the live streaming service. This code collects quality indicator values ​​representing network transmission quality, obtains corresponding data samples, and continuously uploads them to a data server for storage and retrieval. For the network conditions corresponding to a user's terminal device, the set of quality indicator values ​​collected discretely along the time domain constitutes the time-domain quality indicator data stored in the data server's database. Subsequently, the time-domain quality indicator data for each user can be directly retrieved from this database.

[0044] The time-domain quality indicator data can be assigned to a single dimension or to multiple dimensions. Each dimension's time-domain quality indicator data characterizes network transmission quality from that dimension. For example, it can include time-domain quality indicator data corresponding to any one or more dimensions such as bandwidth, network latency, and packet loss rate, used to respectively characterize changes in bandwidth, network latency, and packet loss rate under network conditions. Typically, the quality indicator values ​​corresponding to the above three dimensions can be collected each time. For ease of understanding, the following description will primarily focus on time-domain quality indicator data corresponding to a single dimension. Cases with time-domain quality indicator data corresponding to multiple dimensions will also be appropriately explained in the later embodiments. Here, the time-domain quality indicator data can be understood as time-domain quality indicator data corresponding to a single dimension.

[0045] The database can associate and store each user with its corresponding time-domain quality index data, so that the user can be identified through the time-domain quality index data.

[0046] Step S1200: Transform the time-domain quality index data of each user from the time domain to the frequency domain, so that the values ​​of each quality index are converted into the amplitude of the data points in the spectrum data, and obtain the spectrum data;

[0047] For each user's time-domain quality index data, taking a single-dimensional time-domain quality index as an example, time-frequency transformation can be applied to transform it to the frequency domain to obtain its corresponding spectrum data. For example, the spectrum data can be mapped to... Figure 2The spectrum diagram in the time-frequency coordinate system shown illustrates the correspondence between time, frequency (Hz), and amplitude. The time coordinate corresponds to the timestamp of the specific quality index value collected in the time-domain quality index data, and the frequency coordinate corresponds to the sampling frequency of the quality index data. Figure 2 The color depth in the spectrum represents the amplitude, which is actually the representation of each quality indicator value in the spectrum after time-frequency transformation. This allows for the acquisition of spectrum data for a single user's single-dimensional time-domain quality indicator data. Extending this logic, it's easy to understand that for situations with multiple dimensions of time-domain quality indicator data, time-frequency transformation can be performed separately for each dimension to obtain spectrum data for each user in each dimension.

[0048] The amplitude values ​​in the spectral data are actually discrete. Each discrete amplitude value can be considered as a corresponding data point. That is, after time-frequency transformation, the quality index values ​​in the time-domain quality index data are actually converted into the amplitude values ​​of the corresponding data points in the spectral data. Therefore, the spectral data, reflected in the spectrum diagram, is represented by a large number of data points discretely distributed in the time-frequency coordinate system. The abstracted effect is as follows: Figure 3 The example uses a time axis and a frequency axis to represent the positional relationship between individual data points (X).

[0049] Step S1300: Extract comprehensive feature information for each user's spectrum data. The comprehensive feature information includes multiple inter-point feature information corresponding to each target data point in the spectrum data. The inter-point feature information includes positional feature information between a single target data point and its neighboring data points.

[0050] It's easy to understand that within the same time-frequency coordinate system, any two data points contain both absolute and relative positional information, constituting the positional information between them. This positional information further implies variations in the time and frequency domains. Feature representation can be achieved by extracting features from these variations, and this feature representation is decorrelated with the amplitude. Therefore, feature representation can be applied to each spectrum data point for each user, yielding its corresponding comprehensive feature information.

[0051] For each spectral data point, its comprehensive characteristic information can be determined. For example... Figure 3As shown, each data point in the spectrum data can be considered a target data point. First, the inter-point feature information corresponding to each target data point is determined. Then, the set of inter-point feature information for all target data points constitutes the comprehensive feature information of the spectrum data. The inter-point feature information describes the positional information between the corresponding target data point and its neighboring data points. Specifically, this can be achieved by extracting features from the positional information between the target data point and each of its neighboring data points to represent the point-to-point positional information and determine the corresponding positional feature information. Finally, the set of all point-to-point positional feature information within the entire neighborhood of the target data point is used to construct the inter-point feature information corresponding to that target data point.

[0052] The neighborhood range of the target data point can be preset, such as... Figure 4 As shown, a window with a defined area is determined by expanding the coordinates of the target data point according to the corresponding time and frequency. All data points within this window are considered neighborhood data points of the target data point. The positional feature information between the target data point and each neighboring data point within this window constitutes the point-to-point positional feature information. The set of all point-to-point positional feature information originating from the target data point within the entire window constitutes the point-to-point feature information corresponding to the target data point.

