Video content rapid recommendation method based on edge calculation

Abstract

The invention discloses a video content rapid recommendation method based on edge computing. The method comprises the following steps: S1, scheduling a video recommendation task to an edge computing node; s2, performing personalized video recommendation by using a reinforcement learning algorithm based on a time sequence attention diffusion model; s3, adjusting the calculation complexity of the algorithm according to the calculation capability of the edge calculation node and the network bandwidth condition; s4, pre-caching the video content interested by the user to an edge computing node, and providing the video content for playing when the user selects a video; s5, carrying out distributed calculation in cooperation with the other group of edge nodes; s6, transferring the video recommendation algorithm reasoning task to a high-level edge server; and S7, updating a video recommendation algorithm in real time according to user feedback. According to the method, edge calculation, reinforcement learning, a time sequence attention mechanism and an improved diffusion model are combined, rapid response, accuracy and high efficiency of personalized video recommendation are achieved, and utilization of calculation resources and network bandwidth is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of video recommendation, and particularly to a method for quickly recommending video content based on edge computing. Background Art

[0002] With the rapid development of the Internet and mobile devices, video content has become an important form of information dissemination and entertainment consumption in modern society. Video streaming platforms and video recommendation systems have become an indispensable part of people's daily lives. These platforms and systems usually recommend relevant video content to users based on the users' interests, preferences, behavioral data, etc., so as to improve the users' viewing experience and the activity of the platform. Traditional video recommendation systems mainly rely on centralized data processing methods, perform complex video recommendation calculations through cloud servers, and use big data and machine learning algorithms to analyze the users' viewing history, behavioral patterns, and interest preferences to generate personalized video recommendation lists. However, with the continuous increase in video content and the diversification of user needs, existing recommendation systems face many challenges. Especially in an environment with limited computing resources and network bandwidth, it is difficult to effectively guarantee the recommendation effect and real-time performance.

[0003] In a cloud computing environment, a video recommendation system needs to rely on a powerful central server to process and calculate recommendation tasks. These servers are usually located in remote data centers. Video recommendation calculations need to transmit a large amount of user data and computing tasks from terminal devices to the data center for processing. However, this method is restricted by network bandwidth and data transmission speed. When a user is in a network environment with low bandwidth or high latency, the response speed and recommendation effect of the video recommendation task will be seriously affected. In addition, the latency of data transmission and the uncertainty of computing load in the cloud computing environment may lead to the inability to guarantee the real-time performance of the video recommendation system. Especially for application scenarios such as video streaming that require high real-time performance, the existing technology cannot meet the requirements of fast and accurate recommendations.

[0004] To overcome the limitations of traditional centralized video recommendation systems, in recent years, edge computing has received extensive attention as an emerging computing model. By deploying computing nodes near user devices or at the network edge, edge computing can process and store data close to the data source, thereby reducing data transmission latency and bandwidth consumption. Edge computing can make computing tasks more decentralized, relieve the pressure on the central server, and improve the real-time performance and efficiency of data processing. Combining the advantages of edge computing, the video recommendation system can sink the recommendation computing tasks to the edge nodes and quickly complete the computing locally, avoiding large-scale data transmission and latency problems and improving the response speed of the recommendation results. However, edge computing nodes usually have relatively limited computing power and storage resources, which makes it an important issue how to efficiently utilize edge computing nodes for personalized video recommendation.

[0005] Currently, although some studies have proposed video recommendation methods based on edge computing, there are still many problems. First, existing edge computing recommendation methods usually ignore the temporal characteristics of user interest changes in personalized recommendation. In user viewing behavior, interests change dynamically over time. Existing recommendation systems often only focus on the overall analysis of historical data and fail to consider the impact of time series factors on user interests. Second, although reinforcement learning and deep learning methods have been widely applied to recommendation systems, most existing methods rely on the central server for computing and fail to fully utilize their advantages in the edge computing environment. Recommendation algorithms based on reinforcement learning usually require a large amount of computing resources and storage space, which are difficult to meet by the limited resources of edge nodes. In addition, existing temporal feature modeling methods fail to combine the temporal attention mechanism in deep learning and ignore the important impact of recent user behavior on the current recommended content, resulting in the personalization degree and accuracy of recommendations being difficult to meet user expectations.

[0006] Based on the above background, existing edge computing video recommendation methods still face the following key problems: First, how to efficiently process personalized video recommendation tasks on edge computing nodes to avoid poor recommendation effects due to limited computing resources; second, how to fully consider the temporal changes of user interests and use temporal features to model user interests, thereby improving the personalization and accuracy of recommendations; third, how to combine reinforcement learning and the temporal attention mechanism and use the distributed computing power of edge computing nodes to optimize dynamic recommendation strategies to achieve more efficient video content recommendation.

[0007] Therefore, the main drawbacks of the prior art are as follows: The personalized video recommendation method in the edge computing environment fails to fully consider the change of user interests over time and fails to reasonably utilize the distributed resources of edge computing for efficient recommendation calculation. The reinforcement learning and time series modeling methods of the prior art rely too much on the central server and fail to fully consider the computing and storage limitations of edge computing nodes, resulting in low computing efficiency and real-time performance of the recommendation system. In addition, the existing diffusion models and attention mechanisms fail to effectively combine the historical behavior data of users and fail to optimize the recommendation strategy. Therefore, there is still much room for improvement in the personalization degree and accuracy of the recommendation. Summary of the Invention

[0008] An object of the present invention is to propose a fast video content recommendation method based on edge computing. The present invention combines reinforcement learning, time series attention mechanism and improved diffusion model to perform personalized video recommendation calculation on edge computing nodes, fully utilizes the computing and storage resources of edge computing nodes, optimizes the recommendation strategy, and improves the recommendation effect and real-time performance.

