Cinema live broadcast optimization method and system based on big data analysis

Abstract

The invention provides a cinema live broadcast optimization method and system based on big data analysis, and the method comprises the steps: firstly obtaining historical live broadcast big data of a target cinema, which comprises a plurality of live broadcast data streams composed of user film watching behavior records and live broadcast service state records, and then carrying out the multi-dimensional film watching feature extraction of the historical live broadcast big data, the method comprises the steps of obtaining real-time film watching features and service stability features, then performing dynamic strategy matching on the real-time film watching features and the service stability features based on a pre-trained dynamic optimization model, generating an optimization labeling result, obtaining a live broadcast service adjustment strategy according to the optimization labeling result, and synchronizing the live broadcast service adjustment strategy to a live broadcast service node in real time. The current live broadcast parameter configuration is adjusted according to the live broadcast service adjustment strategy, and finally, the parameters of the dynamic optimization model are updated according to the real-time live broadcast quality index acquired by the adjusted live broadcast parameter configuration, so that the dynamic optimization of cinema live broadcast is realized, and the live broadcast service quality is improved.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and more particularly, to a method and system for optimizing cinema live broadcast based on big data analysis. Background Art

[0002] At present, with the rapid development of cinema live broadcast technology and the increasing demand of audiences for high-quality viewing experiences, cinema live broadcast services are facing many technical challenges that need to be solved urgently. In the operation of traditional cinema live broadcast systems, the configuration of live broadcast parameters and service adjustment mainly rely on manual experience or preset fixed rules. However, this method has obvious limitations.

[0003] On the one hand, manual experience is often subjective and one-sided, and it is difficult to comprehensively and accurately grasp the viewing behavior patterns and demand characteristics of different user groups in different scenarios. The preferences, habits of different users when watching movies, and their sensitivities to live broadcast quality vary greatly. It is very difficult for manual experience to analyze and process these complex and changeable factors in detail, resulting in the inability of live broadcast services to accurately meet the personalized needs of users and affecting the viewing experience of users.

[0004] On the other hand, the preset fixed rules lack flexibility and adaptability. The cinema live broadcast environment is dynamically changing, and factors such as network conditions, device performance, and the number of users are changing all the time. Fixed rules cannot be dynamically adjusted according to these real-time changes. When encountering sudden situations such as network fluctuations and sudden increases in user traffic, problems such as freezing and delay are likely to occur in the live broadcast service, seriously affecting the stability and smoothness of the live broadcast.

[0005] Therefore, the existing cinema live broadcast technology is unable to provide high-quality, stable and personalized live broadcast services when dealing with complex and changeable viewing demands and dynamically changing live broadcast environments. Summary of the Invention

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for optimizing cinema live broadcast based on big data analysis, the method comprising:

[0007] Obtain historical live broadcast big data of a target cinema, the historical live broadcast big data including a plurality of live broadcast data streams, each live broadcast data stream consisting of at least one user viewing behavior record and a corresponding live broadcast service status record;

[0008] Perform multi-dimensional viewing feature extraction processing on the historical live broadcast big data to obtain real-time viewing features and service stability features of each live broadcast data stream;

[0009] Based on the pre-trained dynamic optimization model, perform dynamic policy matching processing on the real-time viewing features and the service stability features to generate an optimized annotation result for the live data stream, and generate a live service adjustment strategy according to the optimized annotation result;

[0010] Synchronize the live service adjustment strategy to the live service node in real time, and trigger the live service node to adjust the current live parameter configuration according to the live service adjustment strategy;

[0011] Collect real-time live quality metrics based on the adjusted live parameter configuration, and update the parameters of the dynamic optimization model according to the real-time live quality metrics.

[0012] In another aspect, an embodiment of the present invention further provides a cinema live broadcast optimization system based on big data analysis, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0013] Based on the above aspects, the embodiments of the present application achieve a global performance improvement of the cinema live broadcast service in a complex dynamic environment. Based on the analysis of historical live broadcast big data, by integrating multi-dimensional viewing features and service stability features, the description of the live broadcast service state has spatio-temporal correlation and behavior perception ability. The pre-trained dynamic optimization model, as the core decision-making engine, not only realizes the intelligent mapping of real-time viewing features and service stability features, but also generates forward-looking optimized annotation results through a dynamic policy matching mechanism, so as to ensure that the live service adjustment strategy can accurately respond to the viewing preferences of different user groups and network fluctuation characteristics. By synchronizing the adjustment strategy to the live service node in real time and dynamically reconstructing the live parameter configuration, the problem of insufficient adaptability caused by fixed parameters in traditional live services is effectively solved, and a full-link real-time closed-loop control from data collection to strategy execution is realized. Particularly importantly, the real-time live quality metrics collected based on the adjusted live parameter configuration can reversely drive the parameter update of the dynamic optimization model, forming an intelligent optimization mechanism with self-evolution ability, which can continuously optimize with the change of viewing behavior patterns and the evolution of network environment, so as to achieve synergistic gains in multiple dimensions such as improving user viewing experience, enhancing service stability and reducing operation and maintenance costs. Description of the Drawings

[0014] Figure 1 It is a schematic flowchart of the execution process of the cinema live broadcast optimization method provided by the embodiment of the present invention.

[0015] Figure 2It is a schematic diagram of exemplary hardware and software components of a theater live broadcast optimization system based on big data analysis provided by an embodiment of the present invention. Detailed implementation manners

[0016] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 It is a flowchart of a method for optimizing theater live broadcast based on big data analysis provided by an embodiment of the present invention. The method for optimizing theater live broadcast based on big data analysis will be introduced in detail below.

[0017] Step S110: Obtain the historical live broadcast big data of the target theater. The historical live broadcast big data includes multiple live data streams, and each live data stream is composed of at least one user viewing behavior record and a corresponding live service status record.

[0018] In this embodiment, it is assumed that the target theater often conducts live screening activities of movies. Specifically, taking the live process of a science fiction movie as an example, the historical live broadcast big data can be obtained. Exemplarily, for each user viewing behavior record, such as the record of user A, it may include the time when he enters the live broadcast, pause operations, fast forward operations, volume adjustment operations, etc. during the live broadcast. The corresponding live service status record may include the network connection situation and the server load situation during the live broadcast. Assuming that 100 users watched this live broadcast, then there will be 100 live data streams composed of such user viewing behavior records and corresponding live service status records. Considering other movie live broadcasts carried out by the theater before, such as action movie live broadcasts and comedy movie live broadcasts, the relevant data of the above different movie live broadcasts together constitute the historical live broadcast big data including multiple live data streams.

[0019] Step S120: Perform multi-dimensional viewing feature extraction processing on the historical live broadcast big data to obtain the real-time viewing features and service stability features of each live data stream.

[0020] In this embodiment, continuing with the example of the science fiction movie live broadcast, first, the user viewing behavior records in the live data stream are processed by time window division. It is assumed that the entire live duration is divided into viewing behavior segments of consecutive time periods of every 10 minutes. For the viewing behavior segment of user A, a pre-trained real-time feature encoder can be called to process it.

[0021] Further, when generating the user interaction frequency feature, analyze the number of interaction operations and the operation interval duration of user A within a single 10-minute time window. For example, within a certain 10 minutes, user A performed 3 pause operations, and the interval durations of each pause operation were 2 minutes, 3 minutes, and 4 minutes respectively. The user interaction frequency feature is generated through these data.

[0022] Further, in terms of generating screen switching preference features, identify the temporal relationship between the screen switching request time points in the operation log of User A and the key frames of the live content. Suppose the set of screen switching request time points extracted from User A's operation log is [15 minutes, 25 minutes, 32 minutes]. The set of key frame time points of the live content, which includes scene switching key frames, special effect start key frames, and plot turning key frames, is parsed from the live content metadata as [10 minutes, 20 minutes, 30 minutes, 40 minutes]. Process each point in the set of screen switching request time points. For example, for the screen switching request time point of 15 minutes, use the binary search algorithm to locate the two closest candidate key frame time points in the ordered list of key frame time points, which are 10 minutes and 20 minutes. After calculating the bidirectional time difference, it is found that the difference from 10 minutes is the smallest, so 10 minutes is the target key frame time point. Calculate the absolute time difference between 15 minutes and 10 minutes, which is 5 minutes. This is the original time deviation amount. Since 15 minutes is later than 10 minutes, it is marked as a positive value according to the annotation rule. Process all screen switching request time points in this way to obtain a time deviation amount sequence with temporal direction identifiers. Then count the first frequency in the preset positive deviation interval, the second frequency in the preset negative deviation interval, and the third frequency in the zero deviation tolerance interval. Suppose the preset positive deviation interval is [3 minutes, 8 minutes], the negative deviation interval is [-8 minutes, -3 minutes], and the zero deviation tolerance interval is [-1 minute, 1 minute]. After counting, it is found that there are 2 times in the positive deviation interval, 1 time in the negative deviation interval, and 0 times in the zero deviation tolerance interval. According to these frequency ratio relationships, calculate the active switching tendency index for scene switching key frames, the delay following index for special effect start key frames, and the predicted switching density index for plot turning key frames. Suppose the total number of scene switching key frames is 5, the total number of special effect start key frames is 3, and the total number of plot turning key frames is 4. The calculated active switching tendency index is 2 / 5 = 0.4, the predicted switching density index is 1 / 3 = 0.33 (a simple weighted summation example here), and the delay following index is 0 / 4 = 0. Then, according to the live content type being a science fiction movie (a multi-scene switching type), configure a weight allocation strategy, increase the weight coefficient of the active switching tendency index, such as increasing it to 0.6, and perform a series of operations such as multiplying the adjusted weight coefficient with the corresponding index to generate a set of screen switching preference features.