[0053] In one embodiment, such as Figure 5 As shown, the positional feature information between a target data point and any of its neighboring data points can be represented by multiple feature values. For example, the first frequency value f1 corresponding to the target data point, the second frequency value f2 corresponding to the neighboring data point, the timestamp t1 of the target data point mapped to the time axis, and the time slot Δt = t2 - t1 corresponding to the absolute time difference between the neighboring data point and the target data point can be used, where t2 is the timestamp of the neighboring data point mapped to the time axis. These feature values ​​are arranged in an orderly manner to form a two-point feature vector, which is used to represent the positional feature information between the target data point and one of its neighboring data points. It is easy to understand that the feature vector not only contains the absolute positional information of the target data point and the neighboring data point in the time-frequency diagram, but also contains the relative positional information between them, thus achieving an effective representation of their positional information. In alternative embodiments, some feature values ​​can be changed, for example, replacing the second frequency value or the first frequency value with the absolute frequency difference between the second frequency value and the first frequency value.

[0054] Based on the above process, it can be understood that each spectrum data can obtain its corresponding comprehensive feature information. The comprehensive feature information can effectively represent the positional features of each data point in the corresponding spectrum diagram, realizing an effective feature representation of each data point in the spectrum diagram. The comprehensive feature information is independent of the amplitude of each data point, is not easily interfered with, and because it represents the positional feature information between points through two-point feature vectors, it also plays an effective role in characterizing the changing trend of network transmission quality in the time domain.

[0055] In other embodiments, when there are multiple time-domain quality index data corresponding to multiple index dimensions such as bandwidth, network latency, and packet loss rate, the corresponding spectrum data can be obtained for each dimension, and thus the comprehensive feature information corresponding to each dimension can be obtained. In this case, each user can obtain multiple comprehensive feature information corresponding to each dimension. Subsequently, the corresponding comprehensive feature information of each dimension can be used separately, and the results of each dimension can be combined to obtain the overall result.

[0056] Step S1400: Determine the feature similarity between two users regarding network transmission quality based on the comprehensive feature information of any two users.

[0057] In one embodiment, when there are multiple users, the similarity between two users can be determined by using the combined feature information of the first and second users to form a user pair, thereby realizing the similarity judgment between the two users.

[0058] In another embodiment, for cases with more than two users, the first user can be matched with each of the other users to determine multiple user pairs, with each user pair containing the first user. Then, for each user pair, the feature similarity between the first user and the other user is determined to achieve similarity determination in the case of multiple users.

[0059] There are several ways to determine the feature similarity between two users.

[0060] In one embodiment, this is achieved through the following process, such as Figure 6 As shown, the process includes:

[0061] Step S1410: For any two users' comprehensive feature information, find out whether there is similar inter-point feature information between the target data points in the comprehensive feature information of the first user and the comprehensive feature information of the second user, and count the number of similar data points.

[0062] Step S1420: Calculate the ratio between the number of similar data points and the total number of target data points in the comprehensive feature information of the first user, and use it as the feature similarity between the first user and the second user regarding network transmission quality.

[0063] Specifically, in this embodiment, based on statistical methods, for the first user and the second user in the same user pair, based on the comprehensive feature information corresponding to the spectrum data of the same indicator dimension, the feature similarity comparison is performed between the inter-point feature information corresponding to each data point in the comprehensive feature information of the first user and the inter-point feature information of each data point in the comprehensive feature information of the second user. By comparing whether there are data points with similar inter-point feature information in the comprehensive feature information of the second user for a data point of the first user, the number of similar data points of the first user relative to the second user is determined. The ratio between the number of similar data points and the total number of inter-point feature information of the first user, that is, the total number of data points, is determined as the feature similarity of the first user with respect to the network transmission quality of the corresponding indicator dimension.

[0064] In the process of calculating the feature similarity between two users using statistical methods, when comparing the similarity of feature information between data points, since each feature information contains feature vectors corresponding to the target data point and its multiple neighboring data points, and each feature vector contains multiple feature values, we can search for similar feature vectors in all feature vectors of the second user's feature information based on one feature vector corresponding to a data point of the first user. When similar feature vectors are found, the confidence level is incremented by 1 for statistical purposes. By traversing each feature vector in the first user's feature information, we can find similar feature vectors in the second user's feature information for each feature vector. The final confidence level can be obtained, which is actually the total number of similar feature vectors between one point feature information of the first user and one point feature information of the second user. When the total number of similar feature vectors is greater than a preset threshold, or when the ratio of the total number of similar feature vectors to the total number of feature vectors in the point feature information of the first user is greater than a preset threshold, it can be confirmed that the point feature information of the first user is similar to the point feature information of the second user. That is, there is a point feature information in the comprehensive feature information of the second user that is similar to the point feature information of the first user. The number of similar data points of the first user is incremented by 1. Following this process, the number of similar data points can be calculated by searching for each point feature information in the comprehensive feature information of the first user in the comprehensive feature information of the second user.

[0065] For the similarity comparison of the feature vectors of the first user and the second user, since a feature vector is composed of multiple feature values, such as the first frequency value, second frequency value, time slot, and timestamp as mentioned earlier, the two feature vectors to be compared can be compared one by one according to their feature types. If the difference between the feature values ​​corresponding to each feature type is less than a preset threshold, the two feature vectors are considered similar, and the confidence score can be accumulated. Otherwise, the two feature vectors are not similar, and the next feature vector is called for similar similarity comparison until all feature vectors in the same feature information have been compared for similarity.