[0009] A fast video content recommendation method based on edge computing according to an embodiment of the present invention includes the following steps: S1. Schedule the video recommendation task to the edge computing node according to the user's geographical location, network condition and device capability; S2. On the edge computing node, based on the user's historical viewing data, historical interest preferences and historical behavior data, use the reinforcement learning algorithm based on the time series attention diffusion model to perform personalized video recommendation, construct the user's time series feature vector, apply the time series attention mechanism to weight the historical behavior data, capture the influence of recent behavior on the current recommendation, use the improved diffusion model to model the user's historical behavior data, predict the video content that the user is interested in, and combine the reinforcement learning algorithm to optimize the recommendation strategy to generate the user's personalized recommendation list; S3. Adjust the computational complexity of the algorithm according to the computing power and network bandwidth condition of the edge computing node, optimize the execution mode of the computing task for resource limitations, and reduce the task complexity when the system load is high; S4. Cache the video content that the user is interested in into the edge computing node in advance. When the user selects a video, immediately provide the video content from the edge computing node for playback; S5. When the computing power of the edge computing node is insufficient, cooperate with another group of edge nodes for distributed computing to process the video recommendation task; S6. Collaborate between edge nodes to transfer the video recommendation algorithm inference task to a high-level edge server; S7. Update the video recommendation algorithm in real time according to user feedback.

[0010] Optionally, S1 specifically includes: S11. Obtain the geographical location information of the user, which is obtained from the user device through the positioning module to determine the geographical area where the user is located; S12. By analyzing the network status of the user, obtain the current network bandwidth and network latency information, and monitor the network quality between the user device and the edge computing node in real time; S13. Evaluate the CPU performance, memory capacity, and graphics processing ability indicators of the user device; S14. According to the network bandwidth and computing power data obtained in steps S12 and S13, schedule the video recommendation task to the edge computing node. When the network bandwidth and device computing power meet the requirements of the recommendation task, the task is assigned to the current edge computing node. When the network bandwidth or computing power does not meet the requirements, the task is assigned to another edge computing node; S15. After the recommendation task scheduling is completed, feedback the processing status of the current recommendation task to the user device, including whether the task scheduling is successful, the status of the selected edge computing node, and the network quality information.

[0011] Optionally, S2 specifically includes: S21. On the edge computing node, obtain the historical viewing data of the user and construct the historical behavior dataset of the user. For each time step t, extract the viewing records related to the time step from the historical behavior dataset, and represent each viewing record as a feature vector , and construct all the viewing records into a time series in chronological order , each feature vector represents the viewing behavior feature vector at time step i. The historical viewing data includes the video content viewed by the user, the viewing duration, the viewing frequency, and the timestamp information at each viewing. The feature vector includes the metadata of the video and the viewing behavior features. The metadata includes the video duration, category, and score. The viewing behavior features include the viewing duration and the viewing frequency; S22. Based on the historical viewing data, construct the time series feature vector of the user. Preprocess the historical viewing data, standardize all feature values using the normalization method, calculate the weighted sum of the feature vectors at each time step, and generate the time series feature vector , and the time series feature vector represents the viewing interest of the user within a time period: ; Wherein, represents the time series feature vector, represents the weighting coefficient of the i-th viewing record, represents the feature vector at time step i, and n represents the total number of time steps; S23. Calculate the attention weight at each time step t The attention weight reflects the importance of historical behavior data at each time step for the current recommendation: ; Among them, represents the attention weight at time step t, that is, the weighted coefficient of the t-th viewing record, and exp represents the natural exponential function. represents the query vector at time step t. represents the query vector at time step i. and represent the time series feature vector; S24. Dynamically calculate the weighted coefficient , and weight each historical viewing record according to the corresponding attention weight to reflect the influence of recent viewing behavior on the current interest, and generate a weighted time series feature vector ; S25. Based on the calculated weighted time series feature vector , apply an improved diffusion model to model the user behavior data, simulate the propagation process of recommended content in the user interest space, and generate the user interest state and interest distribution. The improved diffusion model integrates a time series attention mechanism and combines with a reinforcement learning algorithm to model the user's interest propagation process; S26. Based on the user interest state, combine with a reinforcement learning algorithm to optimize the recommendation strategy; S27. Based on the update result of the reinforcement learning algorithm, generate a personalized recommendation list for the user. The personalized recommendation list is sorted according to the user's current interest state and the predicted interest values for different video contents, and the sorting result is used to generate the video recommendation result. The recommendation list is generated based on the historical behavior, attention weighted features, the interest propagation process generated by the diffusion model, and the strategy optimized by reinforcement learning.