[0023] For the content response delay feature, monitor the difference between the content loading request and the server response timestamp in User A's operation log, and count the number of abnormal events where the content loading delay exceeds a preset threshold (such as 3 seconds). Suppose that during the entire live broadcast, the number of abnormal events where the content loading delay exceeds 3 seconds is 2 times, and generate the content response delay feature based on this.

[0024] Meanwhile, perform anomaly event detection and processing on the live service status records. Suppose during a live broadcast, the timestamp of a live service interruption event is identified as 30 minutes after the start of the live broadcast, and the trigger condition for the service degradation event is that the network bandwidth is lower than 10 Mbps for 5 consecutive minutes. Conduct a stability quantification analysis on the live service status records, extract the network bandwidth sampling value sequence associated with the timestamp of the live service interruption event from the live service status records. For example, the bandwidth sampling data from 5 minutes before to 5 minutes after the interruption event is [8 Mbps, 9 Mbps, 7 Mbps, 6 Mbps, 5 Mbps, 4 Mbps, 5 Mbps, 6 Mbps, 7 Mbps, 8 Mbps]. Calculate the ratio of the standard deviation to the mean of these data to generate the bandwidth fluctuation feature. Analyze the video decoding log segments in the live service status records that match the trigger condition of the service degradation event, identify the video frame data transmitted during the effective period of the trigger condition of the service degradation event, and count the percentage of the number of frames with decoding error flags in the total number of transmitted frames to generate the decoding error rate feature. Traverse the hardware resource usage records in the live service status records, locate the central processor occupancy data points that overlap with the timestamp of the service interruption event, extract the continuous time periods that exceed the preset safety threshold (such as 80%), and calculate the weighted difference between the maximum value and the average value of the central processor occupancy rate during this continuous time period to generate the resource occupancy peak feature. Finally, perform time series alignment processing on these features to generate a second feature set containing the bandwidth fluctuation feature, the decoding error rate feature, and the resource occupancy peak feature. Fuse the first feature set and the second feature set to obtain the real-time viewing features and service stability features of each live data stream.

[0025] Step S130: Based on the pre-trained dynamic optimization model, perform dynamic policy matching processing on the real-time viewing features and the service stability features to generate an optimized annotation result for the live data stream, and generate a live service adjustment policy according to the optimized annotation result.

[0026] In this embodiment, continuing with the relevant real-time viewing features and service stability features of the previous science fiction movie live broadcast as an example, the feature vectors are concatenated to generate a multi-dimensional feature input vector. Assume that the historical feature weight template stored in the policy matching layer of the dynamic optimization model contains the benchmark weight values of each feature dimension under different live broadcast scenarios. For example, the benchmark weight value for the user interaction frequency feature is 0.2, the benchmark weight value for the picture switching preference feature is 0.3, the benchmark weight value for the content response delay feature is 0.1, the benchmark weight value for the bandwidth fluctuation feature is 0.2, the benchmark weight value for the decoding error rate feature is 0.1, and the benchmark weight value for the resource occupancy peak feature is 0.1. Calculate the cosine similarity between each feature dimension in the multi-dimensional feature input vector and the historical feature weight template to generate a scene matching degree index. Assume that after calculation, the benchmark weight values are dynamically adjusted according to the scene matching degree index. The weight of the user interaction frequency feature is adjusted to 0.15, the weight of the picture switching preference feature is adjusted to 0.35, the weight of the content response delay feature is adjusted to 0.12, the weight of the bandwidth fluctuation feature is adjusted to 0.18, the weight of the decoding error rate feature is adjusted to 0.08, and the weight of the resource occupancy peak feature is adjusted to 0.12. Then, normalization processing is performed to obtain the feature importance distribution. The multi-dimensional feature input vector is weighted and fused according to the feature importance distribution to generate an optimized decision vector. The optimized decision vector is input into the annotation output layer of the dynamic optimization model to generate an optimized annotation result including the picture quality optimization direction (such as improving picture clarity), the bandwidth allocation optimization direction (such as increasing bandwidth allocation), and the decoding priority optimization direction (such as increasing the decoding priority of special effect scenes). A live broadcast service adjustment strategy is generated according to these optimized annotation results, such as adjusting the parameters of the video encoder to improve the picture quality, adjusting the parameters of the bandwidth allocation module to increase the bandwidth, and adjusting the parameters of the decoding module to increase the decoding priority of special effect scenes.

[0027] Step S140: Synchronize the live broadcast service adjustment strategy to the live broadcast service node in real time, and trigger the live broadcast service node to adjust the current live broadcast parameter configuration according to the live broadcast service adjustment strategy.

[0028] In this embodiment, for the live service adjustment strategy generated above, continue to analyze it using the science fiction movie live scenario as an example. Assume that the live service adjustment strategy includes picture resolution adjustment parameters (such as adjusting from 720p to 1080p), bitrate control parameters (such as the maximum allowed bitrate is 5Mbps and the minimum guaranteed bitrate is 2Mbps), and buffer interval configuration parameters (such as adjusting the initial buffer threshold from 10 seconds to 15 seconds). Send the picture resolution adjustment parameters to the video encoding module to trigger the video encoder to dynamically adjust the resolution preset value from 720p to 1080p. Synchronize the bitrate control parameters to the bandwidth allocation module. During this process, monitor the real-time network throughput and packet loss rate of the live service node. Assume the real-time network throughput is 3Mbps and the packet loss rate is 1%. According to the maximum allowed bitrate and the minimum guaranteed bitrate in the bitrate control parameters, calculate the recommended bitrate range for the current network condition as [2Mbps, 3Mbps]. Call the bitrate adaptive controller to dynamically adjust the video transmission bitrate within this recommended bitrate range so that the actual bitrate maintains a preset proportional relationship with the network throughput, such as adjusting the actual bitrate to 2.5Mbps. When it is detected that the packet loss rate exceeds the threshold (such as 3%), an emergency bitrate reduction operation will be triggered and the forward error correction coding mechanism will be enabled. Deploy the buffer interval configuration parameters to the data cache module to trigger the adjustment of the initial buffer threshold according to the performance of the user device, such as adjusting the initial buffer threshold from 10 seconds to 15 seconds to ensure smooth playback.

[0029] Step S150, collect real-time live quality metrics based on the adjusted live parameter configuration, and update the parameters of the dynamic optimization model according to the real-time live quality metrics.

[0030] In this embodiment, during the live broadcast of a science fiction movie, real-time live broadcast quality metrics can be collected based on the adjusted live broadcast parameter configuration. Specifically, within a preset observation period (such as the remaining duration of the entire live broadcast), the playback smoothness metric, picture quality score, and interaction response delay metric of the user side are collected. The frequency of stuttering events in the playback smoothness metric is statistically analyzed. Assuming that within the remaining 30 minutes of the live broadcast, stuttering events occur 2 times, a stuttering frequency feature of 2 / 30 (the concept of the number of stuttering times per minute) is generated. Time series analysis is performed on the picture quality score. Assuming that the audience's ratings of the picture quality at different time periods during the live broadcast are [8 points, 7 points, 8 points, 9 points], a quality fluctuation feature is generated by analyzing this data. Percentile calculation is performed on the interaction response delay metric. Assuming that the 90% percentile of the interaction response delay is calculated to be 2 seconds, a delay distribution feature is generated. The difference between the stuttering frequency feature and the historical stuttering baseline data (assuming the historical stuttering baseline is 0.1 stuttering per minute) is calculated to generate the first loss component. The degree of deviation between the quality fluctuation feature and the expected quality stability curve (assuming the expected quality stability curve is that the score fluctuation is within 1 point) is analyzed to generate the second loss component. The proportion of high-delay samples (assuming high delay is defined as a delay greater than 3 seconds) in the delay distribution feature is statistically analyzed to generate the third loss component. The first loss component, the second loss component, and the third loss component are weighted and summed to generate a comprehensive loss function. Assuming that the weight of the first loss component is 0.3, the weight of the second loss component is 0.4, and the weight of the third loss component is 0.3, the comprehensive loss function is calculated. Then, the partial derivative of the comprehensive loss function with respect to the model weights is calculated through the backpropagation algorithm to generate the model parameter adjustment gradient, and the weight parameters of the dynamic optimization model are updated based on this model parameter adjustment gradient so that more accurate optimization operations can be performed in subsequent live broadcasts.

[0031] Based on the above steps, the embodiments of the present application achieve an overall performance improvement of the theater live broadcast service in a complex dynamic environment. Taking historical live broadcast big data as the analysis basis, by integrating multi-dimensional viewing features and service stability features, the description of the live broadcast service state has spatio-temporal correlation and behavior perception capabilities. The pre-trained dynamic optimization model, as the core decision-making engine, not only realizes the intelligent mapping of real-time viewing features and service stability features, but also generates forward-looking optimization annotation results through a dynamic policy matching mechanism, so as to ensure that the live broadcast service adjustment strategy can accurately respond to the viewing preferences of different user groups and the characteristics of network fluctuations. By synchronizing the adjustment strategy to the live broadcast service node in real time and dynamically reconstructing the live broadcast parameter configuration, it effectively solves the problem of insufficient adaptability caused by fixed parameters in traditional live broadcast services, and realizes the full-link real-time closed-loop control from data collection to strategy execution. Particularly importantly, the real-time live broadcast quality indicators collected based on the adjusted live broadcast parameter configuration can reversely drive the parameter update of the dynamic optimization model, forming an intelligent optimization mechanism with self-evolution ability, which can continuously optimize with the change of viewing behavior patterns and the evolution of network environment, so as to achieve collaborative gains in multiple dimensions such as improving user viewing experience, enhancing service stability and reducing operation and maintenance costs.

[0032] In a possible implementation manner, step S120 includes:

[0033] Step S121, performing time window division processing on the user viewing behavior records in the live broadcast data stream to obtain viewing behavior segments in multiple consecutive time periods.