[0066] Based on statistical methods, the similarity between data points is compared by comparing the feature vectors corresponding to the frequency domain features, thereby determining whether two data points are similar. Ultimately, this achieves a quantitative determination of the similarity between the network conditions of two users. The computational efficiency is high. The comparison process is mainly carried out by comparing the numerical values ​​of frequency and time domain features, allowing for a certain degree of offset. It does not require comparing the magnitude of each data point, thus improving the detection range. Moreover, the confidence score statistics are used in the process to assist in determining feature similarity, which has a higher accuracy rate compared to directly measuring feature similarity.

[0067] In another embodiment, it can be as follows Figure 7 As shown, a neural network model based on a dual-tower structure employs two isomorphic text feature extraction models to extract deep semantic information from the comprehensive feature information of the first and second users, respectively, obtaining their respective deep feature vectors. These vectors are then concatenated into a comprehensive feature vector, which is then fully connected through a classifier and mapped to a pre-defined binary classification space to obtain the classification probability of the vector mapping to the positive category. This probability serves as the feature similarity between the first and second users regarding the network transmission quality of the corresponding indicator dimension. The neural network model can be pre-trained to a convergent state by constructing training samples in advance and providing manually labeled data as input to the model. Iterative training can then be performed to achieve convergence. When using comprehensive feature information as input to the model, it can be reasonably standardized according to pre-defined unified encoding rules, which are uniformly applicable during the model's training and online inference phases. For example, multiple feature vectors corresponding to each data point can be expanded and concatenated into a single-row vector, that is, the inter-point feature information of each data point can be represented as a single-row vector. The single-row vectors corresponding to the inter-point feature information of multiple data points form a vector sequence to represent the comprehensive feature information.

[0068] Based on a neural network model, this system can determine the similarity of two users by analyzing their comprehensive feature information for the same metric dimensions and thus determine their corresponding feature similarity. It can also utilize the deep semantic information in the comprehensive feature information to make a deeper understanding and obtain effective feature similarity.

[0069] In one embodiment, for time-domain quality indicator data with multiple corresponding indicator dimensions, the individual feature similarity between each pair of users can be determined according to the above process for each indicator dimension. Then, the feature similarity of each user pair under each indicator dimension is weighted and summed to obtain a comprehensive feature similarity to represent the feature similarity between the two users in that user pair. During the weighted summation, the permissions corresponding to each indicator dimension can be flexibly set, and the sum of each weight can be set to 1 to harmonize the final result to a specific numerical range, such as [0,1], for easy measurement.

[0070] As can be seen from the above embodiments, this application has multiple technical advantages, including but not limited to:

[0071] First, this application performs time-frequency transformation on the user's time-domain quality index data to obtain the corresponding frequency-domain spectrum data. Then, it extracts features from each data point in the spectrum data, using the positional feature information between two data points within a neighborhood to represent the features of each data point. Each data point is represented by its inter-point feature information through the positional feature information between it and its neighboring data points. Based on this, the inter-point feature information of each data point in the spectrum data is determined as the comprehensive feature information corresponding to that spectrum data, thus achieving an effective feature representation of each user's time-domain quality index data in the frequency domain mapping. This allows each spectrum data to be distinguished and compared through its comprehensive feature information. Therefore, the feature similarity can be quantitatively determined based on the pairwise comprehensive feature information, achieving an effective measurement of the similarity of the time-domain quality index data.

[0072] Secondly, when representing the spectral data, this application uses the positional feature information between data points. According to the characteristics of spectral data, the positional feature information is mainly composed of time-domain and frequency-domain information. Therefore, it does not need to rely on the amplitude of each data point for feature statistics. It is not sensitive to amplitude or noise, but only sensitive to frequency in the frequency domain. It can effectively measure the same network change trend. Compared with other methods, the positional feature information between two points on the time spectrum can provide a more accurate and interference-resistant feature representation method, which is superior to other methods.

[0073] Furthermore, based on the effective measurement of time-domain quality index data for each user and the determination of feature similarity between pairs of users, this application can be used to classify the network conditions of a large number of users, making the classification more accurate. Therefore, it is not difficult to understand how the technical solution of this application can be extended to application scenarios such as live streaming. By conducting corresponding simulation tests based on the similar network conditions obtained from the correct classification, network optimization can be subdivided into user groups corresponding to different categories of network conditions, so as to further match different optimal algorithms for different user groups, thereby achieving overall network optimization of the network transmission quality of live streaming services.

[0074] Based on any embodiment of this application, please refer to Figure 8 The process involves transforming the time-domain quality index data of each user from the time domain to the frequency domain, converting the values ​​of each quality index into the amplitude of data points in the spectrum data, and obtaining the spectrum data, including:

[0075] Step S1210: Window and frame-segment the data for each time-domain quality index to obtain the corresponding data frame sequence;

[0076] This embodiment utilizes speech preprocessing technology to perform time-frequency transformation on the time-domain quality index data corresponding to each index dimension. Therefore, for each time-domain quality index data, Hamming windowing is applied for windowing to facilitate frame division, which allows for a certain frame shift. Through the windowing operation, the time-domain quality index data is sampled to obtain the corresponding data frame sequence.

[0077] In one embodiment, the corresponding windowing and framing operations can be performed using the following Hamming window formula:

[0078]

[0079] Where w(n) is the window coefficient at index n of the data sample, and N is the total number of samples.

