Video prediction caching strategy based on Markov correction model
The Markov correction model predicts the access probability of video segments, which solves the problem of slow response and inaccurate popularity evaluation of newly launched videos in P2P streaming media cache, and realizes a video cache strategy with high hit rate and fast response.
Patent Information
- Application Number
- CN202510892555.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-09-02
AI Technical Summary
The existing P2P streaming media caching strategy cannot respond quickly when newly launched videos, and relies on historical access data to cause inaccurate popularity assessment, affecting response delay.
The video prediction cache strategy based on the Markov correction model is adopted, and the state transition matrix is corrected using the exponential weighted average model to predict the access probability of the video segment, and the cached video segment is selected based on the access probability.
It improves the hit rate of video segments, has higher accuracy, can quickly respond to the popularity of newly launched videos, reduce response delay, and adapt to the continuous changes in user click-through rate.
Smart Images
Figure CN120583248A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of streaming media big data, and in particular relates to a video prediction caching strategy based on a Markov correction model. Background Art
[0002] In recent years, with the continuous advancement of information technology, computer network performance has become increasingly advanced. The smooth and high-quality services of mobile terminals have attracted a large number of users, and more and more people are turning to mobile devices to obtain information. This convenient method has put tremendous pressure on traditional media transmission systems. In the client / server (C / S) model, the centralized deployment of servers leads to an over-concentration of user data requests, which can cause a certain degree of network congestion when server bandwidth is limited. However, P2P streaming media systems store data locally on individual nodes and establish an overlay network with hierarchical nodes (super nodes, ordinary nodes), effectively alleviating network congestion. P2P streaming media systems allow users to play videos and audio files with a short buffering period before playing them, playing only as much data as they receive, without having to wait for the entire video or audio file to be fully cached. This allows users to download and watch the selected video at the same time. From opening a video to playing it, from the source server to the proxy server to the user node, the data transmission process within the system is like a stream of water. Therefore, these new media transmission systems are figuratively called streaming media systems.
[0003] The caching strategy of a P2P streaming system requires precise evaluation of popular videos to maximize the hit rate of video segments while reducing response latency. This directly impacts user experience and streaming system performance. Node selection technology in a P2P streaming system can reduce server pressure and improve the quality of service at service nodes. However, existing technologies have the following problems:
[0004] 1. Most existing P2P streaming media caches cache videos based on the popularity of video segments, which cannot respond quickly when new dramas are released.
[0005] 2. In existing solutions, the selection of P2P streaming segments relies on statistics of all historical access data. However, some video segments were once very popular but have recently lost their popularity. The statistics of all historical data will affect the selection of cache segments. Summary of the Invention
[0006] The purpose of the present invention is to provide a video prediction caching strategy based on a Markov correction model to solve the technical problem in the existing technology that for newly launched products and radio frequencies launched earlier, the playback popularity assessment of video segments is inaccurate due to limited statistics on historical access data, which affects the corresponding delay.
[0007] The video prediction caching strategy based on the Markov correction model includes the following steps:
[0008] S1. Based on the user access records (i.e., prior data), extract the user access patterns, and derive the initial state transfer matrix and the user access initial probability;
[0009] S2. Use the exponential weighted average model to modify the state transfer matrix, and add the old state transfer matrix to the new state transfer matrix by weighted summation;
[0010] S3. Obtaining a state transition matrix of the predicted segment based on iterative calculation of the state transition matrix, and then calculating the access probability of the video segment at each time point in the predicted segment;
[0011] S4. Select corresponding video segments for caching based on the access probability.
[0012] Preferably, in step S1, the video is segmented, and the probability of the user initially accessing the corresponding video segment is calculated for each video segment based on the user access record, and the corresponding initial access probability matrix is formed in sequence; then, the initial state transfer matrix of the prior data is obtained based on the probability and correlation relationship of the user successively accessing each video segment, and the user access record includes the user number and the playback record of each user for each video segment.
[0013] Preferably, in step S2, as the system continues to iterate, the user access records are read in a sliding window manner, and the state transition matrix is modified based on the new and old user access records in the sliding window.
[0014] Preferably, in step S2, the state transfer matrix is corrected using an exponential weighted average model, the old state transfer matrix is added to the new state transfer matrix by weighted summation, and the state transfer matrix of the predicted segment is obtained by predicting and calculating the historical data of user access records.
[0015] Preferably, the state transfer matrix formula S at the current moment is n_now The calculation formula is as follows:
[0016]
[0017] Among them, S i t i The state transfer matrix at the moment, i=1,2,…,n; a is the coefficient, m is the exponent of the weight coefficient 1-a, for S i When weighting is performed, i+m=n, S n_now Represents the state transfer matrix formula at the current moment.
[0018] Preferably, according to the reading method of the sliding window, at time t1, the state transfer matrix is extracted from the access records obtained by the sliding window to obtain the "shadow state transfer matrix" S1 at time t1, and then the state transfer matrix at time t1 is calculated as follows:
[0019]
[0020] The state transition matrix S based on time 1 1_now Calculate the probability of a video segment being accessed at each future time point. Preferably, the calculation formula for the probability of a video segment being accessed at the first three time points is as follows:
[0021] P1=P0×S0,
[0022]
[0023] Among them, P1, P2 and P3 are the access probabilities of the video segments at the first three time points respectively.