[0034] For example, if the entire live broadcast duration of a science fiction movie is 120 minutes, it can be divided into consecutive time periods of 15 minutes each, and 8 viewing behavior segments can be generated. Each viewing behavior segment covers various viewing operation information of users within these 15 minutes.

[0035] Step S122, calling a pre-trained real-time feature encoder to extract real-time viewing features from the viewing behavior segments, and generating a first feature set including user interaction frequency features, picture switching preference features, and content response delay features.

[0036] Step S123, performing abnormal event detection processing on the live broadcast service state record to identify the time stamp of the live broadcast service interruption event and the triggering conditions of the service degradation event.

[0037] For example, during the live broadcast of a science fiction movie, the timestamp of the live broadcast service interruption event is identified as 45 minutes after the start of the live broadcast, and the trigger condition for the service degradation event is that the network bandwidth is lower than 10 Mbps for 5 minutes. Thus, based on this timestamp and trigger condition, a stability quantification analysis can be performed on the live broadcast service status record. Specifically, a sequence of network bandwidth sampling values associated with the timestamp of the live broadcast service interruption event can be extracted from the live broadcast service status record. For example, the bandwidth sampling data from 5 minutes before to 5 minutes after the occurrence of the interruption event is [9 Mbps, 8 Mbps, 7 Mbps, 6 Mbps, 5 Mbps, 4 Mbps, 5 Mbps, 6 Mbps, 7 Mbps, 8 Mbps]. Calculate the ratio of the standard deviation to the mean of these data to generate the bandwidth fluctuation feature. First, calculate the mean. The sum of these data is 65 Mbps, and then divide by the number of data, which is 10, to get a mean of 6.5 Mbps. Then calculate the square of the difference between each data and the mean. For example, the difference between the first data 9 Mbps and the mean 6.5 Mbps is 2.5 Mbps, and its square is 6.25 Mbps². Perform such calculations and sum them for all data to get a total of 35 Mbps². Then divide by the number of data, which is 10, to get a variance of 3.5 Mbps². The standard deviation is the square root of the variance, approximately 1.87 Mbps. Finally, calculate the ratio of the standard deviation to the mean. 1.87 Mbps divided by 6.5 Mbps is approximately 0.29, which is the bandwidth fluctuation feature.

[0038] Step S124: Based on the timestamp and the trigger condition, perform a stability quantification analysis on the live broadcast service status record to generate a second feature set including a bandwidth fluctuation feature, a decoding error rate feature, and a peak resource occupancy feature.

[0039] Step S125: Perform feature fusion on the first feature set and the second feature set to obtain the real-time viewing feature and the service stability feature.

[0040] In this embodiment, the first feature set including the user interaction frequency feature, the picture switching preference feature, and the content response delay feature and the second feature set including the bandwidth fluctuation feature, the decoding error rate feature, and the peak resource occupancy feature are subjected to feature fusion to obtain the real-time viewing feature and the service stability feature. For example, each feature in the first feature set and the second feature set can be combined according to a set rule, so that the finally obtained real-time viewing feature and service stability feature can comprehensively reflect the user's viewing experience and the stability status of the service.

[0041] In a possible implementation manner, step S122 includes:

[0042] Step S1221: Invoke a pre-trained real-time feature encoder to parse the user operation log in the viewing behavior segment, extract the number of interaction operations and the operation interval duration of the user within a single time window, and generate user interaction frequency features.

[0043] For example, taking a 15-minute viewing behavior segment as mentioned above, by checking the user operation log, it can be seen that the user performed 3 pause operations, 1 fast-forward operation, and 2 volume adjustment operations within this time period. Calculate the total number of these operations: 3 pauses + 1 fast-forward + 2 volume adjustments = 6 operations, which constitutes the number of interaction operations of the user within this 15-minute time window. Looking at the operation interval duration again, the first pause operation was at the 3rd minute, and the second pause operation was at the 7th minute. Then the operation interval duration from the first pause to the second pause is 7 - 3 = 4 minutes; the operation interval duration from the second pause to the third pause operation (assuming the third pause is at the 12th minute) is 12 - 7 = 5 minutes, etc. These operation interval duration data together constitute the operation interval duration part in the user interaction frequency features. In this way, the user interaction frequency features are completely generated, and these user interaction frequency features can reflect the operation activity level of the user during the viewing process.

[0044] Step S1222: Identify the temporal relationship between the time points of the screen switching requests and the key frames of the live content in the user operation log, calculate the time deviation amount between the screen switching requests and the content key frames, and generate screen switching preference features.

[0045] In a possible implementation manner, step S1222 includes:

[0046] Step S1222-1: Extract the timestamp data of the trigger moments of the screen switching operations from the user operation log to generate a set of screen switching request time points.

[0047] For example, within the above-mentioned 15-minute viewing behavior segment, it is found from the user operation log that the trigger moments of the screen switching operations are the 5th minute, the 10th minute, and the 13th minute respectively. Then the set of screen switching request time points is [5 minutes, 10 minutes, 13 minutes].

[0048] Step S1222-2: Parse the pre-annotated key frame types and the corresponding timestamp sequences in the live content metadata to generate a set of live content key frame time points including scene switching key frames, special effect start key frames, and plot turning key frames.

[0049] In this embodiment, it is assumed that the timestamps of the scene transition key frames parsed from the live content metadata are the 4th minute and the 8th minute; the timestamps of the special effect start key frames are the 6th minute and the 11th minute; the timestamps of the plot turning key frames are the 9th minute and the 14th minute. Then the set of key frame time points of the live content is [4 minutes, 6 minutes, 8 minutes, 9 minutes, 11 minutes, 14 minutes].

[0050] Step S1222-3: Traverse each screen switching request time point in the set of screen switching request time points, and search for the target key frame time point with the smallest time difference from the current screen switching request time point in the set of key frame time points of the live content.

[0051] In this embodiment, taking the screen switching request time point of 5 minutes as an example, calculate the difference between it and each time point in the set of key frame time points of the live content. The difference between 5 minutes and 4 minutes is 5 - 4 = 1 minute; the difference between 5 minutes and 6 minutes is 6 - 5 = 1 minute; the difference between 5 minutes and 8 minutes is 8 - 5 = 3 minutes; the difference between 5 minutes and 9 minutes is 9 - 5 = 4 minutes; the difference between 5 minutes and 11 minutes is 11 - 5 = 6 minutes; the difference between 5 minutes and 14 minutes is 14 - 5 = 9 minutes. It can be seen that the differences from 4 minutes and 6 minutes are the smallest and equal. According to the rule, 6 minutes with a later time sequence is preferentially selected as the target key frame time point. The same operation is also performed on the screen switching request time points of 10 minutes and 13 minutes.

[0052] Step S1222-4: Calculate the absolute time difference between the screen switching request time point and the target key frame time point, and generate the original time deviation amount of a single screen switching request time point.

[0053] For example, for the screen switching request time point of 5 minutes and the target key frame time point of 6 minutes, its absolute time difference is 6 - 5 = 1 minute, and thus it is output as the original time deviation amount of this screen switching request time point.

[0054] Step S1222-5: Perform positive and negative sign annotation processing on the original time deviation amount. When the screen switching request time point is later than the target key frame time point, it is annotated as positive, and when it is earlier than the target key frame time point, it is annotated as negative, to generate a time deviation amount sequence with a time sequence direction identifier.

[0055] For example, since 5 minutes is earlier than 6 minutes, it is annotated as -1 minute. In this way, the original time deviation amounts of all screen switching request time points are annotated to obtain a time deviation amount sequence with a time sequence direction identifier.

[0056] Step S1222-6: Count the first frequency in the preset positive deviation interval, the second frequency in the preset negative deviation interval, and the third frequency in the zero deviation tolerance interval in the time deviation amount sequence.

[0057] In this embodiment, assume that the preset positive deviation interval is [3 minutes, 8 minutes], the negative deviation interval is [-8 minutes, -3 minutes], and the zero deviation tolerance interval is [-1 minute, 1 minute]. After counting, it is found that the frequency in the positive deviation interval is 0 times, the frequency in the negative deviation interval is 1 time (which is the previously calculated deviation amount of -1 minute), and the frequency in the zero deviation tolerance interval is 2 times (assuming there is another deviation amount in this interval).

[0058] Step S1222-7: Generate an active switching tendency index for the user for the scene transition key frames, a delay following index for the special effect start key frames, and a predicted switching density index for the plot turning key frames according to the proportional relationship of the first frequency, the second frequency, and the third frequency.

[0059] In this embodiment, assume that the total number of scene transition key frames is 5, the total number of special effect start key frames is 3, and the total number of plot turning key frames is 4. Calculate the active switching tendency index for the user for the scene transition key frames. Divide the frequency 0 in the positive deviation interval by the total number of scene transition key frames 5, and 0 divided by 5 equals 0. Calculate the delay following index for the special effect start key frames. Weight and sum the frequency 1 in the negative deviation interval with the total number of special effect start key frames 3 (assuming the weighting coefficient is 1 here), and 1 divided by 3 is approximately equal to 0.33. Calculate the predicted switching density index for the plot turning key frames. Perform a dynamic proportional mapping on the frequency 2 in the zero deviation tolerance interval and the total number of plot turning key frames 4 (assuming the mapping relationship is direct division here), and 2 divided by 4 equals 0.5.

[0060] Step S1222-8: Allocate weights to the active switching tendency index, the delay following index, and the predicted switching density index according to the key frame type, and generate a set of picture switching preference features including time sensitivity weights and content relevance weights.