[0080] Step S1220: Perform a short-time Fourier transform on the data frame sequence to convert the values ​​of each quality index into the amplitude of the data points in the spectrum data;

[0081] Furthermore, the short-time Fourier transform algorithm is applied to the data frame sequence to perform time-frequency transformation, so as to transform the time-domain quality index data from the time domain to the spectrum. Thus, in the spectrum diagram, the positional correspondence of data points can be described using two dimensions: the time domain and the frequency domain. As for the quality index data corresponding to each data point, it is transformed into the amplitude of the data point in the spectrum data, that is, the power value.

[0082] In one embodiment, the short-time Fourier transform operation is performed using the following formula:

[0083]

[0084] Among them, X m (ω) represents the frequency domain data of the m-th frame, x(n) represents the time domain data sample, R represents the length of each slide in the time domain, which is equal to the width of the Hamming window minus the width of the frame shift overlap. The constrained window function w(n) is as follows:

[0085]

[0086] The specific sliding process is as follows Figure 9 The example is by Figure 9 It can be seen that a certain degree of overlap, i.e., frame shift, is allowed between the windows corresponding to each data frame.

[0087] Step S1230: For each data frame corresponding to the windowing operation, filter out a preset proportion of low-amplitude data points to obtain the spectrum data.

[0088] Considering that noise interference may cause fluctuations in the collected quality index values, the data frames obtained by windowing can be appropriately filtered to remove some low-amplitude data points. To do this, the amplitude of each data point is first calculated using the following formula:

[0089] |X m (ω)| 2

[0090] After determining the amplitude of each data point, the data points can be filtered. The proportion of low-amplitude data points removed by filtering can be arbitrarily set between 20% and 40%, for example, 30%. After filtering the data points, the purified spectrum data is obtained, which can then be used to extract comprehensive feature information.

[0091] For time-domain quality index data with multiple corresponding index dimensions, the spectral data corresponding to each index dimension can be obtained in the following ways.

[0092] Based on the above embodiments, it is easy to understand that, taking advantage of the characteristic that time-domain quality index data is obtained by discrete acquisition in the time domain, the time-frequency transformation technique is cleverly used to convert the time-domain quality index data into spectrum data, which facilitates the extraction of frequency domain features. In this process, some low-amplitude data points are filtered out to avoid using data points with excessively low amplitudes for feature extraction, making the extracted features insensitive to noise and thus playing an anti-interference role.

[0093] Based on any embodiment of this application, comprehensive feature information is extracted from the spectrum data of each user. Please refer to [link to relevant documentation]. Figure 10 ,include:

[0094] Step S1310: Based on the spectrum data of each user, determine each data point as a target data point, and determine the neighboring data points that fall within the neighborhood range of each target data point;

[0095] When performing feature extraction on each user's spectrum data, each data point in the spectrum data can be used as a target data point for corresponding feature extraction. When performing feature extraction on a target data point, it is necessary to determine the neighborhood range of that target data point and the neighboring data points within that neighborhood range. This allows for the formation of data point pairs based on the target data point and each of its neighboring data points. A feature vector, i.e., a two-point feature vector, is then determined for each data point pair. The neighborhood range, as mentioned earlier, can be obtained by appropriately extending the range along the time and frequency domains based on the target data point.

[0096] Step S1320: Determine the two-point feature vectors corresponding to each target data point and its respective neighboring data points. Each two-point feature vector contains the positional feature information between the target data point and its corresponding neighboring data points.

[0097] When extracting the two-point feature vector between each pair of data points, the main focus is on numerically representing the positional feature information between the data point pairs. Multiple feature data are used to jointly describe the positional information of the target data point and one of its neighboring data points relative to the spectrum mapped by the spectral data, including absolute and relative positional information. As in the previous embodiment, the first frequency value corresponding to the target data point, the second frequency value corresponding to the neighboring data point, the timestamp of the target data point mapped to the time-frequency graph, and the time slot represented by the absolute difference between the timestamps of the target data point and the neighboring data point can be used to jointly construct the positional feature information between the target data point and the neighboring data point, forming a two-point feature vector.

[0098] Step S1330: Merge all two-point feature information corresponding to each target data point to form the point-to-point feature information corresponding to that target data point;

[0099] It is easy to understand that for a target data point, there are multiple neighboring data points in its neighborhood, and thus multiple data point pairs. Therefore, multiple two-point feature vectors can be obtained accordingly. By merging these two-point feature vectors together according to certain encoding rules, the inter-point feature information corresponding to the target data point is formed, which can realize the comprehensive representation of the characteristics of the target data point.

[0100] Step S1340: Merge all the inter-point feature information corresponding to the spectrum data to construct comprehensive feature information, which serves as a feature representation of the network transmission quality of the corresponding user.

[0101] For the same spectrum data, there are multiple data points that can be used as target data points, and each target data point can be used to extract corresponding inter-point feature information. Therefore, by merging the inter-point feature information of all target data points according to certain coding rules, the comprehensive feature information corresponding to the spectrum data can be constructed. This enables the feature representation of the spectrum data of the corresponding user under the corresponding indicator dimension, representing the network transmission quality measured by the user under the corresponding indicator dimension, and providing a data description of the user's network conditions from one aspect.

