A video online recommendation system and method combined with network information

Abstract

A video online recommendation system combined with network information comprises a video quality evaluation module, a network measurement module and a multi-target reordering module; the recommendation algorithm of the video quality evaluation module selects candidate videos; the network measurement module measures the network condition during the request of the video; the multi-target reordering module estimates the residual transmission time of the candidate videos according to the network bandwidth feedback calculated by the network measurement module, reorders the candidate video list selected by the recommendation algorithm, and selects a preset number of video combinations to transmit to the user end. The present application proposes a new short video recommendation system and method, which takes network bandwidth into the video recommendation system. The system can select the video combination that can be successfully transmitted within a specified time and bring the highest actual income. In the case of poor network bandwidth, compared with the traditional recommendation mechanism, the system can improve the video playing time by 160%, that is, the income is improved.

Description

Technical Field

[0001] This invention relates to the field of short videos, and in particular to an online video recommendation system and method that combines network information. Background Technology

[0002] Short videos have begun to account for a large portion of internet traffic and have proven to have significant commercial value. Existing short video platforms have implemented recommendation systems to encourage users to spend more time on them (Xudong Gong, Qinlin Feng, Yuan Zhang, Jiangling Qin, Weijie Ding, Biao Li, Peng Jiang, and Kun Gai. Real-time short video recommendation on mobile devices. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, CIKM '22, page 3103–3112, New York, NY, USA, 2022. Association for Computing Machinery.). Over the past decade, video recommendation models have demonstrated a strong ability to better match user interests and increase platform revenue. However, improving the accuracy of recommendation models now faces the problem of diminishing marginal utility.

[0003] Specifically, for online short video platforms, if a user is interested in a video, they can watch it for a period of time; if they are not interested, they can quickly "scroll" to another video. The recommendation system helps the server dynamically select several potentially attractive videos based on real-time user feedback and transmit them to the user over the network. These recommended videos need to be transmitted before a very short deadline (before the previous video is scrolled through); otherwise, if the deadline is exceeded, the user can only scroll through some locally cached videos, which may reduce their appeal and cause the user to leave. For example, in Kuaishou (one of China's largest short video platforms), the server attempts to recommend and transmit six videos to the user every 3 seconds before the deadline.

[0004] However, due to network fluctuations, not all recommended videos can be successfully received by users.

[0005] Figure 1The results show that, without considering network conditions, existing recommendation models can lose over 90% of video playback time under poor network conditions. While previous research has attempted to accelerate video transmission by dynamically selecting the video bitrate based on the user's network status, it has rarely considered video selection. In particular, because video sizes can vary, even videos that might be more appealing according to the recommendation model may fail to be delivered to the user due to poor network conditions. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an online video recommendation system and method that combines network information, thereby recommending more attractive videos to users.

[0007] To address the aforementioned technical problems, the present invention proposes the following technical solution: an online video recommendation system incorporating network information, comprising a video quality assessment module, a network measurement module, and a multi-objective reordering module; the video quality assessment module uses a recommendation algorithm to select candidate videos; the network measurement module measures the network conditions during the requested video period; the multi-objective reordering module estimates the remaining transmission time of the candidate videos based on the network bandwidth calculated and fed back by the network measurement module, reorders the candidate video list selected by the recommendation algorithm, and selects a preset number of videos to transmit to the user. The selected preset number of videos achieves the maximum total video playback time and can be transmitted within the deadline under the current network bandwidth.

[0008] In the aforementioned online video recommendation system that combines network information, preferably, the video quality assessment module measures video quality, which includes one or more of the following: playback time, like probability, or share probability.

[0009] In the aforementioned online video recommendation system that incorporates network information, preferably, the network measurement module calculates the bandwidth based on the total transmitted bytes and transmission time in a specific user request, and uses this bandwidth as the network bandwidth for playing video in the next user request to calculate the transmission time.