[0061] For example, since it is a live broadcast of a science fiction movie, which belongs to the multi-scene switching type, the weight coefficient of the active switching tendency index is increased according to the weight distribution strategy, such as increasing it to 0.6. For the starting key frame of special effects, assuming that the weight coefficient of the delay following index is adjusted to 0.2 according to the special effects situation in the live content. For the key frame of plot turning point, the weight coefficient of the predicted switching density index is adjusted to 0.2. Then, the adjusted weight coefficients are multiplied by the corresponding indexes. The active switching tendency index multiplied by its weight coefficient is 0 multiplied by 0.6 equals 0; the delay following index multiplied by its weight coefficient is 0.33 multiplied by 0.2 equals 0.066; the predicted switching density index multiplied by its weight coefficient is 0.5 multiplied by 0.2 equals 0.1. Combining these results generates a set of picture switching preference features including time sensitivity weight and content relevance weight, and this set of picture switching preference features can reflect the preference characteristics of users for picture switching of different types of key frames.

[0062] Step S1223, monitor the difference between the content loading request and the server response timestamp in the user operation log, count the number of abnormal events where the content loading delay exceeds the preset threshold, and generate a content response delay feature.

[0063] For example, within the above-mentioned 15-minute movie-watching behavior segment, check the timestamps of the content loading request and the server response in the user operation log. Assume that the content loading request occurs at the 2nd minute and the server response timestamp is at the 6th minute. Therefore, the difference between them is 6 - 2 = 4 minutes. Assume that the preset threshold is 3 minutes, then this is an event where the content loading delay exceeds the preset threshold. Continue to check all content loading requests and server response situations within the entire 15 minutes, count the number of such abnormal events. Assume that a total of 2 times are counted, and these 2 times constitute the content response delay feature, and this content response delay feature can reflect the delay situation of the user when obtaining content.

[0064] In a possible implementation manner, step S1222-2 includes:

[0065] Step S1222-21, read the metadata tag block embedded in the live content video stream, and extract the frame number, frame type label, and offset relative to the live start time of each key frame.

[0066] For example, in the live broadcast of a science fiction movie, assume that the metadata tag block in the video stream contains a series of key frame information. For one of the key frames, its frame number is 100, the frame type label is "scene switching", and the offset relative to the live start time is 5 minutes. In this way, information extraction is performed on all key frames.

[0067] Step S1222-22: Divide the key frames into scene transition key frames, special effect start key frames, and plot turning key frames according to the frame type tags.

[0068] For example, among the numerous key frames extracted, the key frames with the frame type tag of "scene transition" are classified into the category of scene transition key frames; the key frames with the frame type tag of "special effect start" are classified into the category of special effect start key frames; the key frames with the frame type tag of "plot turning" are classified into the category of plot turning key frames.

[0069] Step S1222-23: Convert the frame numbers of each key frame into absolute timestamp data based on the offset, and generate a set of key frame time points of the live content arranged in chronological order.

[0070] For example, assume that the offset of the first scene transition key frame is 5 minutes, then its absolute timestamp is 5 minutes after the start of the live broadcast; the offset of the second scene transition key frame is 12 minutes, and its absolute timestamp is 12 minutes. Such conversions are performed on all types of key frames, and then the timestamps of these key frames are arranged in chronological order to form a set of key frame time points of the live content. For example, this set of key frame time points of the live content may be [5 minutes (scene transition key frame), 8 minutes (special effect start key frame), 12 minutes (scene transition key frame), 15 minutes (plot turning key frame), 20 minutes (special effect start key frame)].

[0071] Step S1222-24: Verify the timestamp continuity of adjacent key frames in the set of key frame time points of the live content. If it is detected that the timestamp jump exceeds the preset frame interval threshold, insert virtual key frame time points in the jump interval and label them as unclassified key frame types.

[0072] For example, assume that the preset frame interval threshold is 3 minutes. When checking the set of key frame time points of the live content, it is found that there is a timestamp jump between the scene transition key frame at 12 minutes and the plot turning key frame at 15 minutes. 15 - 12 = 3 minutes, just reaching the threshold. If the jump exceeds 3 minutes, insert virtual key frame time points in this interval and label them as unclassified key frame types.

[0073] For example, in a possible implementation, step S1222-3 includes:

[0074] Step S1222-31: Perform a time ascending sorting process on the set of key frame time points of the live content to generate an ordered list of key frame time points.

[0075] Continuing with the set of key frame time points of the live content mentioned above [5 minutes (scene transition key frame), 8 minutes (special effect start key frame), 12 minutes (scene semana transition key frame), 15 minutes (plot twist key frame), 20 minutes (special effect start key frame)] as an example, after sorting the time points in ascending order, the ordered list of key frame time points remains [5 minutes (scene transition key frame), 8 minutes (special effect start key frame), 12 minutes (scene transition key frame), 15 minutes (plot twist key frame), 20 minutes (special effect start key frame)].

[0076] Step S1222 - 32, use the binary search algorithm to locate the two candidate key frame time points closest to the current video frame switching request time point in the ordered list of key frame time points.

[0077] In this embodiment, assuming that the video frame switching request time point is 11 minutes, first take the middle element of the ordered list of key frame time points, that is, 12 minutes (scene transition key frame). Since 11 minutes is less than 12 minutes, then take the middle element in the first half of the list [5 minutes (scene transition key frame), 8 minutes (special effect start key frame)], which is 8 minutes (special effect start key frame). In this way, the two candidate key frame time points closest to 11 minutes, namely 8 minutes (special effect start key frame) and 12 minutes (scene transition key frame), are located.

[0078] Step S1222 - 33, calculate the two-way time differences between the current video frame switching request time point and the two candidate key frame time points, and select the candidate key frame time point with the smallest difference as the target key frame time point.

[0079] For example, for the video frame switching request time point of 11 minutes, the difference from 8 minutes (special effect start key frame) is 11 - 8 = 3 minutes; the difference from 12 minutes (scene transition key frame) is 12 - 11 = 1 minute. Since 1 minute is less than 3 minutes, 12 minutes (scene transition key frame) is selected as the target key frame time point.

[0080] Step S1222 - 34, when the differences between the two candidate key frame time points and the current video frame switching request time point are equal, preferentially select the candidate key frame time point with a later time order as the target key frame time point.

[0081] For example, if the screen switching request time point is 9 minutes, the difference from 8 minutes (the starting key frame of the special effect) is 9 - 8 = 1 minute, and the difference from 12 minutes (the scene switching key frame) is 12 - 9 = 3 minutes. However, if the differences are equal, for example, the screen switching request time point is 10 minutes, the difference from 8 minutes (the starting key frame of the special effect) is 10 - 8 = 2 minutes, and the difference from 12 minutes (the scene switching key frame) is 12 - 10 = 2 minutes. At this time, 12 minutes (the scene switching key frame) is preferentially selected as the target key frame time point.

[0082] For example, in a possible implementation manner, step S1222-5 includes:

[0083] Step S1222-51, establish a time deviation amount annotation rule library, and define that a positive value indicates that the user operation lags behind the key frame event, and a negative value indicates that the user operation is ahead of the key frame event.

[0084] For example, when processing the time deviation amount between the screen switching request time point and the target key frame time point, it is annotated according to this time deviation amount annotation rule library.

[0085] Step S1222-52, according to the sequence of the target key frame time point and the screen switching request time point, add a symbol identifier to each original time deviation amount.

[0086] Suppose the screen switching request time point is 11 minutes and the target key frame time point is 12 minutes. Since 11 minutes is earlier than 12 minutes, the original time deviation amount is 12 - 11 = 1 minute, and it is marked as -1 minute according to the rule, indicating that the user operation is ahead of the key frame event.

[0087] Step S1222-53, calculate the absolute value of the annotated time deviation amount to generate a time deviation amount absolute value sequence and a symbol identifier sequence.

[0088] For the time deviation amount marked as -1 minute above, its absolute value is 1 minute. In this way, the time deviation amount absolute value sequence is [1 minute], and the symbol identifier sequence is [-].

[0089] Step S1222-54, associate and store the time deviation amount absolute value sequence and the symbol identifier sequence in chronological order to generate a time deviation amount sequence with a chronological direction identifier.

[0090] For example, in chronological order, the absolute values and sign identifications of the time deviation amounts corresponding to multiple screen switching request time points are associated, such as [-1 minute, 2 minutes, -3 minutes]. Here, -1 minute means that the first screen switching request time point is 1 minute ahead of the target key frame time point, 2 minutes means that the second screen switching request time point is 2 minutes behind the target key frame time point, and -3 minutes means that the third screen switching request time point is 3 minutes ahead of the target key frame time point.

[0091] For example, in a possible implementation manner, step S1222-6 includes:

[0092] Step S1222-61, obtain the upper limit value and the lower limit value of a preset positive deviation interval, and count the number of events where the absolute value of the time deviation amount is within this interval and the sign is positive as the first frequency.

[0093] For example, assume that the preset positive deviation interval is [3 minutes, 8 minutes]. In the time deviation amount sequence [-1 minute, 2 minutes, -3 minutes, 5 minutes, -2 minutes, 7 minutes], only the event of 7 minutes has an absolute value within [3 minutes, 8 minutes] and a positive sign. Therefore, the first frequency is 1.

[0094] Step S1222-62, obtain the upper limit value and the lower limit value of a preset negative deviation interval, and count the number of events where the absolute value of the time deviation amount is within this interval and the sign is negative as the second frequency.

[0095] For example, assume that the preset negative deviation interval is [-8 minutes, -3 minutes]. In the above time deviation amount sequence, only the event of -3 minutes has an absolute value within [-8 minutes, -3 minutes] and a negative sign. Therefore, the second frequency is 1.

[0096] Step S1222-63, define the zero deviation tolerance interval as the event where the absolute value of the time deviation amount is less than or equal to a preset error threshold, and count the number of events that meet the condition as the third frequency.

[0097] For example, assume that the preset error threshold is 1 minute. In the time deviation amount sequence, only the event of -1 minute has an absolute value less than or equal to 1 minute. Therefore, the third frequency is 1.