[0012] Optionally, the S25 specifically includes: S251. Initialize the user interest state as the user interest at the initial moment. The user interest state is the user's current interest representation, and the initial user interest state is determined according to the user's historical viewing records, interest preferences, and the context information at the current recommendation moment; S252. Set the adjacency matrix A of the diffusion process. The adjacency matrix represents the connection relationship between each node in the user interest space. The element in the adjacency matrix represents the propagation weight between node i and node j in the user interest space, reflecting the influence from one node to another node. The nodes include video content and viewing time; S253. Design an iterative propagation process based on an improved diffusion model to simulate the dynamic changes of the user interest state and update the propagation of the user interest state: ; Among them, represents the user interest state at time step t, A represents the adjacency matrix of the diffusion process, that is, the propagation matrix of the improved diffusion model, reflecting the propagation relationship between nodes, represents the user interest state at time step t - 1, represents the weight parameter, adjusting the influence of the temporal feature on the user interest, represents the weighted temporal feature vector, which is the weighted interest of the user behavior at the current time step, represents the propagation weight coefficient, regulating the influence of neighbor nodes on the current state, represents the weight matrix of each neighbor node i, which is the contribution during the propagation process, represents the user interest state of neighbor node i at time step t - 1; S254. At each time step t, calculate the user interest distribution according to the updated interest state The user interest distribution represents the preference degree of the user for different video contents at time step t: ; Among them, represents the user interest distribution at time step t, softmax represents normalizing the interest state vector and outputting the probability distribution, which is the interest of the user in each video content, represents the user interest state at time step t, represents the bias vector, which is the adjustment factor of the user interest; S255. Based on the calculated user interest distribution , simulate the propagation of the recommended content in the user interest space, and iteratively calculate the change of the user interest state through a multi-step propagation process.

[0013] Optionally, the S26 specifically includes: S261. According to the user interest state , construct the current state and the recommendation action The recommendation action represents the video content recommended to the user, which is the video content predicted based on The state represents the interest distribution of the user at the current moment, represented by the vector ; S262. Define the reward function , the reward function is comprehensively evaluated based on the user's interaction behavior, the matching degree between the recommended content and the user's interests, and time factors: ; Among them, represents the reward function, , and represent constant coefficients to adjust the weights of various factors in the reward, represents the feature vector of the state , represents the feature vector of the action , represents the cosine similarity between the user's interests and the recommended content, measuring the matching degree between the two, represents the two-norm operation, represents the initial weight, represents the decay factor, represents the timestamp of the historical behavior, t represents the current time, represents controlling the influence of historical behavior on the current recommended content through time decay, and the influence of more distant historical data on the recommendation result gradually weakens, represents the interaction intensity of the user with the recommended content. The interaction includes the frequency of the user clicking on the recommended content, the total duration of the user watching the recommended content, and the user's rating or liking of the recommended content; S263. For each state-action pair , calculate and update the Q value. The Q value represents the expected return of taking a certain action in a given state: ; Among them, represents the value of taking the action in the state , which is the expected reward for choosing the action, represents the reward function, which is the immediate reward obtained after choosing the action in the current state , and is the user's feedback on the video content, represents the learning rate, controlling the update amplitude, represents the discount factor, measuring the importance of future rewards, represents the policy probability of choosing the action in the next state , represents the value of taking the action in the state , represents the expected Q value based on the policy; S264. After each update of the Q value, generate an optimal recommended action based on the updated Q value , and recommend video content to the user according to the optimal recommended action, where the optimal recommended action is determined by selecting the action with the maximum Q value: ; where, represents the optimal recommended action, and argmax(x) represents the variable value that maximizes x; S265. The reinforcement learning algorithm iteratively updates the recommendation strategy based on user feedback, and updates the Q value each time according to the user's interaction feedback on the recommended content to adjust the user's interest state.

[0014] Optionally, the specific steps of S3 include: S31. Obtain the computing power and network bandwidth of each edge computing node; S32. Calculate the complexity of the recommendation task , where is the computing complexity of task t, indicating the amount of computation required for task t: ; where, represents the complexity of the recommendation task, represents the number of subtasks of task t, represents the weight coefficient of subtask i, represents the computing time complexity of subtask i; S33. Calculate the edge node load index according to the computing power and network bandwidth of the edge computing node: ; where, represents the load index of edge computing node n, represents the computing power of the edge computing node, represents the maximum computation of the edge computing node, represents the network bandwidth, represents the maximum bandwidth of the edge node; S34. Adjust the computing complexity of the recommendation task according to the value of the edge computing node load index. If , degrade the computing complexity of the recommendation task, and calculate the size of the candidate video set for the recommended content, where is the size of the candidate video set for each user in the recommendation task: ; where, Represents the size of the adjusted candidate video set, Represents the size of the default recommended candidate set, and floor represents the floor function, which returns the largest integer not greater than the variable; S35. When the edge computing node load index increases the computational complexity of the recommendation task, improves the recommendation accuracy, and expands the size of the recommendation candidate set: ; Among them, represents the size of the expanded recommendation candidate set, and ceil represents the ceiling function, which returns the smallest integer not less than the variable.

[0015] The beneficial effects of the present invention are as follows: First of all, by scheduling the video recommendation task to the edge computing node, the present invention reduces the delay of data transmission and improves the response speed of recommendation calculation. In traditional video recommendation systems, recommendation calculation usually relies on a remote central server, which results in the real-time performance and accuracy of recommendation results being affected in network environments with low bandwidth or high latency. In the present invention, the video recommendation task is processed on the local edge computing node, eliminating the latency problem caused by remote transmission and improving the response speed of the recommendation system. Especially in bandwidth-constrained environments, users can obtain personalized recommendation content more quickly.