[0102] As can be seen from the above embodiments, the process of feature extraction for each spectrum data is based on numerical operations, and the operations are accurate to the data points. It can be implemented without complex models, and its computational efficiency is high. In the encoding process, the positional information between each data point on the time-frequency graph is used for feature representation, without considering the amplitude of each data point. This realizes the fingerprint extraction of time-domain quality index data in the frequency domain. The data fingerprint can effectively represent the time-domain quality index data, which facilitates the subsequent accurate determination of the feature similarity between any two users.

[0103] Based on any embodiment of this application, after determining the feature similarity between two users according to the comprehensive feature information of any two users, please refer to... Figure 11 ,include:

[0104] Step S2100: Cluster the users based on the feature similarity between each pair of users to determine multiple similar user clusters, and label each similar user cluster with a corresponding category label.

[0105] As described in the preceding embodiments of this application, a feature similarity can be determined for any two users. Therefore, clustering operations can be performed based on the feature similarity of each user pair to determine one or more similar user clusters.

[0106] In one embodiment, multiple users are paired one by one with each of the other users, using a first user as a reference. Thus, the feature similarity of the first user relative to any other user can be determined similarly to any of the embodiments disclosed above, thereby obtaining the feature similarity of the corresponding user pair. Furthermore, by ranking the other users using feature similarity, it is possible to determine which users are highly similar to the first user. Users whose feature similarity exceeds a preset threshold are selected and, together with the first user, form a cluster of similar users. Since the comprehensive feature information of the users in this cluster is highly similar, their network conditions, i.e., network transmission quality, are also similar, thereby enabling further determination of the similarity in network transmission quality among multiple users.

[0107] In another embodiment, for cases with more than two users, any clustering algorithm can be used to perform the clustering operation. The clustering algorithm includes, but is not limited to, any of the following: K-Means clustering algorithm, mean-shift clustering algorithm, density-based clustering algorithm (DBSCAN), expectation-maximum clustering algorithm based on Gaussian mixture model, agglomerative hierarchical clustering algorithm, graph community detection clustering algorithm, etc.

[0108] The K-Means clustering algorithm first selects some clusters and randomly initializes the core user nodes of each cluster as centroids. A user node can be a pair of users and their feature similarity. The centroid is a position with the same vector length as each user node. Then, the distance from each user node to the centroid is calculated, and the user node is assigned to the cluster closest to the centroid. The centroid of each cluster is then recalculated and determined. These steps are repeated until the change in the centroid of each cluster does not exceed a preset range after each iteration. Alternatively, centroids can be randomly initialized multiple times, and the best result is selected.

[0109] Mean-shift clustering is a sliding window-based algorithm that uses a sliding window to find dense regions of user nodes. It's a centroid-based algorithm that locates the centroid of each cluster by updating candidate centroids to the mean of the points within the sliding window. Then, similar windows are removed from these candidate windows, ultimately forming a set of centroids and their corresponding clusters.

[0110] Density-Based Clustering (DBSCAN), similar to Mean Shift Clustering, is also a density-based clustering algorithm. It first determines an adjacency range and a preset threshold. Starting with an unvisited user node, it uses this node as the center point and checks if the number of user nodes within its adjacency range is greater than or equal to the preset threshold. If it is, the user node is marked as the center point; otherwise, it is marked as a noise node. This process is repeated. If a noise node exists within the adjacency range of a center point, it is marked as an edge node; otherwise, it remains a noise node. This process is repeated until all user nodes have been visited.

[0111] The Expectation-Maximization (EM) clustering algorithm based on Gaussian Mixture Models (GMM) first selects the number of clusters (similar to K-Means) and randomly initializes the Gaussian distribution parameters (mean and variance) for each cluster. Alternatively, a relatively accurate mean and variance can be derived from observed data. Then, given the Gaussian distribution for each cluster, the probability of each user node belonging to each cluster is calculated. A user node is more likely to belong to a cluster the closer it is to the center of the Gaussian distribution. Furthermore, based on these probabilities, Gaussian distribution parameters are calculated to maximize the probability of user nodes. These new parameters can be calculated using a weighted average of the user node probabilities, where the weights represent the probability of a user node belonging to a cluster. The first two steps are then iterated until the changes during iteration do not exceed a preset range, completing the clustering process.

[0112] Agglomerative hierarchical clustering algorithms are divided into two categories: top-down and bottom-up. Agglomerative hierarchical clustering (HAC) is a bottom-up clustering algorithm. HAC first treats each user node as a single cluster, then calculates the distance between all clusters to merge clusters until all clusters are aggregated into a single cluster.

[0113] The Graph Community Detection (GCM) clustering algorithm works as follows: First, it initially assigns each user node, considered a vertex, to its own community, and then calculates the modularity M of the entire network. Second, if two communities merge, the algorithm calculates the resulting change in modularity ΔM. Essentially, it selects the community pair with the largest increase in ΔM and merges them. Third, it calculates and records the new modularity M for this cluster. Then, it repeats steps one and two, merging community pairs each time, until it obtains the maximum gain in ΔM. Finally, it records the new clustering pattern and its corresponding modularity score M.