[0098] Step S1222-64, exclude the abnormal deviation events in the time deviation amount sequence that do not fall into the above three intervals, and generate a frequency statistics result after cleaning.

[0099] In this example, 2 minutes and -2 minutes do not fall into the above three intervals. Therefore, when generating the frequency statistics result after cleaning, these two events are not considered. The frequency statistics result after cleaning is that the first frequency is 1, the second frequency is 1, and the third frequency is 1.

[0100] For example, in a possible implementation, step S1222-7 includes:

[0101] Step S1222-71: Calculate the ratio of the first frequency to the total number of scene transition key frames to generate an active switching tendency index reflecting the user's initiative to initiate a switching operation after a scene transition.

[0102] For example, assume that the total number of scene transition key frames is 5 and the first frequency is 1. Then the active switching tendency index is 1 divided by 5, which equals 0.2.

[0103] Step S1222-72: Perform a weighted sum of the second frequency and the total number of special effect start key frames to generate a predicted switching density index reflecting the user's early switching before the special effect starts.

[0104] For example, assume that the total number of special effect start key frames is 3 and the second frequency is 1. For the weighted sum (assuming a weighting coefficient of 1 here), then the predicted switching density index is 1 divided by 3, approximately equal to 0.33.

[0105] Step S1222-73: Perform a dynamic proportional mapping of the third frequency to the total number of plot turning point key frames to generate a delayed following index reflecting the user's switching following the plot turning point.

[0106] For example, assume that the total number of plot turning point key frames is 4 and the third frequency is 1. Assuming the dynamic proportional mapping is direct division here, then the delayed following index is 1 divided by 4, which equals 0.25.

[0107] Step S1222-74: Normalize the active switching tendency index, predicted switching density index, and delayed following index to eliminate the statistical bias caused by the difference in the number of key frame types.

[0108] For example, assume that the active switching tendency index is 0.2, the predicted switching density index is 0.33, and the delayed following index is 0.25. Then, first calculate their sum as 0.2 + 0.33 + 0.25 = 0.78. Then divide each index by the sum. The normalized active switching tendency index is 0.2 divided by 0.78, approximately equal to 0.26. The normalized predicted switching density index is 0.33 divided by 0.78, approximately equal to 0.42. The normalized delayed following index is 0.25 divided by 0.78, approximately equal to 0.32.

[0109] For example, in a possible implementation, step S1222-8 includes:

[0110] Step S1222-81: Configure the weight assignment strategy according to the live content type. When it is detected that the live content is of the multi-scene switching type, increase the weight coefficient of the active switching tendency indicator.

[0111] For example, when it is detected that the live content is of the multi-scene switching type (a science fiction movie belongs to the multi-scene switching type), increase the weight coefficient of the active switching tendency indicator. Suppose the weight coefficient of the active switching tendency indicator is increased to 0.6.

[0112] Step S1222-82: When it is detected that the live content is of the significant special effect density type, increase the weight coefficient of the predicted switching density indicator.

[0113] Suppose special effects are more significant in a science fiction movie, and the weight coefficient of the predicted switching density indicator is increased to 0.25.

[0114] Step S1222-83: When it is detected that the live content is of the strong plot coherence type, increase the weight coefficient of the latency following indicator.

[0115] Suppose the plot coherence of a science fiction movie is strong, and the weight coefficient of the latency following indicator is increased to 0.15.

[0116] Step S1222-84: Perform a multiplication operation on the adjusted weight coefficient and the corresponding indicator to generate a time sensitivity weight vector.

[0117] The active switching tendency indicator multiplied by its weight coefficient is 0.26 multiplied by 0.6 = 0.156; the predicted switching density indicator multiplied by its weight coefficient is 0.42 multiplied by 0.25 = 0.105; the latency following indicator multiplied by its weight coefficient is 0.32 multiplied by 0.15 = 0.048. The time sensitivity weight vector is [0.156, 0.105, 0.048].

[0118] Step S1222-85: Calculate the content correlation weight vector according to the switching success rate of different key frame types in the user's historical viewing data.

[0119] Suppose it is analyzed from the user's historical viewing data that the switching success rate of scene switching key frames is 80%, the switching success rate of special effect start key frames is 70%, and the switching success rate of plot turning key frames is 60%. Normalize these success rates, and the sum is 80% + 70% + 60% = 210%. The content correlation weight of the normalized scene switching key frame is approximately 0.38 (80% divided by 210%), the content correlation weight of the special effect start key frame is approximately 0.33 (70% divided by 210%), and the content correlation weight of the plot turning key frame is approximately 0.29 (60% divided by 210%). The content correlation weight vector is [0.38, 0.33, 0.29].

[0120] Step S1222-86, perform a dot product operation on the time sensitivity weight vector and the content association degree weight vector to generate a set of screen switching preference features.

[0121] For example, the calculation process is 0.156 multiplied by 0.38 + 0.105 multiplied by 0.33 + 0.048 multiplied by 0.29 = 0.05928 + 0.03465 + 0.01392 = 0.10785. The set of screen switching preference features is jointly composed of this calculation result and related weight and index information, and this set of screen switching preference features can comprehensively reflect the user's preference features in terms of screen switching.

[0122] In a possible implementation manner, step S124 includes:

[0123] Step S1241, extract a sequence of network bandwidth sampling values associated with the timestamp of the live service interruption event from the live service status record, filter out the bandwidth sampling data within a set time range before and after each live service interruption event occurs, and calculate the ratio of the standard deviation to the mean of the bandwidth sampling data to generate a bandwidth fluctuation feature.

[0124] For example, during the live broadcast of a science fiction movie, assume that a live service interruption event occurs, and its timestamp is 45 minutes after the start of the live broadcast. Set a time range of 5 minutes before and after the occurrence of the live service interruption event to extract the bandwidth sampling data, that is, the time period from 40 minutes to 50 minutes. The sequence of bandwidth sampling values obtained from the live service status record for this time period is [9Mbps, 8Mbps, 7Mbps, 6Mbps, 5Mbps, 4Mbps, 5Mbps, 6Mbps, 7Mbps, 8Mbps]. When calculating the mean, add up the above data to get a total of 65Mbps, and then divide by the number of data 10 to get a mean of 6.5Mbps. Then calculate the square of the difference between each data and the mean. For example, the difference between the first data 9Mbps and the mean 6.5Mbps is 2.5Mbps, and its square is 6.25Mbps²; the difference between the second data 8Mbps and the mean 6.5Mbps is 1.5Mbps, and its square is 2.25Mbps², and so on for all data, calculate and sum them to get a total of 35Mbps², then divide by the number of data 10 to get a variance of 3.5Mbps², and the standard deviation is the square root of the variance, approximately 1.87Mbps. Finally, calculate the ratio of the standard deviation to the mean, 1.87Mbps divided by 6.5Mbps is approximately equal to 0.29, and the output is the bandwidth fluctuation feature, which can reflect the fluctuation of the network bandwidth near the live service interruption event.

[0125] Step S1242: Analyze the video decoding log segment in the live service status record that matches the trigger condition of the service degradation event, identify the video frame data transmitted during the effective period of the trigger condition of the service degradation event, count the percentage of the number of frames carrying decoding error flags in the video frame data to the total number of transmitted frames, and generate a decoding error rate feature.

[0126] For example, assume that the trigger condition of the service degradation event is that the network bandwidth is lower than 10 Mbps for 5 consecutive minutes, and this condition is met during a certain period of the live broadcast. Analyze the video decoding log segment that matches the trigger condition of this service degradation event, and identify the video frame data transmitted during the effective period of the trigger condition of this service degradation event. Assume that a total of 100 video frames are transmitted during this period. By checking the video decoding log segment, it is counted that the number of frames carrying decoding error flags is 5. Calculate the percentage of the number of frames carrying decoding error flags in the video frame data to the total number of transmitted frames, that is, 5 frames divided by 100 frames equals 5%, and the output is the decoding error rate feature, which can reflect the proportion of video decoding errors during service degradation.

[0127] Step S1243: Traverse the hardware resource usage record in the live service status record, locate the central processing unit occupancy data points that overlap with the time stamp of the service interruption event, extract the continuous time periods in the data points that exceed the preset safety threshold, and calculate the weighted difference between the maximum value and the average value of the central processing unit occupancy rate within the continuous time periods to generate a resource occupancy peak feature.

[0128] For example, in the service interruption event 45 minutes after the start of the live broadcast mentioned above, traverse the hardware resource usage record in the live service status record to find the central processing unit occupancy data points that overlap with this time stamp. Assume that these data points are [82%, 85%, 83%, 81%], the preset safety threshold is 80%, and the continuous time period exceeding 80% is [82%, 85%, 83%]. Calculate the average value within this continuous time period. Add these three data to get a total of 250%, and then divide by 3 to get an average value of approximately 83.3%, and the maximum value is 85%. Here, assume that the weighting coefficient is 1, calculate the weighted difference between the maximum value and the average value, 85% - 83.3% = 1.7%, and the output is the resource occupancy peak feature, which can reflect the peak situation of the central processing unit resource occupancy during the live service interruption event.

[0129] Step S1244: Perform time series alignment processing on the bandwidth fluctuation feature, decoding error rate feature, and resource occupancy peak feature, so that the time intervals corresponding to the bandwidth fluctuation feature, decoding error rate feature, and resource occupancy peak feature are in a synchronous mapping relationship with the time stamp of the service interruption event and the trigger condition of the service degradation event, and generate a second feature set with time series tags.