[0016] Secondly, the present invention effectively utilizes the computing resources of the edge computing node, solving the computing bottleneck problem brought about by the traditional video recommendation system relying on cloud servers for centralized computing. Performing recommendation calculation on the edge computing node can give full play to the distributed computing power of the edge node, relieve the load pressure on the central server, and avoid the decline in recommendation accuracy caused by insufficient computing resources. The recommendation strategy optimization based on reinforcement learning proposed by the present invention, combined with the computing power and network bandwidth conditions of the edge node, can dynamically adjust the complexity of the recommendation task, while ensuring the recommendation accuracy, making full use of the resources of the edge computing node.

[0017] Thirdly, the present invention has made innovations in time series modeling, adopting a time series attention mechanism to capture the dynamic characteristics of user interests changing over time. Traditional video recommendation systems usually perform static modeling on historical data, failing to fully consider the changing trend of user interests over time. The present invention constructs a time series feature vector and applies a time series attention mechanism to weight the historical behavior data of users, capturing the influence of recent user behavior on the current recommended content, improving the personalization and accuracy of recommendations. Especially when modeling the dynamic changes of user interests, it can more accurately reflect the interest preferences of users at different time periods, providing video recommendations that better meet the current needs of users.

[0018] In addition, the present invention also introduces an improved diffusion model, which combines a reinforcement learning algorithm to model the process of user interest propagation. By simulating the propagation of user interest states in the user interest space, the present invention can more effectively predict user interest changes and adjust the recommendation strategy according to the changes in interest states. Through this innovation, the video recommendation system can more accurately capture the potential needs of users for video content, provide a more personalized recommendation list for users, and also adjust the recommendation strategy in real time according to user feedback, thereby continuously optimizing the recommendation effect.

[0019] Finally, through the cooperation between edge computing nodes, when the computing power is insufficient, the system performance can be further improved through distributed computing. By transferring the inference task of the video recommendation algorithm to a high-level edge server, the problem of insufficient computing power of a single node can be effectively alleviated, ensuring the smooth completion of the recommendation calculation task. This distributed computing method not only improves the efficiency of recommendation calculation, but also can dynamically adjust the task allocation according to the actual load conditions of edge computing nodes, ensuring the stability and efficiency of the video recommendation system under various network environments and load conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of a method for fast video content recommendation based on edge computing proposed by the present invention; Figure 2 is a schematic diagram of personalized video recommendation using a reinforcement learning algorithm with a temporal attention diffusion model for a method for fast video content recommendation based on edge computing proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0022] Refer to Figure 1 and Figure 2 , a method for fast video content recommendation based on edge computing, includes the following steps: S1. Schedule the video recommendation task to an edge computing node according to the user's geographical location, network condition, and device capability; S2. On the edge computing node, based on the user's historical viewing data, historical interest preferences, and historical behavior data, use a reinforcement learning algorithm based on a temporal attention diffusion model for personalized video recommendation, construct the user's temporal feature vector, apply the temporal attention mechanism to weight the historical behavior data, capture the impact of recent behaviors on the current recommendation, use an improved diffusion model to model the user's historical behavior data, predict the video content that the user is interested in, and combine the reinforcement learning algorithm to optimize the recommendation strategy to generate the user's personalized recommendation list; S3. According to the computing power and network bandwidth conditions of the edge computing node, adjust the computational complexity of the algorithm, optimize the execution method of the computing task for resource constraints, and reduce the task complexity when the system load is high; S4. Pre-cache the video content that the user is interested in to the edge computing node. When the user selects a video, immediately provide the video content from the edge computing node for playback; S5. When the computing power of the edge computing node is insufficient, cooperate with another group of edge nodes for distributed computing to process the video recommendation task; S6. Collaborate between edge nodes to transfer the video recommendation algorithm inference task to a high-level edge server; S7. According to user feedback, update the video recommendation algorithm in real time.

[0023] In this embodiment, the specific steps of S1 are as follows: S11. Obtain the user's geographical location information through the positioning module from the user device to determine the geographical area where the user is located; S12. By analyzing the user's network status, obtain the current network bandwidth and network latency information, and real-time monitor the network quality between the user device and the edge computing node; S13. Evaluate the CPU performance, memory capacity, and graphics processing ability indicators of the user device; S14. According to the network bandwidth and computing power data obtained in steps S12 and S13, schedule the video recommendation task to the edge computing node. When the network bandwidth and device computing power meet the requirements of the recommendation task, the task is assigned to the current edge computing node. When the network bandwidth or computing power does not meet the requirements, the task is assigned to another edge computing node; S15. After the recommendation task scheduling is completed, feedback the processing status of the current recommendation task to the user device, including whether the task scheduling is successful, the status of the selected edge computing node, and the network quality information.