[0024] Preferably, in step S4, after predicting the access probability of the video segment, the video segment with a higher access probability is selected for caching.
[0025] The video prediction caching strategy based on the Markov modified model can be run based on a cache replacement algorithm and a first-in-first-out algorithm respectively.
[0026] The technical advantages of the present invention are as follows: First, the probability of a video segment being accessed is predicted based on a modified Markov model. Then, segments with a higher access probability are selected for caching. This method obtains a state transition matrix from the number of clicks on a video segment to adapt to the continuous changes in the user's click rate. According to experimental verification, the results obtained by this method have a significantly higher hit rate than other existing technologies, so the prediction accuracy is higher. Since the state transition matrix is updated and corrected based on recent historical data through a time window during the calculation, the number and frequency of recent visits to each video segment can be fully considered, and the popularity of newly launched video segments can be fully considered, and premature historical data can be prevented from affecting the accuracy of the system's prediction of recent video popularity. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 This is a basic flow chart of the video prediction caching strategy based on the Markov correction model of the present invention.
[0028] Figure 2 This is a schematic diagram of reading user access records using a sliding window approach in the present invention.
[0029] Figure 3 and Figure 4The hit rate comparison diagram is obtained by comparing the present invention with the prior art based on the cache replacement (CRA) and first-in-first-out (FIFO) algorithms respectively.
[0030] Figure 5 and Figure 6 This is a hit rate comparison chart obtained by comparing the present invention with the prior art based on the first-in-first-out (FIFO) algorithm. DETAILED DESCRIPTION
[0031] The specific implementation methods of the present invention will be further explained in detail below through the description of embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.
[0032] like Figures 1-6 As shown, the present invention provides a video prediction caching strategy based on a Markov correction model, which includes the following steps.
[0033] S1. Based on the user access records, i.e., prior data, the user access patterns are extracted to obtain the initial state transfer matrix and the initial probability of user access.
[0034] In the Markov prediction model, user access records serve as prior data to calculate the initial user access probability and the corresponding initial state transition matrix. The video is first segmented. For each video segment, the probability of the user initially accessing the corresponding segment is calculated based on the user access records, and the corresponding initial access probability matrix is formed. The user access records include the user ID and each user's playback history for each video segment. In this embodiment, the user access records are shown in Table 1.
[0035] Table 1: User access record table
[0036]
[0037] In the user access records shown in Table 1, the probability of a user initially accessing the first segment is 2 / 15, and the probability of accessing the second segment is 7 / 45. The user initial access probability matrix for each video segment in the video is as follows:
[0038] P0=[2 / 15 7 / 45 2 / 15 2 / 15 4 / 45 1 / 45 1 / 45 4 / 45 1 / 9 1 / 9],
[0039] Among them, P0 is the initial access probability matrix that represents the initial access probability of users for each video segment.
[0040] From the six user access records above, we can see that after a user visits the first segment, the probability of visiting the first segment again is 0, the probability of visiting the second segment is 5 / 6, the probability of visiting the third segment is 1 / 6, and the probability of visiting the remaining segments is 0; after a user visits the second segment, the probability of visiting the first segment is 1 / 7, the probability of visiting the third segment is 5 / 7, and the probability of visiting the fourth segment is 1 / 7. Similarly, the initial state transfer matrix of these six data is calculated and expressed as:
[0041]
[0042] Among them, S0 represents the initial state transfer matrix obtained based on prior data.
[0043] S2. Use the exponential weighted average model to modify the state transfer matrix, and add the old state transfer matrix to the new state transfer matrix by weighted summation.
[0044] Because users who watch videos have strong subjective perceptions, a fixed state transition matrix cannot accurately describe the changing probability of a video segment being accessed during system iterations. Prior probabilities work well in the early stages of system execution, but as the system continues to iterate, the state transition matrix needs to be modified to more accurately predict segment access probabilities.
[0045] To this end, this step uses an exponentially weighted average model to modify the state transfer matrix, adding the old state transfer matrix to the new state transfer matrix through weighted summation. As the system continues to iterate, new user access records are added to the old access records. Due to the large number of user access records, the computational complexity of the state transfer matrix using the cumulative calculation method will gradually increase. Considering the temporal relationship between user access records, this step also uses a sliding window method to read user access records, such as Figure 2 As shown, a “sliding window diagram” is obtained.
[0046] Afterwards, the model is modified using the exponential weighted average model to obtain the state transfer matrix S0 at time t0 from the user access record at time t0; obtain the state transfer matrix S1 at time t1 from the user access record at time t1; and obtain the state transfer matrix S2 at time t2 from the user access record at time t2.
[0047] The state transfer matrix formula S at the current moment n_now The calculation formula is as follows:
[0048]
[0049] Among them, S i t iThe state transfer matrix at the moment, i=1,2,…,n; a is the coefficient, m is the exponent of the weight coefficient 1-a, for S i When weighting is performed, i+m=n, S n_now The state transfer matrix formula representing the current moment can be used to calculate the state transfer matrix of the predicted segment.