[0130] For example, taking the live service interruption event 45 minutes after the start of the live broadcast mentioned above as an example, the bandwidth fluctuation feature (0.29) calculated near this timestamp, the decoding error rate feature (5%) obtained during the effective period of the service degradation event trigger condition, and the peak resource occupancy feature (1.7%) overlapping with the timestamp of this live service interruption event are aligned in time series. These features are synchronously mapped to the 45-minute timestamp and the trigger condition of the service degradation event (the network bandwidth is lower than 10 Mbps for 5 minutes). For example, it is marked that 45 minutes after the start of the live broadcast, when the network bandwidth is lower than 10 Mbps for 5 minutes, the bandwidth fluctuation feature is 0.29, the decoding error rate feature is 5%, and the peak resource occupancy feature is 1.7%. Through such time series alignment processing, a second feature set with time series tags is generated, and this second feature set can comprehensively reflect the service stability-related features under specific live service states (such as live service interruption and service degradation).

[0131] In a possible implementation manner, step S130 includes:

[0132] Step S131, splicing the real-time viewing features and the service stability features to generate a multi-dimensional feature input vector.

[0133] In this embodiment, the real-time viewing features include the user interaction frequency feature, the screen switching preference feature, the content response delay feature, etc., and the service stability features include the bandwidth fluctuation feature, the decoding error rate feature, the peak resource occupancy feature, etc. For example, assume that the value of the user interaction frequency feature is [0.3] (here it represents a value under a certain quantization standard), the value of the screen switching preference feature is [0.5], the value of the content response delay feature is [0.2], the value of the bandwidth fluctuation feature is [0.15], the value of the decoding error rate feature is [0.05], and the value of the peak resource occupancy feature is [0.1]. Splicing these features in order, the multi-dimensional feature input vector obtained is [0.3, 0.5, 0.2, 0.15, 0.05, 0.1].

[0134] Step S132, obtaining the historical feature weight template stored in the policy matching layer of the dynamic optimization model, where the historical feature weight template includes the benchmark weight values of each feature dimension under different live scenarios.

[0135] For example, in the live broadcast scenario of science fiction movies, assume that in the historical feature weight template, the benchmark weight value of the user interaction frequency feature is 0.2, the benchmark weight value of the screen switching preference feature is 0.3, the benchmark weight value of the content response delay feature is 0.1, the benchmark weight value of the bandwidth fluctuation feature is 0.2, the benchmark weight value of the decoding error rate feature is 0.1, and the benchmark weight value of the resource occupancy peak feature is 0.1.

[0136] Step S133: Calculate the cosine similarity between each feature dimension in the multi-dimensional feature input vector and the historical feature weight template to generate a scene matching degree index, dynamically adjust the benchmark weight value according to the scene matching degree index to generate a real-time feature weight value, and perform normalization processing on the real-time feature weight value to obtain a feature importance distribution.

[0137] Taking the user interaction frequency feature as an example, calculate its cosine similarity with the corresponding feature dimension in the historical feature weight template. Suppose after calculation (the specific calculation process involves vector calculation, and the result is described in words here), the cosine similarity of the user interaction frequency feature is 0.8. Calculate the cosine similarities of other feature dimensions in the same way. Suppose the cosine similarity of the screen switching preference feature is 0.9, the cosine similarity of the content response delay feature is 0.7, the cosine similarity of the bandwidth fluctuation feature is 0.85, the cosine similarity of the decoding error rate feature is 0.75, and the cosine similarity of the resource occupancy peak feature is 0.8. Generate a scene matching degree index based on these cosine similarities, and then dynamically adjust the benchmark weight value. For example, for the user interaction frequency feature, the adjusted weight value according to the scene matching degree index may become 0.18 (this is the result obtained according to the scene matching degree index and the set adjustment rules). Adjust the benchmark weight values of other features in the same way. The adjusted weight value of the screen switching preference feature is 0.32, the adjusted weight value of the content response delay feature is 0.09, the adjusted weight value of the bandwidth fluctuation feature is 0.19, the adjusted weight value of the decoding error rate feature is 0.08, and the adjusted weight value of the resource occupancy peak feature is 0.09. Normalize these adjusted weight values. Calculate their sum as 0.18 + 0.32 + 0.09 + 0.19 + 0.08 + 0.09 = 0.95. Then divide each adjusted weight value by the sum. The normalized weight value of the user interaction frequency feature is 0.18 divided by 0.95, approximately equal to 0.19. The weight value of the screen switching preference feature is 0.32 divided by 0.95, approximately equal to 0.34. The weight value of the content response delay feature is 0.09 divided by 0.95, approximately equal to 0.095. The weight value of the bandwidth fluctuation feature is 0.19 divided by 0.95, equal to 0.2. The weight value of the decoding error rate feature is 0.08 divided by 0.95, approximately equal to 0.084. The weight value of the resource occupancy peak feature is 0.09 divided by 0.95, approximately equal to 0.095. These normalized weight values constitute the feature importance distribution.

[0138] Step S134, perform weighted fusion on the multi-dimensional feature input vector according to the feature importance distribution to generate an optimized decision vector.

[0139] For example, according to the feature importance distribution obtained previously, each feature in the multi-dimensional feature input vector is multiplied by the corresponding weight value, and then added together to obtain an optimized decision vector. That is, (0.3×0.19)+(0.5×0.34)+(0.2×0.095)+(0.15×0.2)+(0.05×0.084)+(0.1×0.095) = 0.057+0.17+0.019+0.03+0.0042+0.0095 = 0.2897. Here, 0.2897 is a quantitative representation of the optimized decision vector (in practice, it may be a multi-dimensional vector, and here it is simplified to a numerical value to represent its calculation result).

[0140] Step S135, input the optimized decision vector into the annotation output layer of the dynamic optimization model to generate an optimized annotation result including the optimization direction of video quality, the optimization direction of bandwidth allocation, and the optimization direction of decoding priority.

[0141] For example, in the annotation output layer, the optimization direction of video quality is judged according to the value of the optimized decision vector. If the value of the optimized decision vector meets the set conditions (here, according to the predefined model rules), it may be determined that the video quality needs to improve the clarity, which is the optimization direction of video quality. For the optimization direction of bandwidth allocation, it may be determined according to the value of the optimized decision vector that the bandwidth allocation needs to be increased. For the optimization direction of decoding priority, it may be determined that the decoding priority of special effect scenes needs to be improved, etc. In this way, an optimized annotation result for the live data stream of science fiction movies is comprehensively generated, and these results will be used for formulating subsequent live service adjustment strategies.

[0142] In a possible implementation manner, step S140 includes:

[0143] Step S141, parse the video resolution adjustment parameter, bit rate control parameter, and buffer interval configuration parameter included in the live service adjustment strategy.

[0144] For example, assume that in the live service adjustment strategy, the video resolution adjustment parameter is to increase from 720p to 1080p, the bit rate control parameter stipulates that the maximum allowable bit rate is 5 Mbps, the minimum guaranteed bit rate is 2 Mbps, and the buffer interval configuration parameter is set to increase the initial buffer threshold from 10 seconds to 15 seconds.

[0145] Step S142, send the video resolution adjustment parameter to the video encoding module to trigger the dynamic adjustment of the resolution preset value of the video encoder.

[0146] In this embodiment, after receiving the parameter to increase the picture resolution from 720p to 1080p, the video encoding module adjusts the resolution preset value of the video encoder according to this instruction. During the live broadcast of a science fiction movie, the video encoder originally encoded the video stream according to the resolution preset value of 720p. When receiving the new picture resolution adjustment parameter, it changes the resolution preset value of the encoding process to 1080p, thereby changing the resolution of the output video stream to meet the requirements of the optimized live broadcast service.

[0147] Step S143: Synchronize the bitrate control parameter to the bandwidth allocation module, triggering the dynamic bitrate adaptation algorithm based on the current network condition.

[0148] Among them, step S143 includes:

[0149] Step S1431: Monitor the real-time network throughput and packet loss rate of the live broadcast service node, and generate a network condition evaluation index.

[0150] For example, during the live broadcast of a science fiction movie, continuously monitor the network condition of the live broadcast service node. Suppose the real-time network throughput is 3Mbps and the packet loss rate is 1% obtained through monitoring. According to the preset rules, synthesize these two data to generate a network condition evaluation index. For example, a simple evaluation rule can be set: network condition evaluation index = network throughput - (packet loss rate × 10) (this is just an example rule, and it may be more complex in reality). Then, according to this rule, the calculated network condition evaluation index is 3Mbps - (1% × 10) = 3Mbps - 0.1Mbps = 2.9Mbps.

[0151] Step S1432: Calculate the recommended bitrate range under the current network condition according to the maximum allowable bitrate and the minimum guaranteed bitrate in the bitrate control parameter.

[0152] For example, given that the maximum allowable bitrate in the bitrate control parameter is 5Mbps and the minimum guaranteed bitrate is 2Mbps, combine the previously calculated network condition evaluation index of 2.9Mbps to calculate the recommended bitrate range. Since the network condition evaluation index of 2.9Mbps is greater than the minimum guaranteed bitrate of 2Mbps, the recommended bitrate range is [2Mbps, 2.9Mbps].

[0153] Step S1433: Call the bitrate adaptation controller to dynamically adjust the video transmission bitrate within the recommended bitrate range, so that the actual bitrate maintains a preset proportional relationship with the network throughput.

[0154] Assume that the preset proportional relationship is that the actual bitrate is 80% of the network throughput (this proportional relationship is set according to the requirements of the live broadcast service). If the current network throughput is 3 Mbps, then according to this proportional relationship, the calculated actual bitrate should be 3 Mbps × 80% = 2.4 Mbps. The bitrate adaptive controller will adjust the video transmission bitrate to 2.4 Mbps to provide an appropriate video transmission bitrate while meeting the network conditions.

[0155] Step S1434, when it is detected that the packet loss rate exceeds the threshold, trigger an emergency bitrate reduction operation and enable the forward error correction coding mechanism.