[0024] In this embodiment, the specific steps of S2 are as follows: S21. On the edge computing node, obtain the user's historical viewing data and construct a historical behavior dataset for the user. For each time step t, extract the viewing records related to the time step from the historical behavior dataset, and represent each viewing record as a feature vector , and construct all the viewing records into a time series in chronological order . Each feature vector represents the viewing behavior feature vector at time step i. The historical viewing data includes the video content viewed by the user, the viewing duration, the viewing frequency, and the timestamp information at each viewing. The feature vector includes the metadata of the video and the viewing behavior features. The metadata includes the video duration, category, and rating. The viewing behavior features include the viewing duration and the viewing frequency; S22. Based on the historical viewing data, construct the user's time-series feature vector. Preprocess the historical viewing data, standardize all feature values using a normalization method, calculate the weighted sum of the feature vectors at each time step, and generate the time-series feature vector , and the time-series feature vector represents the user's viewing interest within a time period: ; Among them, represents the time-series feature vector, represents the weighting coefficient of the i-th viewing record, represents the feature vector at time step i, and n represents the total number of time steps; S23. Calculate the attention weight at each time step t , and the attention weight reflects the importance of the historical behavior data at each time step for the current recommendation: ; Among them, represents the attention weight at time step t, that is, the weighting coefficient of the t-th viewing record. exp represents the natural exponential function, represents the query vector at time step t, represents the query vector at time step i, and represent the time-series feature vector; S24. Dynamically calculate the weighting coefficient , and weight each historical viewing record according to the corresponding attention weight to reflect the influence of recent viewing behavior on the current interest, and generate the weighted time-series feature vector ; S25. Based on the calculated weighted time-series feature vector , an improved diffusion model is applied to model user behavior data, simulate the propagation process of recommended content in the user interest space, and generate user interest states and interest distributions. The improved diffusion model integrates a temporal attention mechanism and combines with a reinforcement learning algorithm to model the user's interest propagation process; S26. Based on the user interest state, a recommendation strategy is optimized by combining a reinforcement learning algorithm; S27. Based on the update result of the reinforcement learning algorithm, a personalized recommendation list for the user is generated. The personalized recommendation list is sorted according to the user's current interest state and the predicted interest values for different video contents, and the sorting result is used to generate a video recommendation result. The recommendation list is generated based on the historical behavior, attention weighted features, the interest propagation process generated by the diffusion model, and the strategy optimized by the reinforcement learning.

[0025] In this embodiment, the S25 specifically includes: S251. Initialize the user interest state As the user interest at the initial moment, the user interest state is the current interest representation of the user, and the initial user interest state is determined according to the user's historical viewing records, interest preferences, and the context information at the current recommendation moment; S252. Set the adjacency matrix A of the diffusion process. The adjacency matrix represents the connection relationship between each node in the user interest space. The elements in the adjacency matrix represent the propagation weight between node i and node j in the user interest space, reflecting the influence from one node to another node. The nodes include video content and viewing time; S253. Design an iterative propagation process based on the improved diffusion model to simulate the dynamic change of the user interest state and update the propagation of the user interest state: ; where, represents the user interest state at time step t, A represents the adjacency matrix of the diffusion process, that is, the propagation matrix of the improved diffusion model, reflecting the propagation relationship between nodes, represents the user interest state at time step t - 1, represents the weight parameter, adjusting the influence of temporal features on the user interest, represents the weighted temporal feature vector, which is the weighted interest of the user behavior at the current time step, represents the propagation weight coefficient, adjusting the influence of neighbor nodes on the current state, represents the weight matrix of each neighbor node i, which is the contribution during the propagation process, represents the user interest state of neighbor node i at time step t - 1; S254. At each time step t, according to the updated interest state calculate the user interest distribution, which represents the preference degree of the user for different video contents at time step t: ; wherein, represents the user interest distribution at time step t, softmax represents normalizing the interest state vector and outputting a probability distribution, which is the interest of the user in each video content, represents the user interest state at time step t, represents the bias vector, which is the adjustment factor of the user interest; S255. Based on the calculated user interest distribution , simulate the spread of the recommended content in the user interest space, and iteratively calculate the change of the user interest state through a multi-step propagation process.

[0026] In this embodiment, the S26 specifically includes: S261. According to the user interest state , construct the current state and the recommended action , the recommended action represents the video content recommended to the user, which is the video content predicted based on , the state represents the interest distribution of the user at the current moment, and is represented by the vector ; S262. Define the reward function , the reward function is comprehensively evaluated based on the user's interaction behavior, the matching degree between the recommended content and the user interest, and the time factor: ; wherein, represents the reward function, , and represent constant coefficients to adjust the weights of various factors in the reward, represents the feature vector of the state , represents the feature vector of the action , represents the cosine similarity between the user interest and the recommended content to measure the matching degree between the two, represents the two-norm operation, represents the initial weight, represents the decay factor, represents the time stamp of the historical behavior, t represents the current time, Indicates that the influence of historical behavior on the current recommended content is controlled by time decay, and the influence of more distant historical data on the recommendation result gradually weakens. Indicates the interaction intensity of the user with the recommended content, where the interaction includes the frequency of the user clicking on the recommended content, the total duration of the user watching the recommended content, and the user's rating or liking of the recommended content. S263. For each state-action pair , calculate the Q value and update it. The Q value represents the expected return of taking a certain action in a given state: ; Among them, represents the state and the value of taking action in it, which is the expected reward for choosing the action. represents the reward function, which is the immediate reward obtained after choosing action in the current state , and is the feedback of the user on the video content. represents the learning rate, which controls the update amplitude. represents the discount factor, which measures the importance of future rewards. represents the policy probability of choosing action in the next state . represents the state and the value of taking action in it. represents the expected Q value based on the policy. S264. After each update of the Q value, generate the optimal recommended action based on the updated Q value, and recommend the video content to the user according to the optimal recommended action. The optimal recommended action is determined by selecting the action with the maximum Q value: ; Among them, represents the optimal recommended action, and argmax(x) represents the variable value that maximizes x. S265. The reinforcement learning algorithm iteratively updates the recommendation policy based on the user feedback. Each time, update the Q value according to the user's interaction feedback on the recommended content and adjust the user's interest state.