[0010] A method for online video recommendation that combines network information includes the following steps; 1) The recommendation algorithm in the video quality assessment module selects candidate videos; 2) The network measurement module measures the network bandwidth during the request period; 3) The multi-objective reordering module estimates the remaining transmission time of candidate videos based on the network bandwidth calculated and fed back by the network measurement module, and reorders the candidate video list selected by the recommendation algorithm. 4) The multi-target reordering module selects a preset number of videos to transmit to the user terminal; the selected preset number of videos reaches the maximum total video playback time and can be transmitted within the preset deadline under the current network bandwidth.

[0011] In the above-mentioned online video recommendation method that combines network information, preferably, the step of the network measurement module calculating the transmission time of candidate videos includes: calculating the bandwidth based on the total number of bytes transmitted and the transmission time during the period when a specific user requests a video; Compared with existing technologies, the advantages of this invention are as follows: This invention proposes a novel short video recommendation system and method that incorporates network bandwidth into the video recommendation system. This system can successfully transmit and generate the highest actual revenue from video combinations within a specified time. Even with poor network bandwidth, compared to traditional recommendation mechanisms, this system can increase video playback time by 160%, resulting in increased revenue. Attached Figure Description

[0012] Figure 1 The average video playback time is defined by existing recommendation models that do not consider network conditions and ideal recommendation models that jointly consider network conditions.

[0013] Figure 2 This describes the process by which the online video recommendation system in Example 1, which incorporates network information, evaluates the total playback time of videos.

[0014] Figure 3 The graph shows the revenue improvement rate of the online recommendation system in Example 1 compared to the traditional recommendation system under different bandwidth conditions.

[0015] Figure 4 This is a graph showing the improvement in the success rate of the online recommendation system in Example 1 compared to a traditional recommendation system under different bandwidth conditions.

[0016] Figure 5 This example illustrates the detailed revenue growth of two randomly selected users from the data in Example 1 within the campus network, where each user watched over 2000 videos. Detailed Implementation

[0017] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0018] It should be noted that when a component is described as being "fixed to, attached to, connected to or connected to" another component, it can be directly fixed to, attached to, connected to or connected to the other component, or it can be indirectly fixed to, attached to, connected to or connected to the other component through other intermediate connectors.

[0019] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of the invention. Example 1

[0020] An online video recommendation system incorporating network information includes a video quality assessment module, a network measurement module, and a multi-objective reordering module. The video quality assessment module uses a recommendation algorithm to select candidate videos. The network measurement module measures the network conditions during the requested video period. The multi-objective reordering module estimates the remaining transmission time of the candidate videos based on the network bandwidth calculated and fed back by the network measurement module, reorders the candidate video list selected by the recommendation algorithm, and selects a preset number of videos to transmit to the user. The selected preset number of videos maximizes the total video playback time and can be transmitted within the deadline given the current network bandwidth.

[0021] Video quality assessment module

[0022] The primary goal of short video recommendation is to estimate the video quality for a specific user. Video quality reflects a user's preference for videos. Various metrics exist to measure video quality, such as estimated playback time, like probability, or share probability. This embodiment uses estimated playback time as a metric for video quality. Many models have been proposed to accurately predict video quality by taking into account video features and the user's viewing, like, or sharing history. The video quality assessment module of this invention can be a known model for predicting video quality.

[0023] Multi-objective reordering module

[0024] The main objective of this module is to reorder the initial video list selected by existing recommendation algorithms. Our reordering algorithm needs to work in a multi-objective manner, considering not only video quality but also the likelihood of successful delivery to the user. The problem of the multi-objective reordering module can be formulated as: how to select a set of M videos that achieve the highest possible total video playback duration and can be transmitted within a deadline of T seconds given a certain network bandwidth and video size.