[0156] For example, assume that the set packet loss rate threshold is 3%. During the live broadcast, if it is monitored that the packet loss rate exceeds 3%, for example, reaches 5%, an emergency bitrate reduction operation will be triggered at this time. Assume that according to the preset emergency bitrate reduction rule, the bitrate needs to be reduced to 50% of the current bitrate (this is an example rule, and it is actually determined according to specific strategies). If the current bitrate is 2.4 Mbps, then the bitrate after the emergency bitrate reduction is 2.4 Mbps × 50% = 1.2 Mbps. At the same time, the forward error correction coding mechanism is enabled to reduce the impact of packet loss on the video playback quality.

[0157] Step S144, deploy the buffer interval configuration parameter to the data cache module, and trigger the adjustment of the initial buffer threshold according to the performance of the user device.

[0158] For example, in the live broadcast scenario of a science fiction movie, after the data cache module receives the buffer interval configuration parameter that increases the initial buffer threshold from 10 seconds to 15 seconds, it will be adjusted according to the performance of the user device. Assume that the user device has a large memory capacity and strong processing power. The data cache module can directly set the initial buffer threshold to 15 seconds. If the performance of the user device is poor, such as having a small memory capacity, the data cache module may, according to the set algorithm (such as comprehensively calculating based on factors such as the remaining memory of the device and the video bitrate), set the initial buffer threshold as close as possible to 15 seconds while ensuring the smoothness of video playback, for example, set it to 13 seconds. In this way, the initial buffer threshold is reasonably adjusted according to the performance of the user device to optimize the playback experience of the live broadcast service at the user end.

[0159] In a possible implementation manner, step S150 includes:

[0160] Step S151, collect the user - end playback smoothness index, picture quality score, and interactive response delay index within a preset observation period.

[0161] In this embodiment, it is assumed that the preset observation period is from 30 minutes after the start of the live broadcast to the end of the live broadcast. For the user-side playback fluency index, data is collected through a monitoring tool installed on the user device, such as recording information about stuttering during video playback; the picture quality score is given by the user based on their subjective feelings during viewing, and the value range can be from 1 to 10 points. These score data are collected in real time during the live broadcast; the interactive response delay index refers to the time interval from when the user issues an operation (such as pause, fast forward, etc.) to when the server responds and the user sees the operation result on the user side, and it is also continuously collected within this observation period.

[0162] Step S152, perform a stuttering event frequency statistics on the playback fluency index to generate a stuttering frequency feature.

[0163] For example, within the observation period from 30 minutes to the end of the live broadcast, carefully check the playback fluency index data and count the number of stuttering events that occur. Suppose a total of 5 stuttering events are found to have occurred, and the total duration of the observation period is 30 minutes. Then the stuttering frequency feature is the number of stuttering events divided by the total duration, that is, 5 times divided by 30 minutes, approximately 1 stuttering every 6 minutes (the calculation result here represents the stuttering frequency feature).

[0164] Step S153, perform a time series analysis on the picture quality score to generate a quality fluctuation feature.

[0165] For example, the picture quality score data collected within the observation period can be viewed. For example, the score is 8 points between 30 minutes and 40 minutes, 7 points between 40 minutes and 50 minutes, 8 points between 50 minutes and 60 minutes, etc. Analyze the changes in these scores over time, calculate the difference in scores between adjacent time periods, such as 8 points - 7 points = 1 point, 7 points - 8 points = -1 point. Determine the quality fluctuation feature based on the magnitude and trend of these differences. If the score difference is large and changes frequently, it indicates a large quality fluctuation; if the score difference is small and relatively stable, it indicates a small quality fluctuation.

[0166] Step S154, perform a percentile calculation on the interactive response delay index to generate a delay distribution feature.

[0167] For example, all the interactive response delay data within the observation period can be collected. Suppose there are 20 pieces of these interactive response delay data. These interactive response delay data can be sorted from smallest to largest, and then the delay value corresponding to the set percentile (such as the 90% percentile) is calculated. If the value of the 18th data (20 × 90% = 18, rounded up) is 2 seconds, then the interactive response delay at the 90% percentile is 2 seconds, and this 2 seconds is an important manifestation of the delay distribution feature.

[0168] Step S155: Calculate the difference degree between the stuttering frequency feature and the historical stuttering baseline data to generate a first loss component, analyze the deviation degree of the quality fluctuation feature from the expected quality stability curve to generate a second loss component, and count the proportion of high-latency samples in the latency distribution feature to generate a third loss component.

[0169] For example, assume that the historical stuttering baseline data is 1 stutter per 10 minutes, and the current stuttering frequency feature is 1 stutter per 6 minutes. When calculating the difference degree, subtract the historical stuttering baseline data from the current stuttering frequency, i.e., (1 / 6 - 1 / 10). First, find a common denominator for the fractions. 1 / 6 becomes 5 / 30, and 1 / 10 becomes 3 / 30. Then, (5 / 30 - 3 / 30) = 2 / 30 = 1 / 15, which is the first loss component. For the second loss component, the expected quality stability curve indicates that in an ideal situation, the picture quality score should remain relatively stable with minimal fluctuations. If the actual quality fluctuation feature shows large fluctuations in the score, deviating from the expected quality stability curve, by comparing the deviation degree between the two (the specific calculation method is based on the set deviation calculation rule. Here, assume that the deviation degree value is obtained through complex calculations such as calculating the sum of the squared differences between the fluctuating score and the stable score, and the deviation degree value is 0.2), this 0.2 is the second loss component. For the third loss component, assume that high latency is defined as a latency greater than 3 seconds. In the previously calculated latency distribution feature, if there are 3 data greater than 3 seconds out of a total of 20 data, then the proportion of high-latency samples is 3 / 20 = 0.15, which is the third loss component.

[0170] Step S156: Perform weighted summation on the first loss component, the second loss component, and the third loss component to generate a comprehensive loss function. Calculate the partial derivative of the comprehensive loss function with respect to the model weights through the backpropagation algorithm to generate the model parameter adjustment gradient, and update the weight parameters of the dynamic optimization model based on the model parameter adjustment gradient.

[0171] For example, assume that the weight of the first loss component is 0.3, the weight of the second loss component is 0.4, and the weight of the third loss component is 0.3. The comprehensive loss function is calculated as (1 / 15×0.3 + 0.2×0.4 + 0.15×0.3). First, calculate the multiplication parts: 1 / 15×0.3 = 0.02, 0.2×0.4 = 0.08, 0.15×0.3 = 0.045. Then add the results: 0.02 + 0.08 + 0.045 = 0.145. This 0.145 is the value of the comprehensive loss function. Calculate the partial derivative of the comprehensive loss function with respect to the model weights through the backpropagation algorithm to generate the gradient for adjusting the model parameters. The backpropagation algorithm calculates the partial derivative corresponding to each model weight according to the relationship between the comprehensive loss function and the model weights (this calculation process involves complex mathematical principles and algorithm operations, and here we mainly emphasize the operation based on the previously calculated comprehensive loss function). These partial derivatives form the gradient for adjusting the model parameters. Finally, update the weight parameters of the dynamic optimization model based on the gradient for adjusting the model parameters. For example, if the partial derivative corresponding to a certain model weight is 0.05, according to the set update rule (such as updating the weight by multiplying the set learning rate by the partial derivative), assuming the learning rate is 0.1, then the update amount of this weight parameter is 0.05×0.1 = 0.005. Apply this update amount to the original weight parameter to complete the update of the weight parameters of the dynamic optimization model, enabling the model to continuously optimize itself according to the real-time live broadcast quality metrics to improve the optimization effect of subsequent live broadcast services.

[0172] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a theater live broadcast optimization system 100 based on big data analysis that can implement the ideas of the present application provided by some embodiments of the present application. For example, the processor 120 can be used on the theater live broadcast optimization system 100 based on big data analysis and is used to execute the functions in the present application.

[0173] The theater live broadcast optimization system 100 based on big data analysis can be a general-purpose server or a special-purpose server, both of which can be used to implement the method for optimizing theater live broadcasts based on big data analysis of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0174] For example, the theater live broadcast optimization system 100 based on big data analysis may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the theater live broadcast optimization system 100 based on big data analysis may further include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present application can be implemented according to these program instructions. The theater live broadcast optimization system 100 based on big data analysis further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0175] For ease of explanation, only one processor is described in the theater live broadcast optimization system 100 based on big data analysis. However, it should be noted that the theater live broadcast optimization system 100 in the present application may further include multiple processors. Therefore, the steps performed by one processor described in the present application can also be jointly performed or separately performed by multiple processors. For example, if the processor of the theater live broadcast optimization system 100 based on big data analysis performs step A and step B, it should be understood that step A and step B can also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0176] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned theater live broadcast optimization method based on big data analysis is implemented.

[0177] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A cinema live broadcast optimization method based on big data analysis, characterized in that: The method comprises: Obtaining historical live broadcast big data of a target cinema, wherein the historical live broadcast big data includes multiple live broadcast data streams, each of which is composed of at least one user viewing behavior record and a corresponding live broadcast service status record; Perform multi-dimensional viewing feature extraction processing on the historical live broadcast big data to obtain real-time viewing features and service stability features of each live broadcast data stream; Based on the pre-trained dynamic optimization model, dynamic strategy matching processing is performed on the real-time viewing feature and the service stability feature to generate an optimized annotation result of the live data stream, and a live service adjustment strategy is generated according to the optimized annotation result; Synchronize the live broadcast service adjustment strategy to the live broadcast service node in real time, and trigger the live broadcast service node to adjust the current live broadcast parameter configuration according to the live broadcast service adjustment strategy; Based on the adjusted live broadcast parameter configuration, a real-time live broadcast quality index is collected, and the parameters of the dynamic optimization model are updated according to the real-time live broadcast quality index.