[0027] In this embodiment, the specific steps of S3 include: S31. Obtain the computing power and network bandwidth of each edge computing node; S32. Calculate the complexity of the recommendation task. The The computational complexity for task t, representing the amount of computation required for task t: ; Among them, represents the complexity of the recommended task, represents the number of subtasks of task t, represents the weight coefficient of subtask i, represents the computational time complexity of subtask i; S33. According to the computing power of the edge computing node and network bandwidth , calculate the edge node load index: ; Among them, represents the load index of edge computing node n, represents the computing power of the edge computing node, represents the maximum computation of the edge computing node, represents the network bandwidth, represents the maximum bandwidth of the edge node; S34. According to the value of the edge computing node load index, adjust the computational complexity of the recommended task. If , degrade the computational complexity of the recommended task, calculate the size of the candidate video set for the recommended content, and the is the size of the candidate video set for each user in the recommended task: ; Among them, represents the adjusted size of the candidate video set, represents the default size of the recommended candidate set, and floor represents the floor function, which returns the largest integer not greater than the variable; S35. When the edge computing node load index , increase the computational complexity of the recommended task, improve the recommendation accuracy, and expand the size of the recommended candidate set: ; Among them, represents the expanded size of the recommended candidate set, and ceil represents the ceiling function, which returns the smallest integer not less than the variable.

[0028] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the actual environment of a large online video platform. The platform has more than ten million active users, rich video content, and huge and complex viewing history data of users. On this platform, the interest of users in watching videos changes rapidly, and traditional recommendation systems often have difficulty providing personalized and accurate recommendation content due to insufficient computing power and excessive latency. In addition, due to the variety of video content and the diversity of user interests, traditional algorithms often have certain biases in the accuracy of recommendations. Therefore, the platform decides to adopt a method for rapid video content recommendation based on edge computing to optimize its recommendation system and improve the user experience.

[0029] In the implementation process, the platform first intelligently schedules video recommendation tasks to suitable edge computing nodes according to the geographical location, network bandwidth, and device performance of users. These nodes model the interest state of users by applying a temporal attention diffusion model based on the viewing history data of users, and optimize the recommendation strategy through a reinforcement learning algorithm. The recommendation task for each user is calculated on the most suitable edge node, reducing the situation of long waiting for recommendations in the traditional centralized computing mode and improving the recommendation response speed.

[0030] Specifically, the platform divides 5,000 active users into two groups, and tests them respectively using the traditional recommendation algorithm and the method of edge computing plus temporal attention diffusion model proposed by the present invention.

[0031] Table 1 Comparative experimental data of the method for rapid video content recommendation based on edge computing and the traditional recommendation algorithm ; In terms of the recommendation response time, the recommendation method based on edge computing and temporal attention diffusion model shows significant advantages. The response time of the traditional recommendation algorithm is 4.8 seconds, while the method of the present invention shortens the response time to 2.5 seconds, with an improvement rate of 48%. This improvement shows that edge computing nodes can process recommendation tasks more quickly, reducing the user waiting time and thus improving the user experience. Traditional algorithms usually rely on a centralized computing platform, suffering from problems of computing bottlenecks and transmission delays, while edge computing significantly improves the recommendation response speed through distributed processing and data near-source computing.

[0032] The user satisfaction score has increased significantly, from 3.6 points of the traditional algorithm to 4.5 points, with an improvement rate of 25%. This shows that the recommendation method based on the temporal attention diffusion model can capture the interest preferences of users more accurately, and dynamically adjust the recommendation strategy according to the historical behavior of users, thus improving the relevance and personalization degree of recommendations. When users get recommendation content that better meets their needs, they often show higher satisfaction.

[0033] In terms of video click-through rate, the recommendation method based on edge computing and temporal attention diffusion model also shows great advantages. The click-through rate of the traditional recommendation algorithm is 9.3%, while the method of the present invention increases the click-through rate to 15.8%, with an increase of 69%. This means that users are more inclined to click and watch the video content recommended by the method of the present invention, further verifying the effectiveness of this method in accurate recommendation and personalized recommendation. Due to the adoption of the temporal attention mechanism, the algorithm can pay more attention to the recent user behaviors, making the recommended content more in line with the immediate interests of users.

[0034] From the perspective of system computing load, the system load of the traditional recommendation algorithm is relatively high, reaching 90%, while the method of the present invention significantly reduces the system load to only 60% through the distributed processing of edge computing, with a decrease of 33%. This optimization shows that edge computing not only reduces the pressure on the central computing node but also improves the overall processing efficiency of the system. By distributing tasks to multiple edge computing nodes, the computational workload of the recommendation task is shared, making the system more efficient and stable.