[0114] As described above regarding the various clustering algorithms, any feasible clustering algorithm can be used to cluster user pairs based on their feature similarity, thereby identifying one or more similar user clusters. These similar user clusters consist of multiple users. Since users have established de facto edge connections based on feature similarity, the similar user clusters obtained through clustering are generally composed of users with similar network conditions.

[0115] To facilitate subsequent calls, the clustering structure can be stored in a database, and each similar user cluster can be labeled with a different category label, so that data corresponding to the network conditions of users of the same type can be called by specifying the category label.

[0116] Step S2200: Construct simulated quality index data based on the temporal quality index data of users in similar user clusters corresponding to the specified category labels to implement network condition simulation, and test the audio and video quality scores obtained by different audio and video coding algorithms under simulated network conditions.

[0117] It's easy to understand that each cluster of similar users obtained from clustering can be accessed by specifying a category label. When it's necessary to perform network condition simulation tests on users in a cluster of similar users corresponding to a category label, you can obtain that batch of users by specifying the corresponding category label, and then perform simulation based on the quality index data corresponding to that batch of users, or directly call the quality index data of one user as the simulation data for the network conditions of all users, and start the test under the simulated network conditions.

[0118] The tests primarily target various audio and video coding algorithms, or audio and video transmission algorithms. A comparative testing approach can be used, employing both the first and second audio and video coding algorithms under simulated network conditions. During the testing process, audio and video quality scores are obtained for each algorithm. These scores can be objective quality scores, such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity), or subjective quality scores.

[0119] Step S2300: Select the optimal algorithm from the different audio and video encoding algorithms based on the audio and video quality score, and configure it as the default algorithm for encoding audio and video streams of users in the target similar user cluster for the live streaming service.

[0120] Each audio / video coding algorithm is tested under specific network conditions within a cluster of similar users to obtain its corresponding audio / video quality score. Based on these scores, the best algorithm is selected. Typically, the algorithm with the highest audio / video quality score is chosen as the optimal algorithm for the given network conditions. This optimal algorithm can then be configured as the default algorithm for users within the cluster or for users with similar network conditions, providing audio / video stream encoding and transmission services to these users.

[0121] As can be seen from the above embodiments, by clustering user groups based on feature similarity, multiple similar user clusters are divided, realizing the classification of different network condition levels. This facilitates the testing of audio and video stream coding algorithms under different network conditions. Thus, the optimal audio and video stream coding algorithm can be determined by different network conditions, ensuring that the audio and video stream coding transmission service for users under various network conditions can achieve the best results. Applied to the field of live streaming, it can achieve an overall improvement in the audio and video stream coding transmission service for users of the entire platform, ensuring a superior user experience. All these effects are achieved based on the effective feature similarity determination of user network transmission quality in this application.

[0122] Based on any embodiment of this application, clustering is performed according to the feature similarity between pairs of users to determine multiple similar user clusters. Please refer to [link to relevant documentation]. Figure 12 ,include:

[0123] Step S2110: Take each user pair as a sample point, and randomly select a preset number of sample points as the center points of the corresponding multiple clusters.

[0124] Each user pair is considered a sample point, or user node. Then, a preset number of sample points can be randomly selected. The preset number of sample points depends on the number of clusters to be preset. The number of clusters can be set by the technical personnel in this field as needed. Each randomly selected sample point is used as the center point of one of the clusters, so that each cluster has a center point.

[0125] Step S2120: Perform a classification process for each sample point, wherein the data distance from the sample point to the center point of each cluster is calculated, and the sample point is assigned to the cluster corresponding to the center point with the smallest data distance.

[0126] For each sample point, the data distance to the center point of each cluster is calculated. Specifically, the data distance from the sample point to the center point can be calculated based on feature similarity. The cluster with the smallest distance to the center point is determined, and the sample point is added to the cluster with the smallest distance, becoming a member of that cluster. Each sample point is classified in this way until all sample points have been traversed.

[0127] When calculating data distance, any data distance algorithm can be used, such as Euclidean span, cosine similarity, Minkowski distance, or Pearson correlation coefficient algorithm.

[0128] Step S2130: For each cluster, recalculate its center position to redetermine its center point, and iterate the above classification process until the preset number of iterations is reached or the center point of each cluster remains unchanged in more than two iterations.

[0129] After classifying all sample points, the center position of each cluster is recalculated. The nearest neighboring center point is then determined as the new center point of that cluster, thus re-determining the center points of each cluster. Then, the iteration continues from step S2120, continuously adjusting the center points and members of each cluster through multiple iterations until a preset number of iterations is reached, or until the center points of each cluster remain unchanged for more than two iterations, at which point the clustering process is complete.

[0130] Step S2140: Add the users corresponding to each sample point in each cluster class to the similar user cluster corresponding to that cluster class.

[0131] After completing the above clustering process, for each cluster, the users corresponding to its sample points are obtained to form the similar user clusters corresponding to that cluster. In this way, all users are classified and multiple similar user clusters are obtained. Each similar user cluster contains multiple similar users with comparable network conditions.

[0132] According to this embodiment, by using statistical methods, multiple users can be quickly clustered to obtain multiple similar user clusters, which facilitates subsequent simulation tests to distinguish different similar user clusters.