2. The method for optimizing live streaming of cinemas based on big data analysis according to claim 1, characterized in that: The multi-dimensional viewing feature extraction process is performed on the historical live broadcast big data to obtain the real-time viewing feature and service stability feature of each live broadcast data stream, including: Performing time window division processing on the user viewing behavior records in the live data stream to obtain viewing behavior segments of multiple continuous time periods; Calling a pre-trained real-time feature encoder to extract real-time viewing features from the viewing behavior segment, and generating a first feature set including user interaction frequency features, screen switching preference features, and content response delay features; Performing abnormal event detection processing on the live broadcast service status record to identify the timestamp of the live broadcast service interruption event and the triggering condition of the service degradation event; Performing a stability quantitative analysis on the live broadcast service status record based on the timestamp and the trigger condition to generate a second feature set including bandwidth fluctuation features, decoding error rate features, and resource occupancy peak features; The first feature set and the second feature set are subjected to feature fusion to obtain the real-time viewing feature and the service stability feature.

3. The method for optimizing live streaming of cinemas based on big data analysis according to claim 2 is characterized in that: The calling of the pre-trained real-time feature encoder performs real-time viewing feature extraction on the viewing behavior segment to generate a first feature set including user interaction frequency features, screen switching preference features, and content response delay features, including: Calling a pre-trained real-time feature encoder to perform behavior pattern analysis on the user operation log in the movie-watching behavior segment, extracting the number of user interaction operations and the operation interval duration in a single time window, and generating user interaction frequency features; Identify the timing relationship between the screen switching request time point and the live content key frame in the user operation log, calculate the time deviation between the screen switching request and the content key frame, and generate a screen switching preference feature; The difference between the content loading request and the server response timestamp in the user operation log is monitored, and the number of abnormal events in which the content loading delay exceeds a preset threshold is counted to generate a content response delay feature.

4. The method for optimizing live streaming of cinemas based on big data analysis according to claim 3 is characterized in that: The identifying the timing relationship between the screen switching request time point and the live content key frame in the user operation log, calculating the time deviation between the screen switching request and the content key frame, and generating the screen switching preference feature includes: Extracting the timestamp data of the screen switching operation triggering time from the user operation log, and generating a screen switching request time point set; Parse the key frame types and corresponding timestamp sequences pre-annotated in the live content metadata, and generate a set of live content key frame time points including scene switching key frames, special effect start key frames, and plot turning key frames; Traversing each screen switching request time point in the screen switching request time point set, searching for a target key frame time point having the smallest time difference with the current screen switching request time point in the live content key frame time point set; Calculating the absolute time difference between the screen switching request time point and the target key frame time point to generate an original time deviation of a single screen switching request time point; The original time deviation is annotated with positive and negative signs, when the screen switching request time point is later than the target key frame time point, it is annotated as a positive value, and when it is earlier than the target key frame time point, it is annotated as a negative value, and a time deviation sequence with a timing direction mark is generated; Counting a first frequency in a preset positive deviation interval, a second frequency in a preset negative deviation interval, and a third frequency in a zero deviation tolerance interval in the time deviation sequence; Generate a user's active switching tendency index for scene switching key frames, a delay follow-up index for special effect start key frames, and a predicted switching density index for plot turning key frames according to the proportional relationship among the first frequency, the second frequency, and the third frequency; The active switching tendency index, the delayed following index and the predicted switching density index are weighted according to the key frame type to generate a screen switching preference feature set including a time sensitivity weight and a content relevance weight.

5. The method for optimizing live streaming of cinemas based on big data analysis according to claim 4 is characterized in that: The method of parsing the pre-annotated key frame types and corresponding timestamp sequences in the live content metadata to generate a live content key frame time point set including scene switching key frames, special effect start key frames and plot turning key frames includes: Read the metadata marker block embedded in the live content video stream, and extract the frame number, frame type label and offset relative to the live start time of each key frame; Classifying the key frames into scene switching key frames, special effect start key frames and plot turning key frames according to the frame type labels; Converting the frame number of each key frame into absolute timestamp data based on the offset to generate a set of key frame time points of the live content arranged in chronological order; Verify the timestamp continuity of adjacent key frames in the live content key frame time point set. If it is detected that the timestamp jump exceeds a preset frame interval threshold, insert a virtual key frame time point in the jump interval and mark it as an unclassified key frame type.

6. The method for optimizing live streaming of cinemas based on big data analysis according to claim 2, characterized in that: The performing stability quantitative analysis on the live broadcast service status record based on the timestamp and the trigger condition to generate a second feature set including bandwidth fluctuation features, decoding error rate features, and resource occupancy peak features, including: Extracting a network bandwidth sampling value sequence associated with the timestamp of the live service interruption event from the live service status record, screening out bandwidth sampling data within a set time range before and after each live service interruption event, calculating a ratio of a standard deviation to a mean of the bandwidth sampling data, and generating a bandwidth fluctuation feature; Parsing the video decoding log segment matching the trigger condition of the service degradation event in the live broadcast service status record, identifying the video frame data transmitted during the period when the trigger condition of the service degradation event is effective, counting the percentage of the number of frames carrying decoding error identifiers in the video frame data to the total number of transmitted frames, and generating a decoding error rate feature; Traversing the hardware resource usage records of the live service status records, locating the CPU occupancy data points that overlap with the timestamp of the service interruption event, extracting the continuous time periods that exceed the preset safety threshold in the data points, calculating the weighted difference between the maximum value and the average value of the CPU occupancy in the continuous time period, and generating resource occupancy peak features; The bandwidth fluctuation characteristics, decoding error rate characteristics and resource occupancy peak characteristics are time-aligned to ensure that the time intervals corresponding to the bandwidth fluctuation characteristics, decoding error rate characteristics and resource occupancy peak characteristics maintain a synchronous mapping relationship with the timestamp of the service interruption event and the triggering condition of the service degradation event, thereby generating a second feature set with a time sequence label.

7. The method for optimizing live streaming of cinemas based on big data analysis according to claim 1, characterized in that: The pre-trained dynamic optimization model performs dynamic strategy matching processing on the real-time viewing feature and the service stability feature to generate an optimized annotation result of the live data stream, including: Concatenate the real-time viewing feature and the service stability feature into feature vectors to generate a multi-dimensional feature input vector; Obtaining a historical feature weight template stored in a strategy matching layer in the dynamic optimization model, wherein the historical feature weight template includes a benchmark weight value of each feature dimension under different live broadcast scenarios; Calculating the cosine similarity between each feature dimension in the multidimensional feature input vector and the historical feature weight template to generate a scene matching index, dynamically adjusting the reference weight value according to the scene matching index to generate a real-time feature weight value, and normalizing the real-time feature weight value to obtain a feature importance distribution; Performing weighted fusion on the multi-dimensional feature input vector according to the feature importance distribution to generate an optimized decision vector; The optimization decision vector is input into the annotation output layer of the dynamic optimization model to generate an optimization annotation result including a picture quality optimization direction, a bandwidth allocation optimization direction and a decoding priority optimization direction.

8. The method for optimizing live streaming of cinemas based on big data analysis according to claim 1, characterized in that: The step of synchronizing the live broadcast service adjustment strategy to the live broadcast service node in real time, and triggering the live broadcast service node to adjust the current live broadcast parameter configuration according to the live broadcast service adjustment strategy, includes: Parsing the picture resolution adjustment parameters, bit rate control parameters and buffer space configuration parameters included in the live broadcast service adjustment strategy; Sending the picture resolution adjustment parameter to the video encoding module to trigger dynamic adjustment of the resolution preset value of the video encoder; Synchronize the bit rate control parameters to the bandwidth allocation module to trigger a dynamic bit rate adaptation algorithm based on the current network status; Deploy the buffer zone configuration parameters to a data cache module, triggering adjustment of an initial buffer threshold according to user equipment performance; The step of synchronizing the rate control parameters to the bandwidth allocation module and triggering a dynamic rate adaptation algorithm based on the current network status includes: Monitor the real-time network throughput and packet loss rate of live broadcast service nodes and generate network status evaluation indicators; Calculate the recommended bit rate range under the current network conditions according to the maximum allowed bit rate and the minimum guaranteed bit rate in the bit rate control parameters; Calling a bit rate adaptive controller to dynamically adjust the video transmission bit rate within the recommended bit rate range so that the actual bit rate and the network throughput maintain a preset proportional relationship; When it is detected that the packet loss rate exceeds the threshold, an emergency bit rate reduction operation is triggered and the forward error correction coding mechanism is enabled.

9. The method for optimizing live streaming of cinemas based on big data analysis according to claim 1, characterized in that: The collecting the real-time live broadcast quality index based on the adjusted live broadcast parameter configuration, and updating the parameters of the dynamic optimization model according to the real-time live broadcast quality index, comprises: Collect the user-side playback fluency index, picture quality score and interactive response delay index within the preset observation period; Performing frequency statistics of freeze events on the playback smoothness index to generate freeze frequency features; Performing time series analysis on the picture quality scores to generate quality fluctuation characteristics; Calculating percentiles of the interactive response delay indicator to generate a delay distribution feature; Calculate the difference between the jam frequency feature and the historical jam baseline data to generate a first loss component, analyze the degree of deviation between the quality fluctuation feature and the expected quality stability curve to generate a second loss component, and count the proportion of high-delay samples in the delay distribution feature to generate a third loss component; Perform a weighted summation on the first loss component, the second loss component and the third loss component to generate a comprehensive loss function, calculate the partial derivative of the comprehensive loss function with respect to the model weight through a back propagation algorithm, generate a model parameter adjustment gradient, and update the weight parameters of the dynamic optimization model based on the model parameter adjustment gradient.

10. A cinema live broadcast optimization system based on big data analysis, characterized in that: The theater live broadcast optimization system based on big data analysis includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the theater live broadcast optimization method based on big data analysis as described in any one of claims 1 to 9.

Images