[0035] In terms of recommendation accuracy, the method based on the present invention also greatly improves the recommendation accuracy. The recommendation accuracy of the traditional algorithm is 75%, while the recommendation method based on the temporal attention diffusion model reaches 88%, with an increase of 17%. This shows that the method of the present invention can better understand the long-term and short-term interest changes of users, dynamically adjust the recommended content, and more accurately match the user needs, thereby improving the effectiveness of the recommendation.

[0036] In terms of bandwidth utilization, the bandwidth utilization of the traditional recommendation algorithm is 80%, while the recommendation method based on edge computing is reduced to 65%, with a decrease of 18.75%. Behind this change is that edge computing decentralizes the data processing and recommendation computing tasks to local or edge nodes, thus reducing the burden of data transmission and the demand for network bandwidth. By reducing the amount of data transmission, the network latency can be effectively reduced, and the overall system performance can be improved.

[0037] As mentioned above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A method for fast video content recommendation based on edge computing, characterized in that: The steps include: S1. Schedule video recommendation tasks to edge computing nodes based on the user’s geographic location, network conditions, and device capabilities. S2. On the edge computing node, based on the user's historical viewing data, historical interest preferences and historical behavior data, a reinforcement learning algorithm based on the temporal attention diffusion model is used to perform personalized video recommendations, construct the user's temporal feature vector, apply the temporal attention mechanism to weight the historical behavior data, capture the impact of recent behavior on the current recommendation, use the improved diffusion model to model the user's historical behavior data, predict the video content that the user is interested in, and combine the reinforcement learning algorithm to optimize the recommendation strategy to generate the user's personalized recommendation list; S3. Adjust the computational complexity of the algorithm according to the computing power of the edge computing nodes and the network bandwidth conditions, optimize the execution mode of computing tasks according to resource constraints, and reduce the task complexity when the system load is high; S4, pre-caching the video content that the user is interested in to the edge computing node, and when the user selects a video, the video content is immediately provided from the edge computing node for playback; S5. When the computing power of the edge computing node is insufficient, it cooperates with another group of edge nodes to perform distributed computing to process the video recommendation task; S6, collaborate between edge nodes to transfer the video recommendation algorithm reasoning task to the high-level edge server; S7. Update the video recommendation algorithm in real time based on user feedback.

2. According to the method of claim 1, the method is characterized in that: The S1 specifically includes: S11, obtaining the user's geographical location information from the user device through a positioning module to determine the geographical area where the user is located; S12. Analyze the user's network status, obtain the current network bandwidth and network delay information, and monitor the network quality between the user device and the edge computing node in real time; S13, evaluating the CPU performance, memory capacity and graphics processing capability indicators of the user's device; S14. According to the network bandwidth and computing power data obtained in steps S12 and S13, the video recommendation task is scheduled to the edge computing node. When the network bandwidth and device computing power meet the requirements of the recommendation task, the task is assigned to the current edge computing node. When the network bandwidth or computing power does not meet the requirements, the task is assigned to another edge computing node. S15. After the recommended task scheduling is completed, the processing status of the current recommended task is fed back to the user device, including whether the task scheduling is successful, the status of the selected edge computing node, and network quality information.

3. The method for fast video content recommendation based on edge computing according to claim 1, characterized in that: The S2 specifically includes: S21. On the edge computing node, obtain the user's historical viewing data and construct a historical behavior data set of the user. For each time step t, extract the viewing record related to the time step from the historical behavior data set, and represent each viewing record as a feature vector , construct all viewing records into a time series in chronological order , each feature vector represents a viewing behavior feature vector at time step i, the historical viewing data includes the video content, viewing time, viewing frequency and timestamp information of each viewing, the feature vector includes the metadata and viewing behavior features of the video, the metadata includes the video time, category and rating, and the viewing behavior features include the viewing time and viewing frequency; S22, based on the historical viewing data, construct the user's time series feature vector, pre-process the historical viewing data, use a normalization method to standardize all feature values, calculate the weighted sum of the feature vectors of each time step, and generate a time series feature vector , the time series feature vector represents the user’s viewing interest in a time period: ; in, represents the time series feature vector, represents the weight coefficient of the i-th viewing record, represents the feature vector of time step i, and n represents the total number of time steps; S23. Calculate the attention weight for each time step t , the attention weight reflects the importance of historical behavior data at each time step to the current recommendation: ; in, represents the attention weight at time step t, that is, the weighting coefficient of the t-th viewing record, exp represents the natural exponential function, represents the query vector at time step t, represents the query vector at time step i, and represents the time series feature vector; S24, Dynamic calculation of weighting coefficients , each historical viewing record According to the corresponding attention weight Weighted to reflect the impact of recent viewing behavior on current interest, generating a weighted time series feature vector ; S25, based on the calculated weighted time series feature vector , an improved diffusion model is applied to model user behavior data, simulate the propagation process of recommended content in the user interest space, and generate user interest states and interest distributions. The improved diffusion model integrates a temporal attention mechanism and is combined with a reinforcement learning algorithm to model the user's interest propagation process; S26, based on the user's interest status, optimizing the recommendation strategy in combination with a reinforcement learning algorithm; S27. Based on the update results of the reinforcement learning algorithm, a personalized recommendation list for the user is generated. The personalized recommendation list is sorted according to the user's current interest status and the predicted interest value of different video content, and the sorting result is used to generate video recommendation results. The recommendation list is generated according to the historical behavior, attention weighted features, the interest propagation process generated by the diffusion model and the strategy optimized by reinforcement learning.