[0025] Based on the above formula, this problem can be viewed as an optimization problem and solved using the knapsack algorithm. It iterates through each video, considering two choices when placing it into the knapsack: either include it or don't. The optimal choice is made between these two options. Scenario 1: When x < M, (Mx) videos with the longest estimated playback time are selected from the remaining videos not selected by the knapsack algorithm. Scenario 2: When x = M, NetRec has obtained M videos. Scenario 3: When x > M, the selected videos are reordered, and the top M videos with the longest estimated playback time are selected. Here, x is the number of original videos selected by the recommendation algorithm in the video quality assessment module.

[0026] Network measurement module

[0027] To ensure that the reordering algorithm provides accurate network information for better decision-making, continuous measurement of real-world network bandwidth is necessary. To minimize computational burden and traffic, bandwidth is estimated by tracking the total bytes transmitted and transmission time in a specific user request. This result is recorded and used for future user requests. This passive measurement method offers two key advantages: first, it eliminates the need for additional measurement traffic or software; second, the resulting bandwidth estimate accurately reflects the achievable bandwidth of the link with the video provider.

[0028] A method for online video recommendation that combines network information includes the following steps; 1) The recommendation algorithm in the video quality assessment module selects candidate videos; 2) The network measurement module measures network bandwidth during the request period; it calculates bandwidth based on the total number of bytes transmitted and the transmission time during the video request period for a specific user. 3) The multi-objective reordering module estimates the remaining transmission time of candidate videos based on the network bandwidth calculated and fed back by the network measurement module, and reorders the candidate video list selected by the recommendation algorithm. 4) The multi-target reordering module selects a preset number of videos to transmit to the user terminal; the selected preset number of videos reaches the maximum total video playback time and can be transmitted within the preset deadline under the current network bandwidth.

[0029] 5) Select the IDs of the first M videos and send them to the user's device to retrieve content from the service provider or a nearby content provider. Any videos retrieved after a deadline (e.g., 3 seconds) will not be displayed to the user.

[0030] To verify the effectiveness of the system provided in this embodiment, the actual playback time of the displayed video was recorded as user feedback to evaluate the effectiveness of our model. It is worth noting that this feedback can also be used to train the video quality estimation model.

[0031] Experimental evaluation

[0032] Dataset. To realistically evaluate the effectiveness of this embodiment, we used Kuaishou's public dataset, which records information on over 10,000 videos watched by millions of users of one of China's largest online video providers over two months (July 5, 2020 to September 5, 2020). For each video, the dataset records its original duration (videoDuration), the actual viewing time by the user in each recommendation (playingTime), the total playback time (playingDuration), and the total number of views within the measurement period (showCnt). The estimated playback time of the recommendation model can be approximated as... .

[0033] Evaluation Metric. In this invention, the total video playback time is used as the evaluation metric, referred to as total revenue. Total revenue is calculated by summing the playback times of all videos recommended through the network and successfully sent to users during each evaluation round.

[0034] Test bench setup. Figure 2 The evaluation process is shown.

[0035] When a client requests video from the server, it attaches network information and starts a timer. After the request reaches the server, the user ID and network information are recorded. The server selects the video to send based on the network information and existing data. In our experiment, we counted the playback time of the received video before the timer expired.

[0036] In our experiments, we used three servers as the server side and two ordinary computers as the client side. Since each user in the dataset watched more than 2,000 videos, fair evaluation can be obtained even if only one user is selected. Therefore, users were randomly selected from the dataset for evaluation; these users watched a total of 3,234 videos. Twelve videos were selected as candidates for each request, and 269 requests were sent per round to measure the revenue of our recommendation system and the baseline. All results below were collected from over 300 rounds of experiments conducted from December 2022 to March 2023. The baseline refers to the traditional recommendation system, which selects M videos based on their estimated playback durations from largest to smallest. Our recommendation system first selects N videos based on their estimated playback durations from largest to smallest, then calculates the transmission time of these N videos, and selects M videos for transmission, aiming to ensure that all M videos are transmitted within the deadline and achieve the maximum total playback duration. Typically, N is twice M.