[0133] Please see Figure 13 To meet one of the purposes of this application, a network quality similarity determination device is provided, which is a functional embodiment of the network quality similarity determination method of this application. The device includes: a data acquisition module 1100, a time-frequency transformation module 1200, a feature representation module 1300, and a similarity determination module 1400. The data acquisition module 1100 is used to acquire time-domain quality index data representing network transmission quality corresponding to multiple users. The time-domain quality index data includes quality index values ​​obtained by discrete acquisition along the time domain. The time-frequency transformation module 1200 is used to convert the time-domain quality index data of each user... The process transforms the time domain to the frequency domain, converting the values ​​of various quality indicators into the amplitude of data points in the spectrum data, thereby obtaining the spectrum data. The feature representation module 1300 is used to extract comprehensive feature information for each user's spectrum data. The comprehensive feature information includes multiple inter-point feature information corresponding to each target data point in the spectrum data, and the inter-point feature information includes the positional feature information between a single target data point and its neighboring data points. The similarity determination module 1400 is used to determine the feature similarity between any two users regarding network transmission quality based on the comprehensive feature information of any two users.

[0134] Based on any embodiment of this application, the time-frequency transformation module 1200 includes: a data framing unit, used to perform windowing and framing processing on each time-domain quality index data to obtain a corresponding data frame sequence; a time-frequency transformation unit, used to perform short-time Fourier transform on the data frame sequence to convert the values ​​of each quality index into the amplitude of data points in the spectrum data; and a data filtering unit, used to filter out a preset proportion of low-amplitude data points in each data frame corresponding to the windowing operation to obtain the spectrum data.

[0135] Based on any embodiment of this application, the feature representation module 1300 includes: an inter-point mapping unit, used to determine each data point in the spectrum data of each user as a target data point, and to determine the neighboring data points falling within the neighborhood range of each target data point; an inter-point coding unit, used to determine the two-point feature vectors corresponding to each target data point and each of its neighboring data points, each two-point feature vector containing the positional feature information between the corresponding target data point and the corresponding neighboring data point; a data point-level construction unit, used to merge all the two-point feature information corresponding to each target data point to form the inter-point feature information corresponding to the target data point; and a spectrum-level construction unit, used to merge all the inter-point feature information corresponding to the spectrum data to construct comprehensive feature information, which serves as the feature representation of the network transmission quality of the corresponding user.

[0136] Based on any embodiment of this application, the location feature information includes the frequency values ​​of the corresponding target data point and its corresponding neighboring data points, the time slot between them, and the timestamp corresponding to the target data point.

[0137] Based on any embodiment of this application, the similarity determination module 1400 includes: a similarity statistics unit, used to, for any two users' comprehensive feature information, search whether there is similar inter-point feature information between each target data point in the comprehensive feature information of the first user and the comprehensive feature information of the second user, and count the number of similar data points; and a similarity quantification unit, used to calculate the ratio between the number of similar data points and the total number of target data points in the comprehensive feature information of the first user, as the feature similarity between the first user and the second user regarding network transmission quality.

[0138] Based on any embodiment of this application, the similarity determination module 1400 further includes: a clustering labeling module, used to cluster users based on the feature similarity between pairs of users to determine multiple similar user clusters, and to label each similar user cluster with a corresponding category label; a simulation scoring module, used to construct simulation quality index data based on the temporal quality index data of users in the similar user clusters corresponding to the specified category labels to implement network condition simulation, and to test the audio and video quality scores obtained by different audio and video encoding algorithms under simulated network conditions; and an optimal configuration module, used to select the optimal algorithm among the different audio and video encoding algorithms based on the audio and video quality scores, and configure it as the default algorithm for encoding audio and video streams for users in the target similar user clusters in the live streaming service.

[0139] Based on any embodiment of this application, the clustering labeling module includes: a cluster partitioning unit, used to randomly select a preset number of sample points as the center points of corresponding clusters, taking each user pair as a sample point; a sample point classification unit, used to perform a classification process for each sample point, wherein the data distance from the sample point to the center point of each cluster is calculated, and the sample point is assigned to the cluster corresponding to the center point with the smallest data distance; an adjustment iteration unit, used to recalculate the center position of each cluster to redetermine its center point, and iterate the above classification process until a preset number of iterations is reached or the center point of each cluster remains unchanged in two or more iterations; and a user cluster construction unit, used to add the users corresponding to each sample point in each cluster to the similar user clusters corresponding to that cluster.

[0140] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 14 As shown, the computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When the computer-readable instructions are executed by the processor, the processor can implement a product search category identification method. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the network quality similarity determination method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that... Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0141] In this embodiment, the processor is used to execute... Figure 13 The system contains the specific functions of each module and its sub-modules, and the memory stores the program code and various data required to execute the aforementioned modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / sub-modules in the network quality similarity determination device of this application, and the server can call the server's program code and data to execute the functions of all sub-modules.

[0142] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the network quality similarity determination method of any embodiment of this application.

[0143] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

[0144] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0145] In summary, this application can accurately measure the similarity between different network conditions based on the spectral data of the time-domain quality index data corresponding to different network conditions. It can effectively classify different network conditions, thereby facilitating network optimization according to different types of network conditions. This can serve application scenarios such as live streaming to improve service quality.