4. The method for fast video content recommendation based on edge computing according to claim 3, characterized in that: The S25 specifically includes: S251, initializing user interest status As the user interest at the initial moment, the user interest state is the user's current interest representation, and the initial user interest state Determined based on the user's historical viewing history, interest preferences, and contextual information of the current recommendation moment; S252: Set an adjacency matrix A of the diffusion process, wherein the adjacency matrix represents the connection relationship between each node in the user's interest space, and the elements in the adjacency matrix are represents the propagation weight between node i and node j in the user's interest space, reflecting the influence from one node to another, where the nodes include video content and viewing time; S253. Design an iterative propagation process based on the improved diffusion model to simulate the dynamic changes of the user's interest state and update the propagation of the user's interest state: ; in, represents the user interest state at time step t, A represents the adjacency matrix of the diffusion process, that is, the propagation matrix of the improved diffusion model, which reflects the propagation relationship between nodes. represents the user interest state at time step t-1, represents the weight parameter, which adjusts the impact of temporal features on user interests. represents the weighted time series feature vector, which is the weighted interest of the user behavior at the current time step. Represents the propagation weight coefficient, which adjusts the influence of neighbor nodes on the current state. Represents the weight matrix of each neighbor node i, which is the contribution in the propagation process, represents the user interest state of neighbor node i at time step t-1; S254, at each time step t, according to the updated interest state Calculate the user interest distribution, which represents the user's preference for different video contents at time step t: ; in, represents the user interest distribution at time step t, softmax represents the normalization of the interest state vector, and outputs the probability distribution, which is the user's interest in each video content. represents the user interest state at time step t, represents the bias vector, which is the adjustment factor of user interest; S255: User interest distribution based on calculation ,The propagation of recommended content is simulated in the user interest space, and the changes of user interest states are iteratively calculated through a multi-step propagation process.

5. The method for fast video content recommendation based on edge computing according to claim 3 is characterized in that: The S26 specifically includes: S261. Based on user interest status , build the current state and recommended actions , the recommended action Indicates the video content recommended to the user, based on The predicted video content, the state Represents the user's interest distribution at the current moment, through the vector express; S262. Define reward function , the reward function is comprehensively evaluated based on the user's interactive behavior, the matching degree between the recommended content and the user's interests, and the time factor: ; in, represents the reward function, , and Represents a constant coefficient, which adjusts the weight of each factor in the reward. Indicates status The characteristic vector of Indicates action The characteristic vector of Represents the cosine similarity between user interests and recommended content, measuring the degree of match between the two. represents the two-norm operation, represents the initial weight, represents the attenuation factor, Indicates the timestamp of historical behavior, t indicates the current time, It means that the influence of historical behavior on the current recommended content is controlled by time decay, and the influence of more distant historical data on the recommendation results gradually weakens. Indicates the user's interaction intensity with the recommended content, including the frequency of users clicking on the recommended content, the total time users watch the recommended content, and the user's rating or likes for the recommended content; S263, for each state-action pair , calculate and update the Q value, which represents the expected benefit of taking an action in a given state: ; in, Indicates status Take action The value of is the expected reward for selecting an action, Represents the reward function, at the current state Next select action The immediate reward is the user's feedback on the video content. Represents the learning rate, controlling the update amplitude, represents the discount factor, measuring the importance of future rewards, Indicates that in the next state Next select action The probability of strategy, Indicates status Take action The value of represents the expected Q value based on the strategy; S264: After each Q value is updated, the optimal recommended action is generated based on the updated Q value. , and recommends video content to the user according to the optimal recommended action, where the optimal recommended action is determined by selecting the action with the maximum Q value: ； in, represents the optimal recommended action, argmax(x) represents the variable value that maximizes x; S265. The reinforcement learning algorithm iteratively updates the recommendation strategy based on user feedback, updates the Q value each time according to the user's interactive feedback on the recommended content, and adjusts the user's interest status.

6. The method for fast video content recommendation based on edge computing according to claim 1, characterized in that: The S3 specifically includes: S31. Obtain the computing power of each edge computing node and network bandwidth ; S32. Calculate the complexity of the recommendation task , is the computational complexity of task t, indicating the amount of computation required for task t: ; in, represents the complexity of the recommendation task, represents the number of subtasks of task t, represents the weight coefficient of subtask i, represents the computational time complexity of subtask i; S33. Based on the computing power of edge computing nodes and network bandwidth , calculate the edge node load index: ; in, represents the load index of edge computing node n, Indicates the computing power of the edge computing node, represents the maximum computation of the edge computing node, Indicates the network bandwidth. Indicates the maximum bandwidth of the edge node; S34. According to the value of the edge computing node load index, adjust the computational complexity of the recommended task. , reduce the computational complexity of the recommendation task, calculate the size of the candidate video set for recommendation content, The size of the candidate video set for each user in the recommendation task: ； in, represents the adjusted candidate video set size, Indicates the default recommendation candidate set size, floor indicates the rounding function, which returns the maximum integer not greater than the variable; S35, when the edge computing node load index When , the computational complexity of the recommendation task is increased, the recommendation accuracy is improved, and the size of the recommendation candidate set is expanded: ； in, Indicates the size of the expanded recommendation candidate set, and ceil represents the upward rounding function, which returns the smallest integer not less than the variable.

Images