[0037] Network Environment. The experiment was conducted in three network environments: 1) Laboratory LAN: Both the server and client were deployed in a small laboratory network directly connected via a 1Gbps cable and several switches. 2) Campus Network: The server and client were located in a university campus network with approximately 40,000 users. This network contained approximately 100,000 wired and wireless switches and routers. The server accessed the campus network via a 1Gbps cable, and the client accessed the campus network via a WiFi 5 wireless router. 3) Internet: The server and client were located in Nanjing and Changsha, respectively, in different provinces of China. They communicated through a network provided by a Chinese internet service provider. The server used a wired cable as the access link, and the client accessed the network through a home wireless router.

[0038] Overall performance: Figure 3 The results show that, compared to traditional recommendation systems, NetRec improves revenue across various network environments due to differences in actual network bandwidth between servers and clients. NetRec's revenue increased by 109.29%-12357.31%, 13.26%-1545.86%, and 0.18%-72.82% for network bandwidths of 0-2 MBps, 2-10 MBps, and 10-150 MBps, respectively. Generally, it increases revenue more when network bandwidth is lower and less when network bandwidth is higher. This is because with sufficient network bandwidth, recommended videos have a greater chance of successful transmission even without considering network conditions during the recommendation process. However, when network speeds are slow, transmission conditions need to be carefully considered when recommending videos; otherwise, clients may not receive the video due to network limitations. We output the improvement in successful transmission rate of our algorithm relative to the baseline, such as... Figure 4 As shown, regardless of network environment and bandwidth, our algorithm can effectively reduce the timeout rate. A lower timeout rate means that the recommendation system of this invention can ensure that more videos are successfully presented to the user.

[0039] To demonstrate that the increased revenue of the recommendation system of this invention was not due to selecting a user dataset that was more favorable to us, we also randomly selected two other users who watched videos of 2K or higher during the statistical period. Figure 5 As shown, despite using datasets from different users, NetRec still achieved good revenue, with average revenue increasing by approximately 80% across different rounds of evaluation.

[0040] In this invention, we propose a novel short video recommendation system and method that incorporates network bandwidth into the video recommendation system. This system is capable of successfully transmitting video combinations that generate the highest actual revenue within a specified time. Even under conditions of poor network bandwidth, this system can achieve an improvement of approximately 160% compared to traditional recommendation mechanisms.

Claims

1. A video online recommendation system combined with network information, characterized in that: The system includes a video quality assessment module, a network measurement module, and a multi-objective reordering module. The network measurement module measures the network conditions during the video request period. The multi-objective reordering module estimates the remaining transmission time of candidate videos based on the network bandwidth calculated and fed back by the network measurement module, reorders the candidate video list selected by the recommendation algorithm, and selects a preset number of videos to transmit to the user. The selected preset number of videos achieves the maximum total video playback time and can be transmitted within the deadline under the current network bandwidth. The video quality assessment module measures video quality, which includes one or more of the following: playback time, like probability, or share probability. The recommendation algorithm of the video quality assessment module selects candidate videos. 2.The video online recommendation system integrated with network information according to claim 1, characterized in that: The network measurement module calculates the bandwidth based on the total number of bytes transmitted and the transmission time in a specific user request, and uses this bandwidth as the network bandwidth for playing video in the next user request to calculate the transmission time. 3.A method for online recommendation of videos combined with network information, characterized in that: Includes the following steps; 1) The recommendation algorithm in the video quality assessment module selects candidate videos; 2) The network measurement module measures network bandwidth during the request period; 3) The multi-objective reordering module estimates the remaining transmission time of candidate videos based on the network bandwidth calculated and fed back by the network measurement module, and reorders the candidate video list selected by the recommendation algorithm. 4) The multi-target reordering module selects a preset number of videos to transmit to the user terminal; the selected preset number of videos reaches the maximum total video playback time and can be transmitted within the preset deadline under the current network bandwidth. 4.The method of Claim 3, wherein: The step of the network measurement module to calculate the transmission time of candidate videos includes: calculating the bandwidth based on the total number of bytes transmitted and the transmission time during the period when a specific user requests a video.

Images