Claims

1. A method for determining network quality similarity, characterized in that, include: Acquire time-domain quality index data representing network transmission quality for multiple users, wherein the time-domain quality index data includes quality index values ​​collected discretely along the time domain; The time-domain quality index data are transformed from the time domain to the frequency domain, so that the values ​​of each quality index are converted into the amplitude of the data points in the spectrum data, and the spectrum data is obtained. For each user's spectrum data, extract its comprehensive feature information. The comprehensive feature information includes multiple inter-point feature information corresponding to each target data point in the spectrum data. The inter-point feature information includes multiple two-point feature vectors between a single target data point and its various neighboring data points. Each two-point feature vector contains a feature value representing the positional feature information of the relative positional relationship between the two points in the time-frequency diagram. The similarity of network transmission quality between two users is determined based on their combined feature information.

2. The network quality similarity determination method according to claim 1, characterized in that, Transforming the time-domain quality index data of each user from the time domain to the frequency domain, converting the values ​​of each quality index into the amplitude of data points in the spectrum data, and obtaining the spectrum data includes: Windowing and frame segmentation are performed on each time-domain quality index data to obtain the corresponding data frame sequence; A short-time Fourier transform is performed on the data frame sequence to convert the values ​​of each quality index into the amplitude of the data points in the spectrum data; For each data frame corresponding to the windowing operation, a predetermined proportion of low-amplitude data points are filtered out to obtain the spectrum data.

3. The network quality similarity determination method according to claim 1, characterized in that, For each user's spectrum data, extract their comprehensive feature information, including: Based on each user's spectrum data, each data point is identified as a target data point, and for each target data point, a neighborhood data point falling within its neighborhood range is determined. Determine the two-point feature vectors corresponding to each target data point and its respective neighboring data points. Each two-point feature vector contains the positional feature information between the target data point and its corresponding neighboring data points. Merge all two-point feature information corresponding to each target data point to form the point-to-point feature information corresponding to that target data point; The inter-point feature information corresponding to the spectrum data is merged to construct comprehensive feature information, which serves as a feature representation of the network transmission quality of the corresponding user.

4. The network quality similarity determination method according to claim 1, characterized in that, The location feature information includes the frequency values ​​of the target data point and its corresponding neighboring data points, the time slot between them, and the timestamp corresponding to the target data point.

5. The network quality similarity determination method according to claim 1, characterized in that, The feature similarity between any two users is determined based on their combined feature information, including: For any two users’ comprehensive feature information, search whether the inter-point feature information of each target data point in the comprehensive feature information of the first user is similar to the inter-point feature information of the second user, and count the number of similar data points. The ratio between the number of similar data points and the total number of target data points in the comprehensive feature information of the first user is calculated as the feature similarity between the first user and the second user regarding network transmission quality.

6. The network quality similarity determination method according to any one of claims 1 to 5, characterized in that, After determining the feature similarity between any two users based on their combined feature information, the following steps are taken: Clustering is performed based on the feature similarity between each pair of users to determine multiple similar user clusters, and each similar user cluster is labeled with a corresponding category label. Based on the temporal quality index data of users in similar user clusters corresponding to specified category labels, simulated quality index data is constructed to implement network condition simulation. Under simulated network conditions, the audio and video quality scores obtained by different audio and video coding algorithms are tested. Based on the audio and video quality score, the optimal algorithm among the different audio and video encoding algorithms is selected and configured as the default algorithm for encoding audio and video streams for users in the target similar user cluster in the live streaming service.

7. The network quality similarity determination method according to claim 6, characterized in that, Clustering is performed based on the pairwise feature similarity among the multiple users to determine multiple similar user clusters, including: Each user pair is used as a sample point, and a preset number of sample points are randomly selected as the center points of the corresponding clusters. A classification process is implemented for each sample point, in which the data distance from the sample point to the center point of each cluster is calculated, and the sample point is assigned to the cluster corresponding to the center point with the smallest data distance. For each cluster, recalculate its center position to redetermine its center point, and iterate the above classification process until the preset number of iterations is reached or the center point of each cluster remains unchanged in more than two iterations; Add the users corresponding to each sample point in each cluster to the corresponding similar user cluster of that cluster.

8. A network quality similarity determination device, characterized in that, include: The data acquisition module is used to acquire time-domain quality index data representing network transmission quality corresponding to multiple users. The time-domain quality index data includes quality index values ​​obtained by discrete collection along the time domain. The time-frequency conversion module is used to convert the time-domain quality index data of each user from the time domain to the frequency domain, so that the values ​​of each quality index are converted into the amplitude of data points in the spectrum data, and thus obtain the spectrum data. The feature representation module is used to extract comprehensive feature information for each user's spectrum data. The comprehensive feature information includes multiple inter-point feature information corresponding to each target data point in the spectrum data. The inter-point feature information includes multiple two-point feature vectors between a single target data point and its various neighboring data points. Each two-point feature vector contains a feature value representing the positional feature information of the relative positional relationship between the two points in the time-frequency diagram. The similarity determination module is used to determine the feature similarity between two users regarding network transmission quality based on the comprehensive feature information of any two users.

9. A computer device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images