[0050] S3. Obtain the state transfer matrix of the predicted segment based on iterative calculation of the state transfer matrix, and then calculate the access probability of the video segment at each time point in the predicted segment.
[0051] In this embodiment, Table 2 is a user access record table after a new user accesses, wherein user numbers 7, 8, and 9 are newly added user access records.
[0052] Table 2 New user access record table
[0053]
[0054] According to the sliding window reading method, at time t1, the state transition matrix is extracted from 4 to 9 access records. The "shadow state transition matrix" at time t1 is obtained as follows:
[0055]
[0056] Then calculate the state transfer matrix at time t1 as follows:
[0057]
[0058] In the embodiment, assuming coefficient a=0.5, the following formula can be obtained:
[0059]
[0060] In this way, the probability of a video segment being accessed at each future time point can be calculated.
[0061] The access probabilities of the video segments at the first three time points are given below, as shown in the following formulas:
[0062] P1=P0×S0,
[0063]
[0064] Among them, P1, P2 and P3 are the access probabilities of the video segments at the first three time points respectively.
[0065] S4. Select corresponding video segments for caching based on the access probability.
[0066] This method first predicts the access probability of a video segment based on a modified Markov model; then selects the video segment with a larger access probability for caching.
[0067] The technical effect of this method is verified through experiments, and the specific contents are as follows.
[0068] The video prediction cache strategy MMPM provided by the present invention is run based on the cache replacement (CRA) algorithm and the first-in-first-out (FIFO) algorithm respectively, and compared with the related existing technologies CFCD and PCN. The specific simulation and comparison results are shown in Figure 2. Figure 3-Figure 6 As shown in the results, the video prediction caching strategy MMPM based on cache replacement (CRA) (CRA-MMPM) and the video prediction caching strategy MMPM based on cache replacement (FIFO) (FIFO-MMPM) have significantly higher hit rates than other existing technologies.
[0069] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the scope of protection of the present invention.
Claims
1. A video prediction caching strategy based on a Markov model, characterized by: The following steps are involved: S1. Based on the user access records (i.e., prior data), extract the user access patterns, and derive the initial state transfer matrix and the user access initial probability; S2. Use the exponential weighted average model to modify the state transfer matrix, and add the old state transfer matrix to the new state transfer matrix by weighted summation; S3. Obtaining a state transition matrix of the predicted segment based on iterative calculation of the state transition matrix, and then calculating the access probability of the video segment at each time point in the predicted segment; S4. Select corresponding video segments for caching based on the access probability.
2. The video prediction caching strategy based on the Markov modified model according to claim 1, characterized in that: In step S1, the video is segmented, and the probability of the user initially accessing the corresponding video segment is calculated for each video segment based on the user access record, and the corresponding initial access probability matrix is formed in sequence; then, the initial state transition matrix of the prior data is obtained based on the probability and correlation relationship of the user's successive access to each video segment. The user access record includes the user number and each user's playback record of each video segment.
3. The video prediction caching strategy based on the Markov modified model according to claim 1, characterized in that: In step S2, as the system continues to iterate, the user access records are read using a sliding window method, and the state transition matrix is modified based on the new and old user access records in the sliding window.
4. The video prediction caching strategy based on the Markov modified model according to claim 1 or 3, characterized in that: In step S2, the state transfer matrix is modified using an exponential weighted average model, and the old state transfer matrix is added to the new state transfer matrix by weighted summation. The state transfer matrix of the predicted segment is calculated based on the historical data of user access records.
5. The video prediction caching strategy based on the Markov modified model according to claim 4, characterized in that: The state transfer matrix formula S at the current moment n_now The calculation formula is as follows: Among them, S i t i The state transfer matrix at the moment, i=1,2,…,n; a is the coefficient, m is the exponent of the weight coefficient 1-a, for S i When weighting is performed, i+m=n, S n_now Represents the state transfer matrix formula at the current moment.
6. The video prediction caching strategy based on the Markov modified model according to claim 1, characterized in that: According to the sliding window reading method, at time t1, the state transfer matrix is extracted from the access records obtained by the sliding window to obtain the "shadow state transfer matrix" S1 at time t1. The state transfer matrix at time t1 is then calculated as follows: The state transition matrix S based on time 1 1_now Calculate the probability of a video segment being accessed at each future time point.
7. The video prediction caching strategy based on the Markov modified model according to claim 6, characterized in that: The calculation formula for the access probability of the video segment at the first three time points is as follows: P1=P0×S0, Among them, P1, P2 and P3 are the access probabilities of the video segments at the first three time points respectively.
8. The video prediction caching strategy based on the Markov modified model according to claim 1, characterized in that: In step S4, after predicting the access probability of the video segment, the video segment with a higher access probability is selected for caching.
9. The video prediction caching strategy based on the Markov modified model according to claim 1, characterized in that: It can run based on cache replacement algorithm and first-in-first-out algorithm respectively.